<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Stephen Woodard on Medium]]></title>
        <description><![CDATA[Stories by Stephen Woodard on Medium]]></description>
        <link>https://medium.com/@thanis_ai?source=rss-2ee1be8ae92b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*mCTApMJw83ObHyAhEynihw.jpeg</url>
            <title>Stories by Stephen Woodard on Medium</title>
            <link>https://medium.com/@thanis_ai?source=rss-2ee1be8ae92b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 16:51:40 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@thanis_ai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Organizations Need to Hear What Their Leaders Actually Think]]></title>
            <link>https://medium.com/thanis-insights/organizations-need-to-hear-what-their-leaders-actually-think-e9841d46e0e6?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9841d46e0e6</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[thought-leadership]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[executive-communication]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Tue, 16 Jun 2026 12:44:25 GMT</pubDate>
            <atom:updated>2026-06-16T12:44:25.558Z</atom:updated>
            <content:encoded><![CDATA[<p>As AI-generated leadership content becomes increasingly common, preserving the executive’s voice may become one of the most important communication challenges facing modern organizations.</p><figure><img alt="Photo by &lt;a href=”https://unsplash.com/@wocintechchat?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText&quot;&gt;Christina @ wocintechchat.com M&lt;/a&gt; on &lt;a href=”https://unsplash.com/photos/people-on-conference-table-looking-at-talking-woman-Q80LYxv_Tbs?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText&quot;&gt;Unsplash&lt;/a&gt;" src="https://cdn-images-1.medium.com/max/1024/1*tUBs1PNTbWOhigZBpKsuvw.jpeg" /><figcaption>Leadership Meeting</figcaption></figure><p>Earlier this year, I wrote about something I called the Artificial CTO, which was my way of describing a pattern I kept seeing across LinkedIn, executive updates, strategy posts, and leadership commentary that looked polished on the surface but felt strangely empty underneath.</p><p>At the time, it was mostly an observation from the field. I had spent years around executive teams, transformation programs, cloud migrations, technology modernization efforts, architecture reviews, acquisition discussions, and governance meetings, and I was starting to notice that a lot of leadership communication had developed the same clean, confident, overly structured tone, where every paragraph sounded perfectly shaped, every lesson sounded neatly packaged, and every executive somehow seemed to have discovered the exact same voice at the exact same moment.</p><p>Then the numbers started showing up.</p><p>Originality.ai published a study reporting that AI-generated posts in LinkedIn’s Leadership and Inspiration category were receiving roughly 75 percent higher engagement than content believed to be written by humans. That number matters because it tells us something uncomfortable about the environment leaders are now communicating inside. The algorithm is not necessarily rewarding hard-won experience, messy judgment, careful reasoning, or actual leadership. It may be rewarding the kind of polished, repeatable, emotionally predictable content that AI systems are very good at producing.</p><p>That does not mean leaders should stop using AI. That would be a silly argument, and honestly, not a very useful one. I use AI every day. Most people working in technology use it every day. These tools are helpful, fast, and increasingly capable, and when used properly, they can absolutely help people organize ideas, sharpen language, and move faster through routine communication.</p><p>The problem I became interested in was different.</p><p>If AI can help leaders sound polished, persuasive, strategic, thoughtful, inspirational, and executive-ready in seconds, how does an organization know whether it is still hearing from the leader?</p><p>That question may sound simple, but it gets serious very quickly, because leadership communication is not just content. Inside organizations, leadership communication shapes direction, creates confidence, frames risk, builds momentum, explains priorities, influences investment, and often determines how people understand what is happening around them.</p><blockquote>A board recommendation is not just a document. It can shape how a board thinks about risk.</blockquote><blockquote>An acquisition assessment is not just a memo. It can shape whether a company decides to buy, partner, pause, or walk away.</blockquote><blockquote>A transformation proposal is not just a strategy update. It can shape budgets, staffing, technology choices, timelines, and expectations across an entire organization.</blockquote><blockquote>A security incident communication is not just a response. It can shape trust with customers, employees, regulators, investors, and partners.</blockquote><p>This is why the distinction between writing and thinking matters.</p><p>For most of my career, I have worked in the space between leadership intent and delivery reality, where executives describe a direction, architects try to turn that direction into something real, delivery teams inherit the complexity, and eventually everyone finds out whether the original recommendation was grounded in reality or just sounded good enough to move forward.</p><p>That experience taught me something that I think many organizations quietly understand but rarely say out loud.</p><p>A lot of important decisions are already halfway made before the meeting begins.</p><p>They begin in the document that gets circulated beforehand. They begin in the executive summary. They begin in the business case. They begin in the recommendation that frames the choices. They begin in the assumptions that are accepted without much pressure because the writing sounds confident, the argument appears organized, and the proposal feels complete enough to keep moving.</p><p>That was the problem I wanted to solve with Thanis Executive.</p><p>I did not build Executive because leaders needed another tool to write faster. The market is already flooded with tools that can generate text, summarize documents, rewrite emails, clean up tone, produce slides, and make almost anything sound more professional than it probably deserves to sound.</p><blockquote>The missing piece was not more generation.</blockquote><blockquote>The missing piece was review.</blockquote><p>Before Thanis Executive, a leader preparing an important communication had to do what most of us already do. You reread the document several times, clean up the wording, check whether the message sounds clear, maybe ask a trusted colleague for feedback, maybe run it through a writing tool, and then eventually send it forward, hoping that the recommendation is as strong as the communication makes it appear.</p><p>What usually does not happen, at least not in a structured way, is a deeper review of the reasoning itself.</p><blockquote>What assumption is this recommendation depending on?</blockquote><blockquote>What happens if that assumption is wrong?</blockquote><blockquote>Is the evidence strong enough to support the conclusion?</blockquote><blockquote>Does the recommendation clearly explain ownership and accountability?</blockquote><p>Is the message still in the leader’s voice, or has it slowly drifted into that strange, polished, generic AI tone that sounds impressive but does not feel like anyone in particular?</p><p>Those are the questions Executive was built to ask.</p><p>Thanis Executive is a review framework for leadership communication. It is designed to help executives, founders, directors, advisors, and decision-makers evaluate important written communication before that communication reaches boards, investors, customers, employees, partners, or other stakeholders.</p><p>It reviews the communication through the lens of reasoning quality, recommendation strength, accountability, decision readiness, communication force, and voice ownership.</p><p>Simply stated, Executive helps leaders pressure-test the message before the message starts influencing people.</p><p>That matters because the benefit to the organization is very practical. When leaders communicate more clearly, organizations align faster. When assumptions are surfaced earlier, teams can challenge weak points before they become expensive decisions. When reasoning is stronger, recommendations become easier to defend. When accountability is visible, ownership becomes clearer. When the leader’s voice is preserved, people are not left trying to interpret whether they are hearing actual judgment or machine-shaped language dressed up as leadership.</p><p>One of the capabilities inside Executive that I care about most is Critical Assumption Analysis.</p><p>The idea came from reading and reviewing a wide range of leadership communications where the same pattern kept showing up. A recommendation could sound strong, flow well, and feel persuasive, while still depending heavily on a single assumption sitting quietly underneath the argument.</p><blockquote>Sometimes that assumption was about growth.</blockquote><blockquote>Sometimes it was about customer adoption.</blockquote><blockquote>Sometimes it was about cost savings.</blockquote><blockquote>Sometimes it was about the readiness of the organization.</blockquote><blockquote>Sometimes it was about whether teams had enough people, enough time, enough budget, or enough technical maturity to actually deliver what the recommendation promised.</blockquote><p>The assumption was not always wrong, but it was often invisible.</p><p>Critical Assumption Analysis is designed to bring that assumption into view so the leader can challenge it, strengthen it, validate it, or decide that the recommendation needs more work before it moves forward.</p><p>That is the heart of what Executive does.</p><blockquote>It does not replace the leader.</blockquote><blockquote>It does not generate the strategy.</blockquote><blockquote>It does not pretend to own the decision.</blockquote><p>It gives the leader a clearer view of the reasoning inside the communication, which is exactly where many important decisions begin.</p><p>The reason this matters now is that AI has changed the speed and surface quality of professional communication. A few years ago, if someone sent a polished executive memo, you could reasonably assume a certain amount of effort went into shaping it. Today, polish is cheap. Structure is cheap. Confidence is cheap. A clean executive tone is cheap.</p><blockquote>What is not cheap is judgment.</blockquote><blockquote>What is not cheap is lived experience.</blockquote><blockquote>What is not cheap is accountability.</blockquote><blockquote>What is not cheap is a leader clearly explaining what they believe, why they believe it, what assumptions they are making, what risks they see, and what decision they are willing to stand behind.</blockquote><blockquote>That is why I keep coming back to this simple idea.</blockquote><blockquote>We want to hear from our leaders.</blockquote><blockquote>Not because AI is bad.</blockquote><blockquote>Not because leaders should never use tools.</blockquote><p>Not because every sentence has to be typed manually in some heroic display of purity.</p><p>We want to hear from our leaders because organizations depend on the judgment of the people responsible for making decisions, and when that judgment gets buried under synthetic polish, everyone downstream inherits the risk.</p><p>Thanis Executive was built for that moment before the communication goes out, before the recommendation hardens, before the meeting starts, before the decision gains momentum, and before everyone starts acting on a document that may sound stronger than it really is.</p><p>It gives leaders a way to slow down just enough to ask better questions of their own communication.</p><blockquote>Does this recommendation hold up?</blockquote><blockquote>Does the reasoning support the conclusion?</blockquote><blockquote>What assumption am I relying on?</blockquote><blockquote>Where is the accountability?</blockquote><blockquote>Does this still sound like me?</blockquote><p>That last question may become one of the most important questions in leadership communication over the next few years, because as AI becomes more common, the challenge will not simply be whether leaders can communicate more efficiently.</p><p>They will.</p><p>The harder challenge will be whether organizations can still recognize the person behind the message.</p><blockquote>That is why we built Executive.</blockquote><blockquote>Not to help leaders sound more like AI.</blockquote><blockquote>To help leaders sound more clearly like themselves.</blockquote><p><strong>Note:</strong> Thanis Executive is now available on the Thanis platform at <a href="https://thanis.ai/">Thanis.ai</a>. You can also read the official press release here: <a href="https://world.einnews.com/pr_news/919778695/thanis-introduces-executive-review-framework-for-leadership-communication">Thanis Introduces Executive Review Framework for Leadership Communication</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9841d46e0e6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thanis-insights/organizations-need-to-hear-what-their-leaders-actually-think-e9841d46e0e6">Organizations Need to Hear What Their Leaders Actually Think</a> was originally published in <a href="https://medium.com/thanis-insights">Thanis Insights</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Thanis Academic: A Student Writing Tool Built for Revision, Rubrics, and Responsible AI Use]]></title>
            <link>https://medium.com/thanis-insights/thanis-academic-a-student-writing-tool-built-for-revision-rubrics-and-responsible-ai-use-95ff5fa36dcf?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/95ff5fa36dcf</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai-in-education]]></category>
            <category><![CDATA[students]]></category>
            <category><![CDATA[academic-writing]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Wed, 29 Apr 2026 08:36:16 GMT</pubDate>
            <atom:updated>2026-04-29T08:53:58.854Z</atom:updated>
            <content:encoded><![CDATA[<p>How Thanis helps students improve academic writing without replacing the student, the instructor, or the work itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3YTac_visoJ3TCDxmGQtew.png" /><figcaption>Thanis.ai console Academic Paper Analysis</figcaption></figure><p>Lets be honest, students are already using generative AI.</p><p>That is no longer a prediction, a campus rumor, or some distant technology trend waiting politely outside the classroom door. It is already inside the workflow. In the UK, the 2025 HEPI/Kortext Student Generative AI Survey found that <strong>88% of students had used generative AI to help with assessments</strong>, up sharply from the previous year. The same report found that <strong>18% of students had included AI-generated or AI-edited text directly in their assessed work</strong>, which is exactly where the academic integrity questions become more serious.</p><p>Globally, the pattern looks similar. The Digital Education Council’s 2024 Global AI Student Survey found that <strong>86% of students already use AI in their studies</strong>, with <strong>54% using it weekly</strong>, and ChatGPT was reported as the most widely used tool at <strong>66%</strong>. In other words, this is not niche experimentation anymore. This is mainstream student behavior, and it is happening faster than most academic policies, classroom norms, and assessment workflows can comfortably absorb.</p><p>That creates a real problem, but not the one people usually name first.</p><p>The problem is not simply that students are using AI.</p><p>The problem is that many students are using AI without enough structure, without enough proof of process, and without a clear boundary between support and substitution.</p><p>A student may ask AI to explain a difficult concept, which can be useful. Another student may ask it to edit a paragraph, which can also be useful if handled carefully. But another student may ask it to generate a full paper, polish the language beyond recognition, invent sources, or reshape the work so heavily that the final submission no longer reflects the student’s own thinking. At that point, the tool has not supported the writing process. It has quietly taken possession of it.</p><p>That is the space Thanis Academic is built for.</p><p>Thanis Academic is not designed to write the paper for the student. It is designed to help the student understand, evaluate, and revise the paper they already wrote.</p><p>That distinction is the whole machine.</p><h3>What Thanis Academic actually is</h3><p>Thanis Academic is a structured academic review mode inside Thanis that helps students analyze their writing after they have created a real draft. It looks at the student’s work through an academic lens, then returns guidance about clarity, argument, structure, evidence, rubric alignment, tone, and revision priorities.</p><blockquote>The student still writes the paper.</blockquote><blockquote>The student still owns the thesis.</blockquote><blockquote>The student still checks the sources.</blockquote><blockquote>The student still decides what to revise.</blockquote><p>Thanis does not replace the instructor, and it does not pretend to be the final authority on grading. It gives the student a clearer reading of their work before they submit it, which is often the exact kind of support students need but do not always know how to ask for.</p><p>That is why Academic mode matters. It does not begin with a blank prompt box and leave the student guessing how to use it properly. It begins with a guided project setup, so the system can understand what kind of academic work the student is bringing into the review.</p><h3>The project setup: how Thanis calibrates the review</h3><p>When a student starts a new Academic project, Thanis does not immediately ask for the paper and rush into feedback. It first walks the student through a simple three-step setup process.</p><p>That setup matters because a history essay, a business case study, a lab report, and a reflective assignment should not be evaluated in exactly the same way. They have different expectations, different structures, different evidence standards, and different kinds of academic risk.</p><p><strong>In Step 1, Thanis asks the student:</strong> ‘<strong>What are you writing’?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tC9irC_6UiMbSp178fvopQ.png" /><figcaption>Thanis Project Setup</figcaption></figure><p>The student can choose from project types such as:</p><blockquote>Essay / Argument</blockquote><blockquote>Research Paper</blockquote><blockquote>Case Study</blockquote><blockquote>Lab Report</blockquote><blockquote>Reflection</blockquote><blockquote>Creative / Other</blockquote><p>This first selection gives Thanis a basic understanding of the assignment shape. An argumentative essay needs attention to thesis strength, claim development, counterargument, and persuasive structure. A research paper needs stronger attention to cited evidence, source integration, research framing, and the relationship between claims and support. A lab report needs structure around method, results, interpretation, and scientific clarity. A reflection may need a different standard altogether, where the goal is not a rigid thesis defense but thoughtful connection between experience, learning, and insight.</p><p>That first choice keeps the review Thanis provides the user from being generic.</p><p><strong>In Step 2, Thanis asks the student to select the academic discipline.</strong></p><p>The student can choose from areas such as:</p><blockquote>History &amp; Humanities</blockquote><blockquote>STEM &amp; Science</blockquote><blockquote>Business</blockquote><blockquote>Literature</blockquote><blockquote>Social Science</blockquote><blockquote>Other</blockquote><p>This second selection gives the system more academic context. A literature paper may need close reading, textual interpretation, and attention to critical framing. A social science paper may need clearer research logic, theoretical grounding, or evidence-based analysis. A business paper may need practical reasoning, case evaluation, and decision logic. STEM writing may need precision, method, sequence, and disciplined explanation.</p><p>This is important because students often receive feedback that feels too general: “Be clearer,” “support your argument,” “develop this idea,” “connect this back to the prompt.” Those comments may be accurate, but they do not always show the student what clarity looks like inside a specific academic domain.</p><p>Thanis Academic narrows the frame before it begins reading.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5YawzxfzoSCiM2RJIAA1oQ.png" /><figcaption>Thanis Academic Dicipline Selection</figcaption></figure><p><strong>In Step 3, Thanis asks for optional assignment context.</strong></p><p>This is where the student can paste the assignment prompt, describe the requirements, or upload a rubric file such as a PDF, Word document, or spreadsheet. The student might add context like: “This is a five-page history paper. It must use three primary sources. The instructor wants us to avoid passive voice.” That information helps Thanis review the work against the actual expectations of the assignment instead of offering detached writing advice in the abstract.</p><blockquote>This is where Thanis becomes especially useful.</blockquote><blockquote>A student without a rubric can still get academic feedback.</blockquote><blockquote>A student with a rubric can get more targeted feedback.</blockquote><p>A student with a prompt, project type, discipline, and rubric can get feedback that is much closer to the real assignment environment.</p><p>That does not mean Thanis is “grading” the paper in the instructor’s place. A better way to say it is this: Thanis uses the setup information to pre-calibrate its academic review, so the feedback is shaped around the kind of work the student is actually doing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gWsc10610qruIkkSOlFSAA.png" /><figcaption>Thanis Rubric / Assignment Upload Prompt</figcaption></figure><h3>Why this guided setup matters</h3><p>Most students do not need another empty AI box.</p><p>They need a structured process.</p><p>A blank prompt box places the burden on the student to know what to ask, how to ask it, how much context to provide, and how to interpret the answer. That sounds simple until you remember that many students are still learning how to understand assignment expectations in the first place.</p><p>Thanis Academic reduces that friction by asking the right questions first.</p><blockquote>What are you writing?</blockquote><blockquote>What discipline is this for?</blockquote><blockquote>What assignment context or rubric should the system consider?</blockquote><p>Those questions are not cosmetic. They shape the entire review. They tell Thanis whether to prioritize argument, evidence, method, reflection, analysis, organization, or rubric alignment. They also help students slow down and think about the assignment before they ask for feedback.</p><p>That small pause matters.</p><p>It reminds the student that academic writing is not just “content.” It is purpose, context, discipline, evidence, and judgment.</p><h3>How rubric support works</h3><p>Rubric support is one of the strongest features in Thanis Academic because it connects the student’s draft to the criteria that may actually be used to evaluate it.</p><p>A rubric is not just a grading sheet. It is a map of instructor expectations. It tells the student what matters: thesis, organization, evidence, analysis, source use, formatting, originality, clarity, grammar, or discipline-specific requirements.</p><p>When a student uploads or pastes a rubric, Thanis can review the draft with those criteria in mind. It can help identify where the paper appears aligned, where it may fall short, and which areas deserve revision before submission.</p><p>For example, if the rubric emphasizes argument, Thanis can look at whether the thesis is visible, specific, and sustained throughout the paper. If the rubric emphasizes evidence, Thanis can examine whether the student is merely dropping in sources or actually interpreting them. If the rubric emphasizes organization, Thanis can identify whether the paper builds logically or drifts between points without enough connective tissue.</p><p>This kind of feedback helps students because many academic mistakes are not obvious at the sentence level. A paper can sound polished and still fail the assignment. A paragraph can be grammatically clean and still not support the thesis. A conclusion can be fluent and still introduce a new argument too late.</p><p>Rubric-aware feedback helps students see that difference. Rubric support is easier to understand when you see it applied to an actual draft, so here is a simple example of what Thanis Academic does after a student gives it a paper, assignment context, and a writing goal.</p><h3>What this looks like in practice</h3><p>To make this less abstract, I tested Thanis Academic with a sample student-style paper titled <strong>“The Role of Structured Feedback in Improving Student Writing Outcomes.”</strong> The paper argued that students often struggle not because they lack effort, but because they do not receive clear, actionable feedback early enough in the writing process. It covered common writing challenges, the feedback gap, rubric alignment, clarity, iteration, and the limits of traditional instructor feedback cycles.</p><p>In other words, the draft was not empty or careless. It had a clear academic topic, a logical structure, and a reasonable argument. It made the case that earlier feedback can improve student writing by helping students understand where their work aligns, or fails to align, with assignment expectations. The paper also emphasized that writing quality depends on clarity, structure, rubric awareness, and revision before submission.</p><p>That made it a useful test case because it was not a terrible draft. It was the kind of paper many students might feel reasonably good about before turning it in.</p><p>Thanis Academic returned scores across four areas:</p><blockquote>Clarity: 60%</blockquote><blockquote>Structure: 50%</blockquote><blockquote>Mechanics: 60%</blockquote><blockquote>Evidence: 40%</blockquote><p>That output matters because the scores immediately show the student that the paper is not failing everywhere, but it also is not equally strong across all areas. The issue is not simply “the writing is bad.” The issue is more specific: the draft has a coherent idea, but it needs stronger structure and much stronger evidence.</p><p>That is exactly the kind of feedback students need before they submit.</p><h3>What the Summary tab showed</h3><p>In the Summary tab, Thanis identified the central argument of the paper: earlier, rubric-referenced feedback can help close the gap between student effort and academic performance by enabling meaningful revision before submission.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/910/1*mSmO6X2s3Bqy3MjXVJrNJw.png" /><figcaption><strong>Thanis Academic summarizes the paper’s core argument, then identifies the deeper revision issue: the draft is coherent, but it needs stronger evidence and scholarly grounding.</strong></figcaption></figure><p>That is useful because it tells the student, first, whether the main argument is visible. A student may think their argument is obvious because they have been living inside the draft for hours or days, but an outside reader may see it differently. Thanis confirms that the paper does have a readable central claim and that the overall movement of the paper runs from problem identification, to mechanism analysis, to a call for scalable feedback systems.</p><p>But the Summary tab did not stop at praise. It also raised a sharper consideration: the argument is coherent, but it stays too much at the level of assertion. The paper describes the value of feedback, but it does not yet demonstrate that value through evidence, case data, scholarly grounding, or concrete examples.</p><p>That is the important part.</p><p>A student reading that feedback would not just walk away thinking, “I need better grammar.” They would understand that the deeper problem is academic support. The paper is making claims, but it has not yet earned those claims with enough outside evidence.</p><p>The Reflection section then turns that diagnosis into a revision move: anchor two or three central claims to external sources or empirical examples so the paper can move from description into argument.</p><p>That is the kind of feedback that helps a student revise intelligently instead of randomly.</p><h3>What the Evidence tab showed</h3><p>The Evidence tab was even more direct.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*E5M7XqSBWdRwdwhpoJHKtQ.png" /><figcaption>The Evidence tab pushes beyond surface-level feedback by identifying where the paper needs counterarguments, stronger support, and a more developed academic position</figcaption></figure><p>Thanis noted that the paper’s central argument, that earlier feedback improves writing outcomes, is unlikely to surprise anyone. That may sound blunt, but it is extremely useful academic feedback because many student papers make arguments that are reasonable but too safe. The claim may be true, but if it does not encounter resistance, complication, or competing interpretations, the paper can feel more like a report than an argument.</p><blockquote><em>Thanis also pointed out that the paper did not engage with counterarguments, competing models, or the conditions under which structured feedback might fail or be misapplied.</em></blockquote><p>That feedback is valuable because it moves the student beyond the first version of the idea. It asks the student to think more critically. Under what conditions does early feedback actually help? When might it become too prescriptive? What happens if students rely on feedback tools without developing independent judgment? What if rubric alignment improves performance but narrows originality? What if structured feedback helps some students more than others?</p><p>Those are better academic questions.</p><p>The Evidence tab also noticed that the draft’s “Limitations of Traditional Approaches” section acknowledged constraints, such as instructors not having enough time to provide detailed feedback on multiple drafts, but did not fully interrogate those constraints. In plain language, the paper named a problem but did not examine it deeply enough.</p><p>That is a subtle but important difference.</p><p>A weaker feedback tool might say, “Add more evidence.” Thanis is more specific. It tells the student that one section gestures toward complexity but does not yet do enough with it.</p><p>Then it gives a practical suggestion: introduce at least one condition under which structured feedback is less effective or harder to implement, then address that condition. That would not weaken the paper. It would sharpen it.</p><p>This is where Thanis Academic becomes especially useful for students. It does not simply reward polish. It pushes the writer toward stronger academic thinking. The student still has to do the research, make the decision, revise the section, and choose which evidence belongs in the paper. But Thanis helps identify where the real revision work should happen.</p><p>And that is the difference between a tool that rewrites and a tool that teaches revision.</p><h3>What Thanis can help students improve</h3><p>Thanis Academic is most useful when the student already has a draft and wants a serious revision pass before submitting it.</p><p>It can help with thesis clarity by identifying whether the central claim is specific enough, visible enough, and strong enough to carry the paper. It can help with argument structure by showing where the reasoning feels underdeveloped, where the sequence of ideas weakens, or where a paragraph does not clearly support the main point.</p><p>It can help with evidence by pointing out where sources need more explanation, where quotations or references appear disconnected from analysis, or where a claim may need stronger support. It can help with organization by identifying sections that feel misplaced, repetitive, or underconnected.</p><p>It can also help students make better revision decisions because it does not simply say, “This is bad” or “This is good.” The more useful question is: what should the student fix first?</p><p>A student with a rough draft does not need twenty vague suggestions. They need a revision path. They need to know whether the thesis is the first problem, whether the evidence is the first problem, whether the structure is the first problem, or whether the paper is mostly sound but needs polishing.</p><p>Thanis is designed to help organize that repair work.</p><h3>Why Thanis does not simply rewrite the paper</h3><p>There is a reason Thanis Academic does not position itself as a paper generator.</p><p>The writing process is where academic learning happens.</p><p>When a student struggles to form a thesis, that struggle is not wasted effort. When they revise a paragraph three times to make the logic cleaner, that is not inefficiency. When they realize that a source does not actually support the claim they wanted to make, that is not a failure of productivity. That is learning doing its quiet, difficult work.</p><p>Generative tools can easily short-circuit that process. They can produce fluent text so quickly that the student may stop asking whether they understand it. They can also introduce risks that are not always visible at first glance: factual drift, invented details, generic phrasing, tone changes, repeated structures, and a gradual loss of the student’s own voice.</p><p>Those risks are not imaginary. Stanford HAI has documented serious hallucination concerns in legal AI benchmarking, showing how fluent systems can still produce unreliable answers in high-stakes settings. Research on AI-assisted writing has also raised concerns about voice dilution and homogenization, where model suggestions can make different writers’ work sound more alike.</p><p>That matters in school because the goal is not merely to produce something that looks acceptable.</p><p>The goal is to produce work the student understands, can defend, and can improve from.</p><p>Thanis is built around that principle.</p><h3>The responsible student workflow</h3><p>The best way to use Thanis Academic is straightforward.</p><p>Start with the assignment. Read the prompt. Understand what is being asked. Notice whether the instructor wants analysis, argument, reflection, research, comparison, or application.</p><p>Then write a real draft. It does not have to be perfect. It can be awkward, uneven, too long, too short, or structurally messy. What matters is that the draft contains the student’s own thinking.</p><p>After that, start a project in Thanis Academic. Select the project type. Select the discipline. Add the assignment prompt or upload the rubric if one is available. Then submit the draft for review.</p><p>Once Thanis returns feedback, the student should not treat it as a command. They should treat it as a careful reading. Some comments may be immediately useful. Some may need judgment. Some may point to a deeper issue the student had not noticed yet.</p><p>Then the student revises the work themselves.</p><p>That final part is not optional. It is the line between responsible assistance and replacement.</p><p>Thanis can sharpen the reading. It does not take over the responsibility.</p><p>The student remains responsible for authorship, sources, citations, factual accuracy, academic integrity, and final submission choices.</p><h3>What Thanis does not replace</h3><p>Thanis Academic does not replace the instructor’s judgment.</p><blockquote>It does not guarantee a grade.</blockquote><blockquote>It does not verify every factual claim.</blockquote><blockquote>It does not certify that citations are correct.</blockquote><blockquote>It does not know every classroom policy.</blockquote><p>It does not turn an unfinished idea into a finished academic argument without the student doing the work.</p><p>That boundary should make students and instructors more comfortable, not less. The point is not to smuggle AI into the writing process as an invisible ghostwriter. The point is to make AI support more transparent, more structured, and more responsible.</p><p>Thanis Academic gives students a way to use AI without handing over the paper.</p><p>That is the practical middle ground many classrooms now need.</p><h3>Why this matters now</h3><p>Academic AI use is already here, but the habits around it are still unstable.</p><p>Students are experimenting. Faculty are adapting. Institutions are revising policies. Detection tools remain imperfect. Rubrics are being rewritten. Everyone is trying to figure out where the real boundary belongs.</p><blockquote>Thanis Academic is one answer to that problem.</blockquote><blockquote>It does not ask students to pretend AI does not exist.</blockquote><blockquote>It also does not encourage them to let AI become the author.</blockquote><p>Instead, it gives students a structured way to use AI as an academic reader, revision guide, and rubric-aware feedback system.</p><blockquote>That is a healthier model.</blockquote><blockquote>It keeps the student inside the work.</blockquote><blockquote>It keeps the instructor’s authority intact.</blockquote><blockquote>It keeps the writing process visible.</blockquote><p>And most importantly, it helps students become better writers over time, not just faster producers of acceptable-looking text.</p><p>Because the future of academic writing should not be a contest between prohibition and automation.</p><p>It should be a better workflow.</p><p>One where students can use advanced tools without losing their own judgment, their own voice, or their responsibility for the work they submit.</p><p>That is what Thanis Academic is trying to build.</p><h3>Final thoughts</h3><p>Students are already using AI, and that reality is not going away. The real question now is whether we give them tools that quietly replace their thinking, or tools that help them understand their own work more clearly.</p><p>That is the line Thanis Academic is built around.</p><p>It is not here to write the paper, flatten the student’s voice, or turn every assignment into the same polished, lifeless submission. It is here to help students slow down at the right moments, see the weak points in their draft, understand how their work aligns with the assignment, and make stronger revision choices before they submit.</p><blockquote>For instructors, that matters because the writing process remains visible.</blockquote><blockquote>For students, it matters because the work still belongs to them.</blockquote><p>And for anyone trying to navigate this strange academic moment, where AI is everywhere but trust is still fragile, Thanis offers a different path: not prohibition, not automation, but guided revision with responsibility still intact.</p><p>If you are a student trying to strengthen a paper, a parent trying to understand what responsible AI use can look like, or an instructor curious about tools that support writing without replacing the writer, Thanis Academic is worth exploring.</p><blockquote>Bring a real draft.</blockquote><blockquote>Bring the assignment.</blockquote><blockquote>Bring the rubric if you have one.</blockquote><p>Thanis will help you read the work more clearly, but the final voice, the final judgment, and the final submission still remain yours.</p><p>Try it free <a href="http://www.thanis.ai/">Thanis.ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=95ff5fa36dcf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thanis-insights/thanis-academic-a-student-writing-tool-built-for-revision-rubrics-and-responsible-ai-use-95ff5fa36dcf">Thanis Academic: A Student Writing Tool Built for Revision, Rubrics, and Responsible AI Use</a> was originally published in <a href="https://medium.com/thanis-insights">Thanis Insights</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Thanis, Explained: A Writing Tool That Doesn’t Rewrite Your Work]]></title>
            <link>https://medium.com/thanis-insights/thanis-explained-a-writing-tool-that-doesnt-rewrite-your-work-0abe1750d48f?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/0abe1750d48f</guid>
            <category><![CDATA[ai-writing-tool]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Mon, 27 Apr 2026 18:02:51 GMT</pubDate>
            <atom:updated>2026-04-27T18:02:51.270Z</atom:updated>
            <content:encoded><![CDATA[<p>Why the future of writing might not be faster output, but deeper understanding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OapKe75ghvHAsFBAKi-lsQ.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@roaddream?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Patrick Boucher</a> on <a href="https://unsplash.com/photos/woman-in-black-shirt-sitting-in-front-of-computer-b6XtaQNJKYU?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>We’ve found ourselves in a strange place with writing, one that didn’t arrive all at once but instead crept in gradually, almost quietly, until it became the default way of working without most of us ever really stopping to question how much had actually changed beneath our feet.</p><p>On one side, you still hear the familiar warnings, often delivered with a tone that feels more protective than restrictive, as if someone is trying to preserve something fragile that might not come back once it’s gone:</p><blockquote><em>“Don’t use AI. It’s cheating.”</em></blockquote><p>And then, just as confidently, usually from a completely different voice, sometimes in the very same environment:</p><blockquote><em>“You should be using AI. It’s the future.”</em></blockquote><p>And somewhere in between those two statements sits a student, or a writer, or really anyone trying to put something meaningful into words, staring at a blank page with a kind of quiet hesitation, not because they don’t know what to say, but because they’re no longer entirely sure how they’re supposed to say it.</p><p>In a series of conversations I had with academic professors at the University of Florida, people who are not theorizing about artificial intelligence from a distance but are actively working to integrate it into real classrooms with real students, I kept hearing a pattern emerge that felt less like a technical problem and more like a human one.</p><p>Students aren’t rejecting AI because they think it’s useless, and they’re not avoiding it because they don’t see the potential. In many cases, they’re hesitant because they’ve spent years being told not to trust it, not to rely on it, to treat it as something that exists just outside the boundaries of acceptable work, and then almost overnight, that message shifts.</p><p>Now they’re told to use it.<br>To explore it.<br>To incorporate it into their thinking.</p><p>And that kind of reversal doesn’t land cleanly, especially when no one has really taken the time to show them what thoughtful, responsible use actually looks like in practice.</p><p>Some students lean in immediately, curious and energized, eager to see what the tool can do and how far it can take them. Others hold back, cautious in a way that feels less like resistance and more like uncertainty, unsure where the line exists between assistance and replacement. And many, probably most, sit somewhere in between, using it just enough to get through an assignment, but not enough to understand it deeply.</p><p>And that middle ground, where usage exists without clarity, is where things begin to break down.</p><h3>The real problem isn’t AI</h3><p>It’s how we’ve started using it.</p><p>We didn’t just introduce a new tool into the writing process. We quietly reshaped the process itself, removing steps that once felt inconvenient but were actually doing more work than we realized.</p><p>Writing used to feel slower, sometimes frustratingly so, but there was a rhythm to it:</p><pre>idea → struggle → draft → revise → understand</pre><p>Now, in many cases, it looks more like this:</p><pre>idea → prompt → output → tweak</pre><p>And on the surface, that feels like progress. It’s faster, cleaner, more efficient, and in a world that rewards speed, it’s hard to argue against that. But something subtle disappears in that transition, something that doesn’t make noise when it leaves.</p><p>Friction.</p><p>And friction, as it turns out, is where thinking actually happens.</p><h3>When the answer comes too easily</h3><p>One of the most interesting things that came up in those classroom discussions wasn’t whether AI was useful, because it clearly is, but what happens when the answer shows up too quickly, too confidently, and with just enough polish to make you stop questioning it.</p><p>Students would input complex scenarios, sometimes clinical, sometimes analytical, and the model would return something structured, something coherent, something that looked and felt right.</p><p>And then everything would pause.</p><p>Not because the answer was perfect, but because the student wasn’t entirely sure whether they actually understood it.</p><p>That’s where the real teaching began.</p><p>Not in generating the answer, but in challenging it.</p><p>In asking:</p><ul><li>Does this actually make sense?</li><li>Would I stand behind this?</li><li>What might be missing here?</li></ul><p>Because as powerful as these models are, they are not infallible. They can produce outdated information, incomplete reasoning, and occasionally fabricate details that appear plausible until you take the time to look closely.</p><p>Which leads to a simple but critical mindset:</p><p>Trust, but verify.</p><p>Not as a warning, but as a discipline.</p><h3>The quiet trade we’re making</h3><p>What we’ve gained in speed, we may be losing in depth.</p><p>AI tools are incredibly effective at producing output, at taking a prompt and transforming it into something readable, structured, and often impressively polished in a matter of seconds. But writing has never just been about producing output.</p><p>It’s about forming a thought, testing it, refining it, and ultimately understanding it.</p><p>When the tool gives you the answer before you’ve fully worked through the question, your role begins to shift in a way that’s easy to miss. You move from being the author of the idea to being the editor of the output, and while that might feel efficient, it changes the nature of the work entirely.</p><p>You stop wrestling with the idea.</p><p>You stop questioning your own assumptions.</p><p>You start optimizing something that was handed to you.</p><p>And that’s a very different skill.</p><h3>What educators are actually trying to teach</h3><p>The goal in education has never been to simply finish the assignment.</p><p>It’s to understand what you wrote.</p><p>That’s why so many educators find themselves in this tension. They see the opportunity that AI presents, the ability to enhance learning, to improve efficiency, to introduce new ways of engaging with complex material. But they also see the risk, the potential for students to bypass the very process that leads to understanding.</p><p>They’re not trying to stop progress.</p><p>They’re trying to preserve the part of learning that actually matters.</p><h3>The gap no one is solving</h3><p>Most tools in this space are built to generate, rewrite, and optimize output. And for certain use cases, that’s incredibly useful.</p><p>But that wasn’t the problem I was trying to solve.</p><p>I didn’t want something to write for me.</p><p>I wanted something that could help me understand my own writing better, something that could point out where my thinking broke down, where clarity slipped, where structure weakened the idea, without stepping in and replacing the work entirely.</p><p>That’s a completely different category.</p><h3>What Thanis actually is</h3><p>That gap is what led to <a href="https://www.thanis.ai/what-is-thanis">Thanis</a>.</p><p>Thanis is a writing feedback system designed to help you improve what you’ve already written, not by generating new text, but by analyzing your draft and returning structured insights that allow you to refine it yourself.</p><p>It looks at your writing and surfaces things like where clarity begins to drift, where tone shifts unintentionally, where structure weakens your argument, and where ideas don’t fully connect.</p><p>And then it stops.</p><p>It doesn’t rewrite the paragraph.<br>It doesn’t “fix” the sentence for you.</p><p>It gives you the idea, and then you decide what to do with it.</p><h3>What it doesn’t do</h3><p>Thanis does not generate full drafts.<br>It does not automatically rewrite your work.<br>It does not attempt to replace your voice with something more polished.</p><p>Because the goal isn’t to produce something that looks better at a glance.</p><p>The goal is to help you become someone who writes better over time.</p><h3>Final thoughts….</h3><p>If the goal is to write faster, we’re already there.</p><p>We solved that.</p><p>But if the goal is to become better writers, to actually understand what we’re saying, to build clarity over time instead of outsourcing it entirely, then we’re still figuring that part out.</p><p>And maybe the answer isn’t letting AI write for us.</p><p>Maybe it’s using it in a way that forces us to think more carefully, write more deliberately, and improve in ways that actually last.</p><p>If you want to see how Thanis works:</p><p><a href="https://www.thanis.ai/what-is-thanis">https://www.thanis.ai/what-is-thanis</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0abe1750d48f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thanis-insights/thanis-explained-a-writing-tool-that-doesnt-rewrite-your-work-0abe1750d48f">Thanis, Explained: A Writing Tool That Doesn’t Rewrite Your Work</a> was originally published in <a href="https://medium.com/thanis-insights">Thanis Insights</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Writing “Proof Gap” is About to Break the Industry]]></title>
            <link>https://medium.com/write-a-catalyst/the-ai-writing-proof-gap-is-about-to-break-the-industry-d299e7fe8d8e?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/d299e7fe8d8e</guid>
            <category><![CDATA[ai-writing-tool]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[compliance]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Mon, 30 Mar 2026 06:33:37 GMT</pubDate>
            <atom:updated>2026-03-30T06:33:37.166Z</atom:updated>
            <content:encoded><![CDATA[<p>Right now, the academic and professional worlds are falling into a dangerous compliance trap. We are asking writers to “disclose” AI use, but we have zero way to prove <em>how</em> it was actually used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ziBq23XjaOlWjIYTYwWuBQ.png" /><figcaption>The EU AI Act is already structuring AI into enforceable categories — including transparency and monitoring obligations. What’s still missing is how those obligations are actually fulfilled at the level of writing.</figcaption></figure><p><strong>We have successfully created a “Cleanup Economy.”</strong> If you scroll through Medium or LinkedIn today, you’ll see an army of writers teaching other writers how to coax a machine into writing. We use multi-stage workflows, “humanizer” tools, and blacklists of banned words just to stop AI from sounding like a robot.</p><p>The result? The first draft is instant, but everything after that is damage control. We are spending 60 minutes “fixing” a 30-second draft just to make it usable.</p><p>But there is a much bigger problem looming than just “bad writing.” While we are busy arguing over prompts, the regulatory ground is shifting beneath our feet.</p><h3>The Regulatory Hammer: From Guidance to Enforcement</h3><p>I’ve spent 20 years building enterprise systems, and I recognize this pattern. It starts with a suggestion, moves to an expectation, and ends in enforcement. We are now entering the <strong>Enforcement Phase</strong>.</p><p>The EU AI Act is no longer a philosophical document; it’s an operational one. It mandates that AI-generated content be identifiable and traceable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Hg_Oik2VX_SWz6YXM8q2UQ.png" /><figcaption>What used to be optional is now enforced at submission. The system is starting to care how content is created.</figcaption></figure><p>At the same time, the academic world is slamming the door shut. FAccT 2026 and major journals now require formal “Usage Statements” at submission. If you can’t declare exactly how you used AI, you face an immediate desk rejection. Here is why every writer and editor should be worried: <strong>Disclosure is currently just a ritual.</strong> A checkbox that says “I used AI for grammar” is just a statement. It isn’t evidence. It isn’t verifiable. It doesn’t show what changed, when it changed, or <em>why</em> it changed. As publishers like Elsevier have made clear, you are still 100% accountable for the integrity of the work, even if the AI hallucinated the citations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IoMJLqo7RttZqvwPTpMDEA.png" /><figcaption>Disclosure does not reduce accountability. Authors are still fully responsible for the content — even when AI is involved. What’s missing is a way to prove how that content was created. Source elsevier.com</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vc_czej_3p_VepUrrcgMyw.png" /><figcaption>Even with AI involved, the responsibility remains entirely human. Accuracy, integrity, and authorship are still enforced — without any structured way to prove what actually happened. Source elsevier.com</figcaption></figure><blockquote>This is the <strong>Proof Gap</strong>. We are being asked to provide transparency, but the tools we use — ChatGPT, Claude, Gemini — are opaque. They give you a final output and leave you to sort out the mess. They generate text, but they don’t capture the <strong>sequence of decisions.</strong></blockquote><h3>The Transition: From Generation to Governance</h3><p>If we want to survive this shift, we have to stop asking chatbots to invent our stories and start using systems that protect our authorship.</p><p>Governance cannot be bolted on at the end of a document; it has to be embedded at the point of execution. This is exactly why I built <strong>Thanis</strong>.</p><p>Thanis doesn’t ask you to be a better prompt engineer. It doesn’t care about “humanizer” hacks. Instead, it operates inside the writing process itself. It treats writing as a sequence of reviewable decisions.</p><p>While a standard AI chatbot just throws more text at you, Thanis identifies specific “Editorial Breaches” — like the moment your setting shifts from a cabin to a hotel without explanation. It names the problem so <em>you</em> can fix it with intention.</p><h3>The New Standard of Authorship</h3><p>In 2026, the winners won’t be the people who can generate the most content. They will be the ones who can <strong>account</strong> for it.</p><p>The “Proof Gap” is closing. You can either stay in the Cleanup Economy, manually scrubbing generic AI text and hoping you don’t get flagged, or you can move to a governed workflow where your voice is backed by evidence.If You Want to Explore This Further</p><p>If this post resonates with you, there are two ways to go deeper:</p><ul><li>The <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5381798"><strong>THANIS whitepaper</strong> </a>breaks down the architecture behind this approach, including how containment, traceability, and decision-level provenance actually work in practice</li><li>The <a href="http://thanis.ai/"><strong>THANIS platform</strong></a> is already live and can be used today to run your own writing through this kind of governed workflow</li></ul><p>Bring a draft that matters to you, not a throwaway paragraph, not a test prompt, but something with real weight behind it. Run it through THANIS and study the response. What comes back is not just revision advice. It is a clearer record of what the writing is doing, where it holds, where it slips, and what can actually be defended.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d299e7fe8d8e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/write-a-catalyst/the-ai-writing-proof-gap-is-about-to-break-the-industry-d299e7fe8d8e">The AI Writing “Proof Gap” is About to Break the Industry</a> was originally published in <a href="https://medium.com/write-a-catalyst">Write A Catalyst</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Prompting. Start Writing.]]></title>
            <link>https://medium.com/write-a-catalyst/stop-prompting-start-writing-814b89b1c78f?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/814b89b1c78f</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[creativity]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Sat, 21 Mar 2026 03:40:19 GMT</pubDate>
            <atom:updated>2026-03-21T03:40:19.815Z</atom:updated>
            <content:encoded><![CDATA[<p>The endless cycle of ChatGPT “hacks” is killing your writing. Here’s how to break it.</p><figure><img alt="Frustrated writer sitting at his computer after hours of prompting." src="https://cdn-images-1.medium.com/max/1024/1*L6Zz238vwebJB5g61qq7xg.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@andemjnr?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Edemekong Andem</a> on <a href="https://unsplash.com/photos/person-with-hands-over-face-at-laptop-S32ulDfL5ZA?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>If you scroll through Medium on any given morning, you will inevitably drown in a sea of identical, desperate headlines.</p><p>“The 5 ChatGPT Prompts to 10x Your Content.”<br>“How to Stop AI from Sounding Robotic.”<br>“I Used ChatGPT Every Day for a Year and Forgot How to Write.”</p><p>There are tens of thousands of these articles clogging up our feeds, published by an army of newly minted “prompt engineers” all selling the same AI poison.</p><p>We have successfully birthed an entire digital economy of writers writing instructional guides on how to coax a machine into writing, so that other writers can avoid the actual act of writing. On Medium alone there are over 10,000+ published articles giving advice on how to use ChatGPT to help you write better.</p><p>And if you read closely, the advice is always the same.</p><p>Some writers preload the model with old samples so it can imitate their voice. Others build custom blacklists of suspicious words that make readers think, “this was written by AI.” Others use multi-stage workflows where the model drafts, critiques itself, rewrites under tighter constraints, and then gets edited again.</p><p>The details vary, but the structure is identical:</p><p><strong>The first draft is fast. Everything after that is damage control.</strong></p><p>A recent article by <a href="https://medium.com/u/8a486cb2d236">James Presbitero Jr.</a> arrives at the same conclusion from a different angle. After reviewing hundreds of AI-assisted articles, he found that “humanizer” tools often make writing harder to fix, not easier. The words change, but the underlying structure still feels synthetic. The surface improves. The core doesn’t.</p><p>Read <a href="https://medium.com/u/8a486cb2d236">James Presbitero Jr.</a> article titled <a href="https://medium.com/technical-excellence/ai-humanizer-tools-wont-save-your-writing-in-2026-here-s-what-actually-works-70787110067e">“AI Humanizer Tools Won’t Save Your Writing In 2026. Here’s What Actually Works.”</a></p><p>That matters, because it confirms something many writers already feel but haven’t fully named:</p><p>This is not a writing breakthrough. It is a cleanup economy.</p><h3>The industry is solving the wrong problem</h3><p>Over the last few days, I spent hours going through these workflows.</p><p>Not just skimming them. Actually running them. Testing the prompts. Following the steps that “advanced users” claim will finally make AI writing feel human.</p><p>And here’s what became obvious:</p><p>None of these systems are actually about writing better.</p><p>They are about <strong>fixing what the model keeps getting wrong</strong>.</p><p>That’s a very different problem.</p><p>Because if your workflow requires:</p><ul><li>training the model on your past writing</li><li>maintaining a blacklist of banned words</li><li>running diagnostic prompts</li><li>rewriting paragraph by paragraph under constraints</li></ul><p>…then you’re not using a writing tool.</p><p>You’re building a manual control system around a generator that doesn’t behave the way you need it to.</p><p>And that is where the time starts leaking.</p><h3>The false promise of speed</h3><p>AI writing feels fast because the first step is instant.</p><p>You type a prompt.<br>You get a draft.<br>It looks complete.</p><p>That’s the hook.</p><p>But writing is not the first draft.</p><p>Writing is everything that happens after:</p><ul><li>tightening weak sections</li><li>removing generic phrasing</li><li>restoring tone</li><li>aligning structure</li><li>making it sound like something a real person would actually say</li></ul><p>And this is where the illusion breaks.</p><p>Because the moment you try to turn that “instant draft” into something real, the process slows down dramatically.</p><p>You are no longer writing.</p><p>You are correcting.</p><h3>The prompting treadmill</h3><p>Take a simple example.</p><p>You want to write a real story. A family trip. It rained for two days. The kids got restless. There was one quiet moment by the fire that actually mattered.</p><p>You already know the story.</p><p>So you ask ChatGPT to write it.</p><p>It responds instantly.</p><p>At first glance, it looks good.</p><p>Then you read it.</p><p>But something feels off.</p><p>The rain becomes poetic instead of frustrating.<br> The kids become generic.<br> The moment by the fire turns into a cliché about gratitude or togetherness.</p><p>Nothing is obviously broken.</p><p>But none of it feels like yours.</p><p>And so you try to fix it.</p><p>And, You prompt again:</p><ul><li>“Make it more natural.”</li><li>“Use my tone.”</li><li>“Avoid clichés.”</li><li>“Don’t sound like AI.”</li></ul><p>You paste in writing samples.<br>You add constraints.<br>You blacklist words.<br>You ask it to critique itself.<br>You rewrite sections.<br>You try again and again..</p><p>And this is where the loop begins.</p><p>You are no longer writing the story.</p><p>You are trying to control the output of a system that keeps drifting away from what you meant.</p><p>I gave ChatGPT a simple prompt. Nothing complicated. Just a short description of a family trip and asked it to help shape the story and make it sound more human.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wryphlRJa9ypAXiOoqEQ3w.png" /><figcaption>Screenshot from my ChatGPT draft</figcaption></figure><p>Here is what ChatGPT gave me</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*af6W92YzoYV0vEG7cpmNKQ.png" /></figure><p>At first glance, it feels fine..</p><p>The sentences flow. The structure holds. It even reads like something you might expect from a personal essay. If you weren’t paying attention, you could almost convince yourself this is usable.</p><p>But then i read it again.</p><p>The problem is… This isn’t my story.</p><p>The rain didn’t feel like a “soothing backdrop.”<br>The kids weren’t just “filled with energy.”<br>And that ending? That clean, reflective wrap-up about everything coming together?</p><p>None of that ever happened.</p><p>What actually happened was messier. More uneven. More specific. There were moments that didn’t resolve cleanly. Tension that didn’t turn into a neat lesson. Small details that mattered more than the summary.</p><p>And that’s exactly what got erased.</p><p>The model didn’t misunderstand the facts.<br>It flattened them and didnt understand at all what that trip was about.</p><p>It took something lived and turned it into something <em>recognizable</em>.</p><p>Not real at all..</p><h3>Why this is worse than a bad draft</h3><p>Now this isn&#39;t a problem really.. Because a bad draft is easy to fix.</p><p>But you can see what’s broken. You can rebuild it. You can feel where it goes wrong.</p><p>However, This is different.</p><p>This is polished enough to pass.<br>Its generic enough to blur.<br>Clean enough to trick you into thinking you’re close.</p><p>But you’re not close.</p><p>You’re now stuck in a strange position:</p><p>You didn’t write it.<br>But now you have to fix it.</p><p>And fixing it means:</p><ul><li>stripping out the fake tone</li><li>removing the generic phrasing</li><li>rebuilding the moments that mattered</li><li>re-inserting your actual voice</li></ul><p>At that point, you’re not saving time.</p><p>You’re doing <strong>two jobs instead of one</strong>.</p><h3>Step off the loop</h3><p>At some point, you have to stop and ask a different question.</p><p>Not: “How do I prompt this better?”</p><p>But: <strong>“Why am I doing all of this just to write something I already know?”</strong></p><p>Because that’s the real issue.</p><p>You’re not struggling to come up with the story. You’re struggling to force a system to stop rewriting it into something generic.</p><p>And the more you try to fix it, the more time you spend managing output instead of shaping your own work.</p><p>Here is a scary statistic : Using ChatGPT For a <strong>500-word story</strong>:</p><ul><li><strong>Initial draft:</strong> 30–45 seconds</li><li><strong>Prompting + iteration loop:</strong> 25-60 minutes</li><li><strong>Total time to something “usable”:</strong><br> = <strong>~30–90 minutes</strong></li></ul><p>That’s the moment where the workflow breaks.</p><p>Not because the model failed.</p><p>Because the tool you’re using is built for generation, not for writing.</p><p>And once you see that clearly, the next step becomes obvious.</p><p>Stop asking a chatbot to invent your story for you.</p><p>Start with the story itself.</p><p>Write the draft in your own words, with your own rhythm, your own details, and your own sense of what actually mattered. Then bring that draft into a tool that was built to help writers refine what is already theirs.</p><h3>That is the difference with Thanis.</h3><p><a href="http://thanis.ai/">Thanis</a> does not ask you to become a better prompt engineer. It does not make you preload old samples, maintain a blacklist of suspicious words, or keep negotiating with a model until the output feels almost human. It starts from a simpler premise: the writer should remain the writer.</p><p>Instead of generating a synthetic version of your memory and forcing you to clean it up afterward, Thanis works with the draft you actually wrote. It helps you see where the structure weakens, where the tone slips, where the phrasing goes flat, and where the story begins to drift away from its center.</p><p>That changes the entire experience.</p><p>You are no longer trying to recover your voice from a system that replaced it.</p><p>You are strengthening the voice that was already on the page.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pNwjb3korynIG8uo5EaI1g.png" /><figcaption>No prompts. No rewrites. Just your draft and what actually needs attention.</figcaption></figure><p>What makes this useful is that the guidance is not vague.</p><p>It is not telling me, in some abstract way, to “improve the story” or “make it stronger.” It is breaking the draft down into dimensions that actually matter while writing: clarity, structure, mechanics, and risk.</p><p>That matters because not every problem in a piece is the same kind of problem.</p><p>Some issues are structural. A scene arrives too early, a middle section feels thin, a resolution lands before the pressure has fully built.</p><p>Some are mechanical. The pacing flattens, sentence lengths become too uniform, or the emotional insight gets buried inside one long clause.</p><p>And some are clarity problems, which are often the most dangerous because they are the easiest to miss while drafting.</p><p>Clarity, in this context, is not about writing “simply.” It is about whether the reader can stay oriented inside the piece. Whether the setting holds.</p><p>Whether the movement from one detail to the next makes sense. Whether the story is carrying the reader forward, or quietly making them stop and ask the wrong question.</p><p>That is where <a href="http://thanis.ai/">Thanis</a> starts to feel different from a general AI tool.</p><p>A typical AI chatbot will usually give you more text. It will paraphrase, expand, summarize, or rewrite. Thanis does something more valuable here: it identifies the specific kind of weakness that is interfering with the story, and it names it without taking the draft away from you.</p><p>In this case, the clearest example is clarity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/942/1*dwje8P25qQ07M5nA9xVStw.png" /><figcaption>This is the kind of issue you don’t notice while drafting, but the reader feels immediately.</figcaption></figure><p>The issue here is not grammar.</p><p>It is not even tone.</p><p>It is a break in reader orientation.</p><p>I wrote the story as if it all lived in one place: the cabin, the rain, the kids stuck inside, the fire at night. But in the middle of the draft, I casually introduced a hotel. Just one line. Then I moved on.</p><p>As the writer, I barely noticed it.</p><p>But as the reader, you feel it immediately.</p><p>Now the mind is no longer inside the scene. It has stepped back and started asking a logistical question: <em>Wait, are we at the cabin or the hotel?</em> The moment that happens, the story loses pressure.</p><p>That is what Thanis is catching.</p><p>Not with a rewrite.<br>Not with a generic suggestion.<br>Not by trying to “humanize” the paragraph.</p><p>It is showing me exactly where the piece stops being fully legible from the reader’s point of view.</p><p>And once that becomes visible, the difference is enormous.</p><p>With ChatGPT, I would still be in the loop, prompting for a better version and hoping the model happened to resolve the inconsistency on its own.</p><p>With Thanis, the problem is named directly. I can cut the line, clarify the setting, or reframe the piece around the correct location. The choice stays with me. But now I’m making an editorial decision, not guessing inside a prompt box.</p><p>That is the real difference.</p><p>One workflow generates more text and leaves you to sort it out.</p><p>The other helps you see what is actually wrong so you can fix it with intention.</p><h3>Thanis is not giving you more text. It is giving you better editorial visibility.</h3><p>Clarity is one example, but it is not the only one.</p><p>What makes Thanis useful is that it breaks revision into the kinds of problems writers actually run into while drafting.</p><p>Sometimes the issue is clarity. A detail appears without context. A location shifts. A sentence introduces confusion the writer no longer notices because they are too close to the material.</p><p>Sometimes the issue is structure. The overall shape of the piece works, but the middle is carrying too little weight, or a paragraph generalizes where it should become more specific. The result is a story that technically moves, but does not fully build pressure before the ending arrives.</p><p>Sometimes the issue is mechanics. The tone is natural, but the pacing stays too flat. Sentence lengths become too uniform. The sentence that should land the hardest gets buried under too many clauses and arrives without enough room to breathe.</p><p>Those are different kinds of weaknesses, and they need different kinds of attention.</p><p>That is what Thanis does well. It does not throw more generated text at the problem. It helps the writer identify what kind of problem they are dealing with in the first place.</p><p>That changes revision from a vague feeling of “something is off” into a more useful question:</p><p>What exactly is off, and what kind of edit does this piece actually need?</p><p>That is the difference between a system that generates language and a tool that helps writers improve it.</p><h3>The difference becomes obvious when you step back</h3><p>At this point, it’s not really about whether one tool is “better” than another.</p><p>It’s about what the tool is actually designed to do.</p><p>Up to now, we’ve been trying to force general-purpose AI systems to behave like writing tools. We prompt them, guide them, correct them, and iterate until the output gets close enough to use.</p><p>With Thanis, the direction flips.</p><p>You don’t start with generated text and try to fix it.<br>You start with your own writing and improve it.</p><p>That sounds like a small shift. It isn’t.</p><p>It changes where your time goes, what kind of work you’re doing, and how much control you actually have over the final piece.</p><p>If you strip everything down, the difference looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HyapzQgOF0DvG9WVYjVaFQ.png" /><figcaption>This table shows. One generates text. The other helps you finish it.</figcaption></figure><p>I do not think writers need more hacks. Writers dont need to learn better prompts..</p><p>They do not need another thread about “secret prompts,” another humanizer tool, or another workflow designed to make generated text feel slightly less artificial. That whole ecosystem is built around a bad assumption: that the answer is always one more layer of prompting.</p><p>I think the better answer is simpler.</p><p>Write the piece.<br>Keep your voice.<br>See the weak spots clearly.<br>Revise with intention.<br>Finish stronger than you started.</p><p>That is what Thanis is for.</p><p>If you’re tired of prompting your way toward something usable, bring your next draft into Thanis and see what happens when the tool is finally working with the writer instead of replacing one.</p><p>If that sounds like a better way to work, test it for yourself.</p><p>Take your next story, your next essay, your next memory you’ve been trying to pin to the page, and run it through Thanis instead of a prompt box.</p><p>You may find that what you needed was never more generation.</p><p><a href="http://www.thanis.ai/">Thanis.AI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=814b89b1c78f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/write-a-catalyst/stop-prompting-start-writing-814b89b1c78f">Stop Prompting. Start Writing.</a> was originally published in <a href="https://medium.com/write-a-catalyst">Write A Catalyst</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Language Control Plane: Why We are Scaling Risk Instead of Intelligence]]></title>
            <link>https://medium.com/write-a-catalyst/the-language-control-plane-why-we-are-scaling-risk-instead-of-intelligence-91fa4921c3bf?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/91fa4921c3bf</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai-writing-tool]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Wed, 18 Mar 2026 03:35:57 GMT</pubDate>
            <atom:updated>2026-03-18T03:35:57.861Z</atom:updated>
            <content:encoded><![CDATA[<p>Why scaling Generative AI without a control plane is just high-speed technical debt.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UnyztcARDx9NLcuNJ5nKuA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@timwitzdam?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Tim Witzdam</a> on <a href="https://unsplash.com/photos/phone-screen-asks-what-can-i-help-with-yWyBU5v3FLI?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Over the past year, I have watched the enterprise world undergo the fastest infrastructure shift in history. We have moved from “What is an LLM?” to “How do I deploy 1,000 of them?” in record time.</p><p>But as a Solutions Architect, I’ve noticed a disturbing pattern emerging across engineering teams, marketing departments, and academic institutions alike. We have essentially optimized the <strong>Data Plane</strong> of language (generation), but we have completely ignored the <strong>Control Plane</strong> (governance).</p><p>The result? We aren’t just scaling content; we are scaling <strong>unverified risk.</strong></p><h3>The “Smoothness” Trap</h3><p>The industry is currently addicted to the “smoothness” of synthetic output. A technical whitepaper that used to take three weeks now takes three seconds. Now hear me out… On the surface, it looks like a massive productivity gain. However, if you look at the “Operational Reality,” the contrast is staggering. Recent reports from firms like Forrester and RAND Corp highlight a harsh truth: up to <strong>80–95% of AI projects are failing to deliver measurable ROI</strong>.</p><p>This isn’t because the technology is “broken.” It’s because the work hasn’t disappeared — it has merely shifted downstream. We have traded the <em>difficulty of creation</em> for the <em>crushing burden of verification</em>. Teams are spending more time manually fact-checking, de-biasing, and reconstructing hallucinated logic than they would have spent writing from scratch.</p><p>We are accumulating “Risk Debt” — decisions made without audit trails that surface only during a compliance failure or a security breach.</p><h3>The Shadow “Decision Layer”</h3><p>Generative AI has introduced a distinct, sharper layer into our organizations: it produces language that is treated as truth and nudges workflows toward specific outcomes. Even when the AI does not execute the final action, it provides the “first draft” of the decision. In high-velocity environments, <strong>the first draft usually wins.</strong> Consider a fintech startup using a standard RAG stack for support. A model retrieves a past “one-time exception” refund for a VIP and confidently recommends: <em>“Approve refund based on precedence”</em>. The model doesn’t understand the nuance of an “exception” versus “policy”; it just sees a semantic match. The support rep, incentivized by speed, hits approve. This isn’t a hallucination; it’s a <strong>context collapse</strong>. It is financial leakage scaled by automation.</p><h3>Why “AI Centers of Excellence” are Failing</h3><p>The default reaction to this risk is to form a centralized “AI Center of Excellence” (CoE) or appoint an “AI Czar”. In my experience, this group quickly becomes the “Department of No”. They try to review every prompt and model, creating a massive bottleneck.</p><p>The inevitable result is <strong>Shadow AI</strong>. Frustrated teams bypass the CoE to use personal accounts or unapproved APIs, creating more hidden risk. You cannot govern the future of language with a “Terms of Use” PDF and a bottleneck.</p><h3>The Architecture of Language Governance</h3><p>We need to stop asking “Did a robot write this?” and start asking “Does this output meet our semantic and logical standards?” To do that, we must move toward <strong>Distributed Decision Ownership</strong>.</p><p>Mature, AI-Centric organizations are shifting toward “AI Pods” — assigning <strong>Decision Owners</strong> within specific business lines. These aren’t model integrators; they are the people accountable for the business intent, the tolerance for error, and the “kill switch” mechanisms.</p><ul><li><strong>The Security Team</strong> sets the guardrails (e.g., “No PII in the prompt”).</li><li><strong>The Decision Owner</strong> sets the logic (e.g., “Refunds over $50 require human approval”).</li></ul><p>This treats governance as an engineering constraint, not a compliance retrofit.</p><h3>Beyond the Magic Box</h3><p>The industry is bifurcating. On one side are organizations treating AI as a “magic box” for productivity, quietly accumulating hidden risks. On the other are AI-Centric organizations treating AI as a decision engine, wrapping it in the same rigorous controls they use for financial ledgers.</p><p>The differentiator for the next decade will not be who has the largest context window. It will be who can run automated decisions with discipline. You cannot solve this by adding more prompt engineers. The scaling constraint is not prompts; it is accountability.</p><p>If you want to move beyond pilots and build true enterprise capability, you must assign Decision Owners. You must stop outsourcing your authority to machines that know how to sound successful, but do not know how to <em>be</em> successful.</p><p>It’s time to stop prompting and start governing.</p><p><strong>Author’s Note:</strong> This article was developed using <a href="http://thanis.ai/"><strong>Thanis</strong></a>, a platform I’m building to bridge the “Trust Gap” in AI-augmented writing. By applying the <strong>Decision Owner</strong> framework to my own creative process, Thanis ensures that every claim is verified and every edit remains under human accountability. We are currently in early access for teams looking to formalize their Language Governance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=91fa4921c3bf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/write-a-catalyst/the-language-control-plane-why-we-are-scaling-risk-instead-of-intelligence-91fa4921c3bf">The Language Control Plane: Why We are Scaling Risk Instead of Intelligence</a> was originally published in <a href="https://medium.com/write-a-catalyst">Write A Catalyst</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Uncalculated Variable: A Reflection on Luck]]></title>
            <link>https://medium.com/shaped-souls/the-uncalculated-variable-a-reflection-on-luck-69ee5bbf59cb?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/69ee5bbf59cb</guid>
            <category><![CDATA[spirituality]]></category>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[shaped-souls]]></category>
            <category><![CDATA[life-lessons]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 19:43:20 GMT</pubDate>
            <atom:updated>2026-03-16T19:43:20.466Z</atom:updated>
            <content:encoded><![CDATA[<h4>Does Luck really exist? Writers Challenge</h4><figure><img alt="Beautiful rainbow over open fields" src="https://cdn-images-1.medium.com/max/1024/1*0y8u1iTO-8dxK8sXne5Dfg.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@briantaylor?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Brian Taylor</a> on <a href="https://unsplash.com/photos/black-and-white-abstract-painting-xmfFJB_GihI?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>I’ve never been the type of person to assume that luck was perpetually on my side. In my world, I’ve always viewed life through the lens of outcomes — there are good outcomes and bad ones. Whether “luck” is the hidden hand guiding us toward a better result is something I’m still not entirely sure of, but looking back at the architecture of my life, I can see the moments where the math simply didn’t add up.</p><p>I think of being seventeen years old, sitting at a red light, lost in the music. I wasn’t doing anything wrong; I was just there. And then, at full speed, another car slammed into mine. By every logical metric, that should have been a tragedy of broken bones and hospital stays. The car was totaled — a metal shell crushed by impact, yet I walked away without a single serious injury.</p><p>Was that luck? Was it a higher power? Or was it just an extra layer of protection from a universe granting an outcome I hadn’t calculated for? I like to think that in those moments, whether we call it luck or divine intervention, something steps in to balance the scales when we are at our most vulnerable.</p><p>But luck doesn’t just show up in the wreckage of a car accident. For me, the most profound “luck” has always been found in the people who walk into my life unexpectedly, right when I need them the most.</p><p>I think about those casual encounters that suddenly spiral into a lifetime of friendship. I remember being invited to a large group dinner by chance, a friend had simply asked if I wanted to tag along. I ended up at a long table of fifteen people, and I started chatting with the person sitting next to me. His name was Frank. He was young, energetic, and when he told me he was a doctor, I’ll admit I didn’t quite believe him at first. He seemed too young for the gravity of that title.</p><p>But we kept talking. Before the night ended, he handed me his card, and it turned out he wasn’t just a doctor; he ran his own practice. That chance seating arrangement, facilitated by a last-minute invitation, turned into a friendship that has spanned over fifteen years. Frank became my personal physician, and today, he is the doctor I trust with my twins.</p><p>I have no idea what alignment of stars or coincidence brought him into my orbit, or me into his, but I know that my life was fundamentally bettered by it. These people influence us in ways we never thought possible, opening doors — and often opening our minds — to versions of ourselves we hadn’t yet met. It makes you realize that while you can build the most secure systems in the world, the most valuable assets are the ones you never saw coming.</p><p>So, do I believe in luck? Yes, I believe there is a certain element of it woven into the universe. I don’t know if some people are “blessed” with it more than others, or if the stories of those “cursed” with bad luck are true. But I prefer to believe in the opposite. I like to think that every once in a while, the universe simply opens up, shines a light upon you, and grants you that unexpected wish.</p><p>It does seem that in a world of calculated risks and planned trajectories, luck is the beautiful, unmapped variable that reminds us we aren’t traveling this road alone.</p><p>A huge thank you to <a href="https://medium.com/u/6da965be0b80">Sherene Jensen</a> and the <strong>Shaped Souls</strong> publication for providing this space for self-exploration. This prompt forced me to contemplate the moments in my own life, from totaled cars to life-changing friendships where the math simply didn’t add up. It’s a powerful reminder that while we try to architect our lives with precision, there is always room for the universe to shine its luck upon us when we least expect it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69ee5bbf59cb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/shaped-souls/the-uncalculated-variable-a-reflection-on-luck-69ee5bbf59cb">The Uncalculated Variable: A Reflection on Luck</a> was originally published in <a href="https://medium.com/shaped-souls">Shaped Souls</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Built Thanis Deep Trace on AWS to Detect Writing Patterns Across a Living Archive]]></title>
            <link>https://medium.com/thanis-insights/how-i-built-thanis-deep-trace-on-aws-to-detect-writing-patterns-across-a-living-archive-dc2a9722abbd?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/dc2a9722abbd</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[software-architecture]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Fri, 13 Mar 2026 14:14:34 GMT</pubDate>
            <atom:updated>2026-03-15T16:34:51.846Z</atom:updated>
            <content:encoded><![CDATA[<p>Technical Deep Dive and Architecture Overview</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oLbJQQGkK0TMVedT9mWeMQ.png" /></figure><p>Most writing tools are built to react to the draft in front of you.</p><p>That is useful up to a point. A single-pass system can flag a muddy paragraph, a sentence that drifts, or an ending that collapses under its own summary. But it still misses the more interesting problem, which is usually not the sentence. It is the recurrence.</p><p>Writers repeat structural decisions. We repeat kinds of endings, sentence shapes, image families, emotional framing habits, explanatory reflexes, and pacing defaults. A draft can hide that because the pattern feels natural from the inside. The archive does not hide it nearly as well.</p><p>That is why I built Deep Trace.</p><p>Deep Trace is the archive-aware pattern analysis layer inside Thanis. Standard Thanis analysis reads one draft and tells you how stable, clear, and coherent that draft is. Deep Trace adds a second layer: it compares the current piece against earlier saved work and looks for repeated writing habits that a single-draft read cannot reliably surface.</p><p>The hard part was not getting a model to talk about writing. That part is easy. The hard part was building a system that could pull the right historical context, constrain the model tightly enough to avoid generic AI-writing-coach output, and return something precise enough to be useful to a real writer.</p><p>That required a different architecture than a normal “prompt in, answer out” workflow.</p><p>And this led to many days pulling my hair out :)</p><h3>The problem I was solving</h3><p>There were three constraints from the start.</p><p>First, Deep Trace could not just re-read the current document and pretend that was longitudinal analysis. It needed real historical proof from earlier saved work.</p><p>Second, it could not become a ghostwriter. The value was not that it could rewrite the sentence more cleanly. The value was that it could show the writer the recurring move beneath the sentence.</p><p>Third, it had to stay fast enough to use. A beautiful archive-aware system that takes three minutes to answer is still too expensive in attention.</p><p>So the design target became:</p><ul><li>use the current draft as context</li><li>retrieve prior writing from the archive</li><li>identify two recurring patterns</li><li>prove them with historical evidence</li><li>return a lean result in plain language</li><li>do it fast enough that a writer will actually click Run again</li></ul><p>That sounds tidy in bullet points. In practice, it became a containment problem.</p><h3>The AWS Architecture</h3><p>At a high level, the Deep Trace workflow looks like this:</p><ol><li>The writer opens a saved document in Studio.</li><li>They trigger Deep Trace from the frontend.</li><li>The backend route pulls the current document context plus historical excerpts from earlier saved work.</li><li>That bundle is sent to Amazon Bedrock for constrained pattern analysis.</li><li>The result is validated, normalized, persisted, and returned to the UI.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QVtCBrbaYSeJ8VMUY76OzA.png" /></figure><p>The key phrase there is <strong>historical excerpts</strong>. Deep Trace is not just “another pass” on the current draft. It is pulling prior material from the archive and forcing the model to support its pattern claims with evidence from earlier pieces.</p><p>That distinction is the feature.</p><h3>The storage layer</h3><p>The writing archive lives in a Postgres-backed persistence layer. Each saved piece is broken into chunks and stored with:</p><ul><li>user_id</li><li>doc_id</li><li>doc_title</li><li>content</li></ul><p>Deep Trace queries against that chunk store.</p><p>For the current document, it pulls a small evenly distributed sample of chunks. That gives the model enough of the live draft to understand where it is without spending tokens on the whole document.</p><p>For history, it deliberately excludes:</p><ul><li>the current doc_id</li><li>the current doc_title</li></ul><p>That second exclusion matters. Earlier versions of the same title can trick a model into thinking it has found a cross-document pattern when it has really just found nearby variants of the same piece. I wanted historical proof from genuinely earlier work, not another copy of the current problem wearing a different timestamp.</p><p>The historical retrieval query then does three things:</p><ul><li>selects a spread of earlier titles</li><li>limits how many versions per title it samples</li><li>limits how many chunks per document it sends forward</li></ul><p>That was one of the biggest design levers in the whole system. Too much history and Bedrock slows or fails. Too little history and the pattern analysis becomes shallow or fake. The answer turned out to be selective sampling, not brute-force inclusion.</p><h3>Why I chose Bedrock for this layer</h3><p>Deep Trace sits inside Thanis, which already runs on AWS. Using Amazon Bedrock let me keep the analysis path inside the same cloud boundary and avoid building a separate orchestration story around external model infrastructure.</p><p>Claude became the reasoning layer because it is strong at comparative reading and structured output when tightly constrained. That “when tightly constrained” part matters. If you let the model talk freely about writing, it will happily produce ten paragraphs of polished, generic commentary that sound useful and are almost impossible to trust.</p><p>So I did not let it talk freely.</p><h3>The real engineering work: output containment</h3><p>The first versions of Deep Trace had a familiar problem. The outputs sounded intelligent and were still not good enough.</p><p>The system would identify a pattern, but the language drifted. It got workshop-ish. It got padded. It got explanatory in the wrong way. Sometimes it was too flattering. Sometimes it rewrote the writer instead of diagnosing the page. Sometimes it returned structure the frontend could not reliably render.</p><p>So the core work became less about “prompting for better feedback” and more about <strong>containing the output shape</strong>.</p><p>I ended up using three layers of control.</p><h3>1. Tight schema design</h3><p>The route does not ask for free-form writing advice. It asks for a fixed structure:</p><ul><li>3 short strengths</li><li>exactly 2 patterns</li><li>each pattern must include:</li><li>category</li><li>pattern_key</li><li>pattern_title</li><li>the_pattern.diagnosis</li><li>the_pattern.frequency</li><li>the_pattern.reader_cost</li><li>the_evidence</li><li>the_pivot</li></ul><p>That shape matters because it narrows the model’s room to wander.</p><h3>2. Prompt-level voice containment</h3><p>The prompt tells the model explicitly what it is not allowed to become:</p><ul><li>not a cheerleader</li><li>not a workshop instructor</li><li>not a therapist</li><li>not a consultant</li><li>not a generic AI writing coach</li></ul><p>It also forces a few stylistic constraints that turned out to matter a lot:</p><ul><li>praise must be brief</li><li>diagnosis comes first</li><li>no fake archive counts</li><li>no stock encouragement</li><li>no padded abstraction</li></ul><p>That helped push the output toward something closer to Thanion’s voice: exact, unsentimental, and useful.</p><h3>3. Runtime validation and coercion</h3><p>Even with a schema, model outputs drift.</p><p>So the route validates the returned fingerprint, checks the evidence against the actual retrieved text, normalizes missing prefixes, and coerces older or flatter JSON shapes into the structure the panel expects.</p><p>That coercion layer turned out to be more important than I expected. The model might conceptually “get it” and still return the wrong nesting shape. Rather than fail every run, the route repairs what can be repaired and only rejects output when the core structure or evidence is truly unusable.</p><p>That was one of the moves that made the feature practical.</p><h3>Historical proof, not current-draft hallucination</h3><p>One of the easiest ways to build a weak pattern system is to let it prove the pattern from the same piece it is analyzing. That produces feedback that sounds cross-document without actually being cross-document.</p><p>So Deep Trace does not allow that.</p><p>The current draft is used for context and for the “In this draft” step. Historical proof comes from earlier saved writing only. That is enforced in both retrieval and validation.</p><p>Once that rule was in place, the quality of the results changed dramatically. The model had to stop pretending and start pointing to prior evidence.</p><p>That is the moment Deep Trace became more than a dressed-up revision helper.</p><h3>The retrieval-performance problem</h3><p>The biggest operational challenge was speed.</p><p>Early versions were too slow because they carried too much:</p><ul><li>too many historical excerpts</li><li>overly long chunk clips</li><li>too many required output fields</li><li>verbose prompt instructions</li><li>too much schema description</li></ul><p>All of that adds latency. All of it increases the chance of malformed or truncated output.</p><p>So I cut aggressively.</p><p>The current version trims:</p><ul><li>current draft excerpts</li><li>historical excerpt count</li><li>excerpt clip length</li><li>output verbosity</li><li>nonessential schema burden</li></ul><p>I also removed a lot of output that looked attractive but was not essential. Deep Trace did not need a flood of editorial prose to be good. It needed sharper structure and less drag.</p><p>That is a theme I trust generally in systems work: most of the time, the feature gets stronger when the output gets leaner.</p><h3>What the system actually found in my own archive</h3><p>The architecture only matters if the result is worth reading.</p><p>One recent run surfaced a pattern it named <strong>The Chain Sentence That Never Lands</strong>. That came from earlier pieces where I had built long clause chains with “and,” carrying the sentence well past the point where its strongest image should have been allowed to stand.</p><p>That was useful because the diagnosis was not just “too long.” It was about where attention was being spent and why the best detail was arriving late.</p><p>Another run caught a different pattern: the tendency to add the sentence that explains what the scene already showed. Again, not a generic style complaint. A recurring move. A repeated reflex. One that showed up across multiple saved pieces and only became obvious when the archive was read longitudinally.</p><p>That is the difference Deep Trace was supposed to make.</p><p>Not better adjectives.<br>Not prettier prose.<br>Not synthetic wisdom.</p><p>A better read on recurrence.</p><h3>What students and academic users get from the same system</h3><p>The archive problem is not limited to memoir or essays. Students repeat habits too.</p><p>They summarize before arguing. They conclude before the evidence has earned the claim. They explain the point because they do not trust the paragraph to carry it. They repeat academic scaffolding that sounds safe but weakens the read over time.</p><p>Deep Trace is useful here because it gives a student something more specific than “improve the writing.” It shows the repeated move. That changes the revision from a vague cleanup exercise into something much more teachable.</p><p>This is also why rubric support matters on the academic side. Once assignment criteria are in the loop, Thanis can read not just for general clarity and coherence, but for the kind of work the assignment is actually asking for. That pushes the system closer to genuine draft governance rather than broad, generic feedback.</p><h3>Why I think this architecture matters and why it worked for Thanis.</h3><p>There is a larger pattern here beyond writing tools.</p><p>Most AI systems are being optimized to produce text faster. I think some of the more valuable systems will be the ones that can read accumulated text more honestly.</p><p>That requires a different architecture. Not just a stronger model. A better retrieval story. Better containment. Stronger validation. Smaller, sharper outputs. Less hallucinated confidence. More evidence.</p><p>That is what Deep Trace pushed me to build this feature into Thanis</p><p>I did not want a machine that sounds like it understands writing.<br>I wanted to build a system that can prove a recurring pattern across stored work and return something a writer can actually use.</p><p>That was the real target all along.</p><h3>Acknowledgments</h3><p>Early feedback on Deep Trace came from <a href="https://medium.com/u/6da965be0b80">Sherene Jensen</a> whose thoughtful testing and perspective on how writers actually interact with analysis tools helped shape the early iterations of this feature.</p><p>The conceptual foundation for this work also draws from earlier research by <a href="https://medium.com/u/ab961d4af13f">Thanion, Ph.D.</a> whose exploration of <em>syntax geometry</em> and structural patterning in language helped frame many of the ideas that eventually informed the Deep Trace approach.</p><p>Both contributions played an important role in bringing the feature from concept to implementation.</p><h4><strong>Deep Trace™ and Thanis™ are trademarks of Obligra LLC.</strong></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dc2a9722abbd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thanis-insights/how-i-built-thanis-deep-trace-on-aws-to-detect-writing-patterns-across-a-living-archive-dc2a9722abbd">How I Built Thanis Deep Trace on AWS to Detect Writing Patterns Across a Living Archive</a> was originally published in <a href="https://medium.com/thanis-insights">Thanis Insights</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Thanis Deep Trace Showed Me the Parts of My Writing I Was Still Hiding From]]></title>
            <link>https://medium.com/thanis-insights/thanis-deep-trace-showed-me-the-parts-of-my-writing-i-was-still-hiding-from-2d96fd017c21?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/2d96fd017c21</guid>
            <category><![CDATA[revision]]></category>
            <category><![CDATA[creative-process]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[writing-tools]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Fri, 13 Mar 2026 13:25:54 GMT</pubDate>
            <atom:updated>2026-03-15T16:35:36.330Z</atom:updated>
            <content:encoded><![CDATA[<p>There is a level of writing feedback that most tools never reach.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1GhMj4TiOaQqq3adZxZ06w.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@joa70?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Joachim Schnürle</a> on <a href="https://unsplash.com/photos/magnifying-glass-focuses-on-a-dictionary-page-2C8_YqfxS5w?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>They can tell you whether a sentence is clear. They can tell you whether a paragraph is too long. They can tell you the draft is “strong” or “risky” or “needs revision,” which is often just a cleaner way of saying very little.</p><p>What they usually cannot do is tell you what keeps returning across your work, what pattern has quietly become habit, and what that habit is costing the page.</p><p>That is the real problem Thanis Deep Trace was built to solve.</p><p>A normal analysis reads the draft in front of you. Deep Trace reads the draft in front of you and then asks a harder question: what keeps showing up in the writing before this one, and what does that repetition reveal?</p><p>That distinction matters. A lot.</p><p>Because one draft can lie to you. One draft can look sharp, coherent, controlled, emotionally honest, and still be carrying the same unconscious move you have made six times before. When that happens, the problem is no longer the paragraph. The problem is the pattern.</p><p>I ran Deep Trace against my own archive recently, and the output was uncomfortably exact. Not flattering. Not cruel. But Exact.</p><p>I used the story i wrote called <a href="https://medium.com/shaped-souls/crawlspace-f38b69c651b0">‘Crawlspace’ </a>posted last October</p><p>Here is one of the strengths it surfaced first:</p><blockquote><em>“You write physical detail that earns its place.”</em></blockquote><p>That line really mattered to me because it named something I care about deeply in my own work. I do not want detail sitting there like upholstery. I want it to do labor. I want it to carry pressure, setting, discomfort, memory, and consequence all at once. Deep Trace pointed to a line from one of my pieces:</p><blockquote><em>“The heat immediately followed us from room to room, stubbornly clinging to our backs like another shirt”</em></blockquote><p>That is the kind of sentence I trust. It is physical before it is interpretive. It lets the body arrive before the explanation.</p><p>But then Deep Trace did what it was supposed to do. It stopped complimenting me and started reading me.</p><h3>The first pattern it caught: the sentence that keeps going because I do not want to stop it yet</h3><p>One of the recurring patterns it found was this:</p><p><strong>The Chain Sentence That Never Lands</strong></p><p>That title is plain, almost bluntly plain, which is part of why it works. There is no jargon hiding inside it. No workshop language. No abstract fog. It names the problem as a reader feels it.</p><p>The diagnosis was simple: I keep building long sentences by chaining clauses together with “and” until the strongest detail arrives too late, buried inside momentum that has already spent the reader’s attention.</p><p>It pointed to one of my own lines:</p><blockquote><em>“the cat had moved farther than it first seemed, and every time I felt nearly close, her eyes blinked again just out of reach, and distance, very predictably, stretches when your cheek is pressed to the earth.”</em></blockquote><p>And the note underneath it was sharper than most human feedback I have received on sentence rhythm:</p><blockquote><em>“The best image, cheek pressed to earth, arrives at the end of a chain that has already spent the reader’s attention.”</em></blockquote><p>That is not a grammar correction. That is not style-policing. That is structural reading.</p><p>It also showed me another example from older work:</p><blockquote><em>“Two theses moving at once, long weeks that never really ended, and a room where the lights felt the same at 3 p.m. or 3 a.m.”</em></blockquote><p>Again, the issue is not that the line is bad. The issue is that all three pressures arrive at the same speed and with the same weight, so none of them gets the full force it might have had on its own.</p><p>This is exactly the kind of thing a writer often defends as rhythm. I know I have. It feels alive while drafting. It feels like motion. It feels like the sentence is carrying urgency. Deep Trace showed me that urgency and accumulation are not always the same thing. Sometimes the sentence is alive. Sometimes it just has not decided where to stop.</p><p>That is a much more useful distinction than “tighten this paragraph.”</p><h3>The second pattern was worse, because it caught something I thought was honesty</h3><p>The other pattern it found was this:</p><p><strong>Telling the Reader What the Story Already Showed</strong></p><p>That one cut deeper because I recognized it instantly.</p><p>I have a habit, especially near the end of a piece or near the close of a pressure point, of adding the sentence that tells the reader what the scene has already made available. Not because I think the reader is foolish. Because while drafting, that sentence feels responsible. It feels like I am clarifying the emotional truth of the moment.</p><p>Deep Trace showed me what that move looks like from the outside.</p><p>It pulled this line from one of my earlier pieces:</p><blockquote><em>“I was not chasing productivity. I was running from stillness.”</em></blockquote><p>That line is strong. I still think it is strong. But Deep Trace’s point was not that the sentence failed. The point was that the image and emotional situation before it had already done the work. The sentence arrived to make sure the reader got it, and in doing so stepped slightly in front of the scene.</p><p>It also flagged this line from another piece:</p><blockquote><em>“So this is the story I’m willing to tell in public: not a performance, not a ledger of sins, just a record…”</em></blockquote><p>Again, true. Again, honest. Again, not necessarily needed where it arrived.</p><p>That is the kind of feedback I built Deep Trace for.</p><p>Not because it humbles the writer. Because it shows the writer where honesty has tipped into explanation, where the need to be understood has stepped between the scene and the reader, where the final sentence is really the sentence after the final sentence.</p><p>Writers do this all the time. Students do it too. They explain the paragraph they should have trusted. They summarize the insight the evidence already carried. They name the feeling before the image has had time to earn it. On the page it looks like caution. Underneath, it is often a kind of fear.</p><p>Deep Trace can see that pattern more clearly than a single-draft read can.</p><h3>This is why I use it now</h3><p>I do not use Deep Trace while the piece is still soft and half-born. I do not run it after every minor edit. I use it when the draft is real enough to expose me.</p><p>That is when it becomes valuable.</p><p>It shows me the habit beneath the sentence. The tendency beneath the page. The thing I am doing not once, but repeatedly, and often most intensely where the writing matters most.</p><p>And that, to me, is the difference between a tool that merely comments on writing and a tool that actually helps a writer grow.</p><p>Most AI writing tools still assume the answer is to generate faster, rewrite cleaner, smooth the edges, flatten the irregularities, and produce something that sounds generally competent and broadly acceptable. I think that is the wrong ambition.</p><p>Writers do not only need help producing text. They need help seeing themselves on the page with more honesty than comfort usually allows.</p><p>That is what Deep Trace is for.</p><ol><li>It is not there to replace the line. It is there to show me what kind of line I keep making.</li></ol><p>2. It is not there to ghostwrite a paragraph. It is there to reveal the recurring pressure inside the paragraph.</p><p>3. It is not there to flatter me into thinking the draft is fine. It is there to make me harder to fool.</p><h3>Why students need this too</h3><p>Students are often told they need to “improve their writing,” which is vague enough to be nearly useless.</p><p>What they usually need is something narrower and more exact. They need to see the repeated move.</p><p>The sentence that over-explains.<br>The paragraph that summarizes before it argues.<br>The conclusion that retreats into abstraction.<br>The emotional label that arrives before the evidence.<br>The structural habit that keeps making the same draft weaker in the same way.</p><p>That is where Deep Trace becomes useful in academic work too. Not as a machine that writes the paper. As a system that makes the student’s own habits visible enough to revise against.</p><p>That is a more serious use of AI than generation, and frankly a more defensible one.</p><h3>The deeper point</h3><p>I built Deep Trace because I wanted a tool that could do more than react to the draft in front of me. I wanted something that could read the writing across time. Something that could notice the recurrence, the pressure, the return. Something that could tell me, plainly, when a pattern had hardened enough to become visible.</p><p>That is what this run did.</p><p>It showed me where I keep adding the sentence after the scene has already finished speaking.<br>It showed me where I let syntax keep moving because I have not yet chosen the real stop.<br>It showed me, in other words, the parts of my writing I was still hiding from by calling them style.</p><p>And that is why Deep Trace matters.</p><p>Not because it makes the prose prettier.<br>Not because it saves time.<br>Not because it gives me one more report.</p><p>Because sometimes the most useful thing a writing tool can do is show you the habit you have started mistaking for yourself.</p><h3>Acknowledgments</h3><p>Early feedback on Deep Trace came from <a href="https://medium.com/u/6da965be0b80">Sherene Jensen</a> whose thoughtful testing and perspective on how writers actually interact with analysis tools helped shape the early iterations of this feature.</p><p>The conceptual foundation for this work also draws from earlier research by <a href="https://medium.com/u/ab961d4af13f">Thanion, Ph.D.</a> whose exploration of <em>syntax geometry</em> and structural patterning in language helped frame many of the ideas that eventually informed the Deep Trace approach.</p><p>Both contributions played an important role in bringing the feature from concept to implementation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2d96fd017c21" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thanis-insights/thanis-deep-trace-showed-me-the-parts-of-my-writing-i-was-still-hiding-from-2d96fd017c21">Thanis Deep Trace Showed Me the Parts of My Writing I Was Still Hiding From</a> was originally published in <a href="https://medium.com/thanis-insights">Thanis Insights</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Prompting Trap: Why Writers Need More Than Better Prompts]]></title>
            <link>https://medium.com/thanis-insights/the-prompting-trap-why-writers-need-more-than-better-prompts-22f3874bc9f7?source=rss-2ee1be8ae92b------2</link>
            <guid isPermaLink="false">https://medium.com/p/22f3874bc9f7</guid>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[ai-writing]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Stephen Woodard]]></dc:creator>
            <pubDate>Mon, 09 Mar 2026 15:27:28 GMT</pubDate>
            <atom:updated>2026-03-09T17:04:53.749Z</atom:updated>
            <content:encoded><![CDATA[<p>Spend ten minutes on Medium or Substack and the pattern becomes hard to miss. ChatGPT isnt going to make you a better writer.</p><figure><img alt="Frustrated writer trying to use ChatGPT" src="https://cdn-images-1.medium.com/max/1024/1*nAgb5eRqC0GdtDoXqfIkHw.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@berctk?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Berke Citak</a> on <a href="https://unsplash.com/photos/a-person-typing-on-a-laptop-on-a-table-adrO5seSbBE?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Writers are no longer just writing. More and more often, they are writing about how to get AI to write better, how to make ChatGPT sound more human, how to preserve voice, how to make a paragraph feel less robotic, how to stop a draft from sounding like it was polished by the same machine that polished ten thousand others. The sheer volume of this alone tells me something important, and what it tells me is not simply that people are curious. It tells me the problem is real, widespread, and still very much unresolved.</p><p>Medium’s ChatGPT topic alone now shows about 95,000 stories and roughly 280,000 followers. That is not a fringe discussion anymore. That is a full-blown publishing pattern, and when I pair that with broader data showing how widespread AI-assisted writing has already become, it starts to look much less like a passing trend and much more like the new default. What I have come to believe, after watching this unfold closely, is that we are still asking the wrong question.</p><p>The question is no longer, “Can AI help me write faster?”</p><p>Of course it can.</p><p>The better question, and in my view the much more serious one, is why so many writers are now spending hours trying to fix what AI gave them in the first place. That question matters a great deal more, because it gets closer to the actual problem.</p><h3>What the average writer is actually dealing with now</h3><p>If you write regularly, whether that means essays, newsletters, technical posts, class assignments, blog posts, or even LinkedIn articles, this probably feels familiar in a very immediate way.</p><p>You open ChatGPT because you are stuck on a first paragraph, or because you are tired, or because you have a deadline and simply need to get moving. You ask it for a draft, an outline, a stronger intro, a cleaner conclusion, or a way to make your writing sound more professional. It gives you something fast, and at first glance it often looks decent, sometimes even impressively decent. The grammar is fine. The structure is clean. The sentences are smooth. But then, once you slow down and read it more carefully, another feeling starts creeping in.</p><p>It sounds a little too polished.<br>A little too agreeable.<br>A little too generic.<br>A little too eager to wrap everything up in a neat bow.</p><p>So then the second job begins, and this is the part I think people are talking about more honestly now.</p><p>You start cutting lines that sound fake. You remove transitions you would never actually use. You rewrite the opening because it does not sound like you. You fact-check a confident claim that appeared out of nowhere. You tighten the middle because the model drifted into filler. You add back a detail from your own experience because the draft somehow managed to say a lot without saying anything truly specific. What looked quick at first slowly turns into another kind of work.</p><p>This happens constantly now, and I suspect more people are living in this loop than want to admit it.</p><p>A student asks AI to help with a paper on <em>The Great Gatsby</em>, and the model produces a clean five-paragraph response with references to symbolism, wealth, and the American Dream. It looks fine on the surface, maybe even reassuringly fine when the student is tired and behind. But once they compare it to the actual rubric, the weakness becomes obvious. It is mostly summary. It does not really analyze. It does not build a strong claim. It does not sound like the student’s own thought process. So now the student has to repair it.</p><p>A newsletter writer asks for a stronger opening paragraph, and the model gives them something sleek, competent, and completely forgettable, something that sounds like every other “thought leadership” intro already floating around the internet. So they spend twenty minutes trying to get their own voice back into it.</p><p>A marketing manager uses AI to draft a client update. The sentences are perfectly acceptable, technically fine, and polished enough to send. But the tone is off just enough to feel insincere. Too polished. Too neutral. Not how they would naturally explain something to a real customer. So again, more cleanup.</p><p>A college student pastes in a rough discussion post and says, “Make this better.” The AI turns it into something longer and more formal, but now the professor can tell the voice shifted. The student saved time on drafting, but lost confidence on submission.</p><p>That, from what I have seen and from what I have heard over and over again, is the real workflow for a lot of people now.</p><p>The draft gets faster.<br> The trust gets slower.</p><h3>The market responded with prompts, hacks, and rituals</h3><p>What happened next makes perfect sense.</p><p>Once people realized that general-purpose AI often gave them something usable but not fully trustworthy, they started building workarounds, and in many ways the current flood of prompt culture is a direct response to that frustration.</p><p>That is why the internet is now crowded with posts like:</p><p>“7 prompts to make ChatGPT sound human”</p><p>“How I trained ChatGPT to write in my voice”</p><p>“Use this one trick to avoid robotic writing”</p><p>“The exact prompt I use to write better essays”</p><p>“How to humanize AI content in minutes”</p><p>“Prompt engineering for better writing”</p><p>I understand why that happened. I really do. When a tool gives you almost what you need, but not quite, you start inventing little rituals to close the gap. You add more adjectives. You ask for a warmer tone. You tell it to sound less robotic. You ask it to write like a teacher, a founder, a student, a journalist, a copywriter, a human being. You keep nudging and refining because you can feel that the draft is close, but not close in the way that matters.</p><p>And yes, sometimes it helps. I do not think it is fair or accurate to say these prompting tricks never work. Some of them clearly do. But what I have come to believe is that this entire prompt-heavy culture is also a sign of something missing. Writers are spending enormous amounts of time trying to force a general-purpose model to behave like a dedicated writing partner. They are trying to coax judgment, voice, consistency, and self-awareness out of a tool that was not really built around those things in the first place.</p><p>That is not a stable long-term answer.</p><p>It may be common. It may even be normal by now. But it is still a workaround.</p><h3>The problem is not just generation. It is governance.</h3><p>This is the point I keep coming back to, and the more I have researched this and spoken with people living inside these workflows every day, the more convinced I have become that this is where the real shift is happening.</p><p>AI did not simply change how people draft. It changed what they now have to manage after the draft appears, and I think that second part is still badly under-discussed.</p><p>The hidden cost of AI writing is not just hallucination, though hallucination is certainly part of it. The hidden cost is everything that comes after generation:</p><p>checking whether the facts are real,<br>checking whether the tone still sounds like you,<br>checking whether the argument actually improved,<br>checking whether the writing got stronger or just smoother,<br>checking whether the revision moved closer to the assignment, the audience, or the original point.</p><p>That is not a small issue. That is the work.</p><p>What I gradually came to believe, and what became much clearer through months of research, observation, and conversations with writers, editors, academics, and students, was that these frustrations were not random. People described them differently depending on where they sat, but the underlying failures were surprisingly consistent. One writer would say the draft sounded polished but empty. A student would say the paper looked better but still did not answer the prompt. An editor would say the copy was competent, but strangely generic. A founder would say the post was fine, but no longer sounded like something they would actually say. The more I listened, the more I researched, and the more I worked through these workflows myself, the more obvious it became that this was not one problem. It was a repeatable pattern.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eHs0wHgVa1DntmtXp5jdcw.png" /><figcaption>The Problem Matrix</figcaption></figure><p>What matters to me is not simply that these issues keep showing up, but that each one carries a cost that people feel almost immediately. Hallucinations create credibility damage and rework. Voice dilution weakens authorship and brand. Tone drift slows approvals. Repetition increases editing time. Prompt chasing kills throughput. Writing that clearly “sounds like AI” raises rejection risk, even when the underlying idea is decent. And when results cannot be repeated reliably, the workflow stops being dependable in any serious way. What I have learned through all of this is that these are not minor side effects sitting off to the edge of the process. Very often, they are the process.</p><p>I have come to think of this as verification labor.</p><p>It is the invisible effort people now spend cleaning, validating, correcting, and restoring writing after AI has helped produce it. And in a surprising number of cases, that labor quietly eats up the time people thought they were saving.</p><p>A writer saves thirty minutes on a draft and then spends forty minutes fixing it.</p><p>A student gets a cleaner paragraph, but then has to rewrite it so it actually answers the question.</p><p>A professor reads an essay that is technically fine but strangely hollow, and now has to decide whether the issue is weak thinking, AI dependence, or both.</p><p>An editor receives a piece that is grammatically polished but has no texture, no tension, and no real sense of authorship. The sentences behave, but the writing does not feel lived in.</p><p>That is where we are.</p><h3>Why this matters for students</h3><p>Students may be feeling this pressure more than anyone else, and what concerns me is not simply that they are using AI. That part is obvious. What concerns me more is how often they are using it without a reliable framework for knowing whether the result is actually helping them think more clearly or simply helping them submit something cleaner.</p><p>Take a very ordinary example.</p><p>A student has a history assignment due at 11:59 p.m. They are tired, behind, and staring at a blank page. They ask AI to “write an intro about the causes of World War I.” The tool gives them something polished and immediate, and for a brief moment the student feels relief. But the relief does not last, because now a different set of questions starts showing up.</p><p>Does this match what we covered in class?<br>Is the teacher asking for summary or argument?<br>Did the model just invent that line?<br>Does this sound like something I would actually write?<br>Is this too formal for a short response?<br>If I submit this, will I even be able to explain it out loud tomorrow?</p><p>That is a very modern kind of anxiety, and I suspect it is more common than people want to say plainly.</p><p>The tool solved the blank page problem, but it introduced a different one: distance from the work itself.</p><p>And if a student keeps doing that often enough, they may get faster at producing assignments while getting weaker at building arguments, noticing weak logic, or hearing their own voice on the page. That is not mentoring. That is substitution with side effects.</p><p>What students need is not just a machine that can give them cleaner paragraphs. They need help understanding where their reasoning is thin, where they are summarizing instead of analyzing, where their evidence does not support their point, and where the final draft no longer sounds like something they can stand behind. That is a much more specific and much more serious problem than “make this better.”</p><h3>Why this matters for writers</h3><p>Professional writers deal with a similar problem, but it shows up a little differently.</p><p>For them, the risk is usually not that the writing is incorrect in an academic sense. More often, the risk is that the writing becomes interchangeable, and once that starts happening often enough, something more important gets lost.</p><p>A founder uses AI to draft a post and suddenly sounds like every other founder.</p><p>A journalist uses it for structure and has to double-check whether a line overstates the claim.</p><p>A content writer gets a fast first draft but loses the exact phrasing that made the idea feel personal.</p><p>A blogger ends up with something that reads fine but could have been written by almost anybody.</p><p>That loss may seem subtle at first, but it adds up, and I think many writers can feel it long before they know exactly how to name it.</p><p>Over time, if the tool keeps smoothing your phrasing, widening your claims, and flattening your rhythm, then it is not just helping you draft. It is slowly reshaping how you sound in public.</p><p>I think a lot of writers already feel that, even if they do not always say it directly.</p><p>They read a draft and think, “This works, but it does not feel like me.”</p><p>Or maybe, more honestly, “This says what I meant, but not in the way I would have said it.”</p><p>That is not a small difference.</p><p>That difference is voice.</p><p>And once voice starts getting diluted inside a workflow, the writer has to do yet another kind of repair work, not just editing for correctness, but editing for identity. That, to me, is one reason so many people now spend their time learning prompt tricks. They are trying to preserve something the tools keep threatening to smooth away.</p><h3>The numbers matter, but so does what they point to</h3><p>What I have researched only deepened this concern, because the broader publishing environment already shows how normal AI-shaped writing has become.</p><p>GPTZero’s review of popular Substacks found that 10% used some detectable AI and 7% relied on it more significantly across posts. Ahrefs found that 74.2% of 900,000 newly created webpages they analyzed contained some AI-generated content. Those numbers do not mean everything online is fake. They do mean AI-shaped writing is now normal enough that questions of trust, authorship, and defensibility are no longer niche concerns.</p><p>That changes the standard.</p><p>We are no longer operating in a world where AI writing is occasional or experimental. We are operating in a world where it is common, mixed into everyday workflows, and often difficult to separate from the rest of the writing process.</p><p>Which means the next meaningful question is not whether people will use AI.</p><p>They already are.</p><p>The question is what kind of infrastructure they will need around it.</p><h3>Why I built Thanis</h3><p>That is the question that pushed me toward building Thanis.</p><p>I did not want to build another prompt library. I did not want to build another wrapper that promises better results if you just phrase the request the right way. And I did not want to keep watching writers and students burn hours trying to reverse-engineer trust out of systems that were built mainly for generation.</p><p>My view is fairly simple.</p><p>The next real category in this space is not better prompting. It is better writing governance.</p><p>That means tools that help people do more than generate text. Tools that help them examine what changed, where the writing drifted, whether the reasoning weakened, whether the tone remained consistent, whether the piece still aligns to the rubric, the assignment, the audience, or the writer’s own standards.</p><p>In plain terms, I think people need something more practical than “prompt harder.”</p><p>They need help answering questions like:</p><p>Did this revision actually improve my argument?<br>Did I just lose my own voice in the cleanup process?<br>Does this paragraph match the assignment, or did it become generic?<br>Where did the reasoning get thinner?<br>Can I defend this piece if someone asks how it was developed?</p><p>Those are not abstract concerns. They are ordinary, everyday concerns now.</p><p>That is the gap Thanis was designed to address.</p><h3>What I believe is coming next</h3><p>I do not think the future of writing belongs to whoever generates the most text the fastest.</p><p>I think it belongs to the people who can still stand behind what they publish.</p><p>The people who can move quickly without handing over judgment.<br>The people who can use AI without letting it flatten their voice.<br>The people who can show how a piece evolved instead of shrugging and saying the machine helped somehow.<br>The people who still know the difference between a clean sentence and an honest one.</p><p>That is where I believe this is heading.</p><p>Not away from AI, but away from blind trust in generation alone.</p><p>The market spent the last phase chasing output. I believe the next phase will care much more about proof.</p><p>Proof that the work still sounds like you.<br>Proof that the reasoning held together.<br>Proof that the revision improved the draft instead of weakening it.<br>Proof that speed did not come at the expense of credibility.</p><p>That, in the end, is what I have been working on.</p><p>Not a way to produce more words.</p><p>But way to help people keep what matters while they write with machines.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22f3874bc9f7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thanis-insights/the-prompting-trap-why-writers-need-more-than-better-prompts-22f3874bc9f7">The Prompting Trap: Why Writers Need More Than Better Prompts</a> was originally published in <a href="https://medium.com/thanis-insights">Thanis Insights</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>