<?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 Marina on Medium]]></title>
        <description><![CDATA[Stories by Marina on Medium]]></description>
        <link>https://medium.com/@vii.marina?source=rss-41506d6f3981------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*f1apdULA0ud9SBy6F8ogxQ.jpeg</url>
            <title>Stories by Marina on Medium</title>
            <link>https://medium.com/@vii.marina?source=rss-41506d6f3981------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 06:03:12 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vii.marina/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[How to Display the Full Assignee History of a Jira Issue in Confluence]]></title>
            <link>https://medium.com/@vii.marina/how-to-display-the-full-assignee-history-of-a-jira-issue-in-confluence-911aa0c58f35?source=rss-41506d6f3981------2</link>
            <guid isPermaLink="false">https://medium.com/p/911aa0c58f35</guid>
            <category><![CDATA[jira]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[jira-integration]]></category>
            <category><![CDATA[atlassian]]></category>
            <category><![CDATA[confluence]]></category>
            <dc:creator><![CDATA[Marina]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 10:41:36 GMT</pubDate>
            <atom:updated>2026-11-05T10:41:36.746Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A step-by-step guide to building a macro that shows who held each ticket and when across any set of Jira issues.</em></p><p>There’s a piece of information buried inside every Jira issue that most teams never surface: the complete list of everyone who was ever assigned to it. You can dig it out one ticket at a time through the activity tab, but the moment you need that history across a set of issues, for a retrospective, a handoff audit, or a bottleneck analysis, Jira leaves you with nothing.</p><p>A community member named Sam Tang put the problem plainly in the Atlassian Community:</p><blockquote>“I want to list out a table to display some tickets’ assignee history… Ticket ID, ticket summary, assignee history: ID-1 ticket 1 Tom &gt; Jerry &gt; Shannon.”</blockquote><p>That request is completely reasonable and surprisingly hard to fulfill out of the box. So we built a Confluence macro that does exactly this, pulling the full assignee chain for any set of Jira issues and rendering it in a clean, linked table. Here’s the full walkthrough.</p><h3>What we’re building</h3><p>The finished macro queries Jira’s REST API, walks through each issue’s changelog for assignee changes, and renders a table in Confluence with three columns: issue key (linked to Jira), summary, and the full assignee chain from first to current.</p><p>It’s built with <strong>User Macro for Confluence Cloud</strong> and goes through four iterations, each one adding something meaningful.</p><h3>STEP 1: The bare-bones version</h3><p>We start with the simplest possible implementation — a static, block-type macro with no body.</p><p>To make it flexible from day one, we add a single <strong>String</strong> parameter called jql (not required), with this default value:</p><pre>JQL DEFAULT<br>reporter = currentUser() AND updatedDate &gt;= &quot;-30d&quot;</pre><p>That default pulls issues the current user reported and that were updated within the last 30 days — a sensible starting point that works without any configuration.</p><pre>VELOCITY TEMPLATE — STEP 1<br>## Search for issues with key, summary, and changelog included<br>#set ( $issues = $JiraManager.get(&quot;/rest/api/3/search/jql?fields=key,summary,assignee&amp;expand=changelog&amp;jql=${parameters.jql}&quot;).issues )<br> <br>#foreach ( $i in $issues)<br>  &lt;p&gt;$i.key:&lt;br&gt;<br>  #foreach ( $h in $i.changelog.histories )<br>    #foreach ($item in $h.items)<br>      #if ($item.fieldId == &quot;assignee&quot;)<br>        $!item.fromString → $!item.toString&lt;br&gt;<br>      #end<br>    #end<br>  #end<br>&lt;/p&gt;<br>#end</pre><p>A few things worth noting. The $JiraManager.get() call uses OAuth, which means the first time a user runs the macro, they’ll be prompted to authorize access. After that, requests run on behalf of that user, so Jira’s own permission model stays fully intact.</p><p>The $! prefix on $item.fromString and $item.toString is doing important work. Without it, Velocity throws a rendering error on any issue that was ever unassigned, since fromString is null in that state. The ! operator silently swaps null for an empty string.</p><blockquote><strong><em>Tip: </em></strong><em>After creating the macro, add it to a page via the Macro Browser or by typing /Issue Assignee History. If it doesn’t appear right away, a hard reload (Ctrl/Cmd + R) clears the client-side cache.</em></blockquote><h3>STEP 2: Handling empty results and showing the current assignee</h3><p>The first version outputs raw changelog entries — useful as a proof of concept, but it misses the current assignee entirely and has no graceful fallback when the JQL returns nothing.</p><pre>VELOCITY TEMPLATE — STEP 2<br>## Search for issues with key, summary, and changelog included<br>#set ( $issues = $JiraManager.get(&quot;/rest/api/3/search/jql?fields=key,summary,assignee&amp;expand=changelog&amp;jql=${parameters.jql}&quot;).issues )<br> <br>## Check for an empty result<br>#if(!$issues)<br>  No issues for JQL &lt;code&gt;${parameters.jql}&lt;/code&gt;<br>  #stop ## End of execution, there is nothing to render<br>#end<br> <br>#foreach ( $i in $issues)<br>  &lt;b&gt;$i.key:&lt;/b&gt;<br>  ## Current Assignee<br>  #if ($i.fields.assignee)<br>    $i.fields.assignee.displayName<br>  #else<br>    &lt;i&gt;(Unassigned)&lt;/i&gt;<br>  #end<br><br><br>  #foreach ( $h in $i.changelog.histories )<br>    #foreach ($item in $h.items)<br>      #if ($item.fieldId == &quot;assignee&quot;)<br>        #if ($item.fromString)<br>← $item.fromString<br>        #else<br>← &lt;i&gt;(Unassigned)&lt;/i&gt;<br>       #end<br>    #end<br>  #end<br>#end<br>&lt;br&gt; ## new line for new issue<br>#end</pre><p>The #stop directive cuts execution cleanly when there are no matching issues, rather than rendering an empty block. The history now reads right to left — current assignee first, previous ones trailing with a ← arrow. Save with Ctrl/Cmd + S and reload the page to see the updated output.</p><h3>STEP 3: Clickable usernames, chronological order, and a lozenge for unassigned states</h3><p>At this point, the macro works correctly, but the assignee chain isn’t guaranteed to be in chronological order, and the usernames are plain text. This step cleans up both.</p><p>Three changes make the difference:</p><p><strong>Sort the history chronologically</strong> using Velocity’s $SortTool:</p><pre>SORT BY CREATION DATE<br>#set ($sortedHistory = $SortTool.sort($i.changelog.histories, [&quot;created&quot;]))</pre><p><strong>Link usernames</strong> using Confluence Storage Format’s &lt;ri:user&gt; tag, replacing display names with account IDs:</p><pre>LINKED USER REFERENCE<br>&lt;ac:link&gt;&lt;ri:user ri:account-id=&quot;${item.from}&quot;/&gt;&lt;/ac:link&gt;</pre><p><strong>Style unassigned states</strong> with an Atlassian UI lozenge. The aui-lozenge-subtle class keeps it visually unobtrusive:</p><pre>UNASSIGNED LOZENGE<br>&lt;span class=&quot;aui-lozenge aui-lozenge-subtle&quot;&gt;unassigned&lt;/span&gt;</pre><pre>VELOCITY TEMPLATE — STEP 3 (FULL)<br>#set ( $issues = $JiraManager.get(&quot;/rest/api/3/search/jql?fields=key,summary,assignee&amp;expand=changelog&amp;jql=${parameters.jql}&quot;).issues )<br> <br>#if(!$issues)<br>  No issues for JQL &lt;code&gt;${parameters.jql}&lt;/code&gt;<br>  #stop<br>#end<br> <br>#foreach ( $i in $issues)<br>  &lt;b&gt;$i.key:&lt;/b&gt;<br> <br>  #set ($sortedHistory = $SortTool.sort($i.changelog.histories, [&quot;created&quot;]))<br>  #foreach ( $h in $sortedHistory )<br>    #foreach ($item in $h.items)<br>      #if ($item.fieldId == &quot;assignee&quot;)<br>        #if ($item.from)<br>&lt;ac:link&gt;&lt;ri:user ri:account-id=&quot;${item.from}&quot;/&gt;&lt;/ac:link&gt; →<br>        #else<br>&lt;span class=&quot;aui-lozenge aui-lozenge-subtle&quot;&gt;unassigned&lt;/span&gt; →<br>        #end<br>      #end<br>    #end<br>  #end<br><br><br>  #if ($i.fields.assignee)<br>&lt;ac:link&gt;&lt;ri:user ri:account-id=&quot;${i.fields.assignee.accountId}&quot;/&gt;&lt;/ac:link&gt;<br>  #else<br>&lt;span class=&quot;aui-lozenge aui-lozenge-subtle&quot;&gt;unassigned&lt;/span&gt;<br>  #end<br>&lt;br&gt;<br>#end</pre><p>The direction flips here too — arrows now run left to right (→), with the current assignee at the end. That matches how most people read a handoff chain naturally.</p><h3>STEP 4: A proper table with linked issue keys</h3><p>The last step wraps everything in a table and makes the issue key a direct link into Jira. The challenging part is that we need the Jira base URL, but all we have access to is Confluence’s $baseUrl. Since Confluence Cloud lives at /wiki, stripping that suffix gives us the Jira root:</p><pre>DERIVE JIRA URL FROM CONFLUENCE BASEURL<br>#set ( $jiraURL = $StringUtils.replace($baseUrl, &quot;/wiki&quot;, &quot;&quot;) )<br>## https://subdomain.atlassian.net/wiki → https://subdomain.atlassian.net</pre><p>We also add white-space: nowrap to the issue key cell so keys like PROJ-1234 never wrap mid-string in a narrow table column.</p><pre>VELOCITY TEMPLATE — FINAL<br>#set ( $issues = $JiraManager.get(&quot;/rest/api/3/search/jql?fields=key,summary,assignee&amp;expand=changelog&amp;jql=${parameters.jql}&quot;).issues )<br>#set ( $jiraURL = $StringUtils.replace($baseUrl, &quot;/wiki&quot;, &quot;&quot;) )<br> <br>#if(!$issues)<br>  No issues for JQL &lt;code&gt;${parameters.jql}&lt;/code&gt;<br>  #stop<br>#end<br> <br>&lt;table&gt;<br>&lt;thead&gt;<br>&lt;tr&gt;<br>  &lt;th&gt;Issue&lt;/th&gt;<br>  &lt;th&gt;Summary&lt;/th&gt;<br>  &lt;th&gt;Assignee history&lt;/th&gt;<br>&lt;/tr&gt;<br>&lt;/thead&gt;<br>&lt;tbody&gt;<br> <br>#foreach ( $i in $issues)<br>&lt;tr&gt;<br>  &lt;td style=&quot;white-space: nowrap;&quot;&gt;<br>    &lt;a href=&quot;${jiraURL}/browse/${i.key}&quot;&gt;$i.key&lt;/a&gt;<br>  &lt;/td&gt;<br>  &lt;td&gt;$i.fields.summary&lt;/td&gt;<br>  &lt;td&gt;<br>    #if ( $i.changelog.histories )<br>      #set ($sortedHistory = $SortTool.sort($i.changelog.histories, [&quot;created&quot;]))<br>      #foreach ( $h in $sortedHistory )<br>        #foreach ($item in $h.items)<br>          #if ($item.fieldId == &quot;assignee&quot;)<br>            #if ($item.from)<br>    &lt;ac:link&gt;&lt;ri:user ri:account-id=&quot;${item.from}&quot;/&gt;&lt;/ac:link&gt; →<br>            #else<br>    &lt;span class=&quot;aui-lozenge aui-lozenge-subtle&quot;&gt;unassigned&lt;/span&gt; →<br>            #end<br>          #end<br>        #end<br>      #end<br>    #end<br>    #if ($i.fields.assignee)<br>    &lt;ac:link&gt;&lt;ri:user ri:account-id=&quot;${i.fields.assignee.accountId}&quot;/&gt;&lt;/ac:link&gt;<br>      #else<br>    &lt;span class=&quot;aui-lozenge aui-lozenge-subtle&quot;&gt;unassigned&lt;/span&gt;<br>    #end<br>  &lt;/td&gt;<br>&lt;/tr&gt;<br>#end<br> <br>&lt;/tbody&gt;<br>&lt;/table&gt;</pre><p>The result is a fully functional assignee history table. Issue keys are linked directly to Jira, summaries are pulled live from the API, and a chronological chain of linked user references is built from first assignee to current.</p><p><strong>Where to find the finished macro</strong></p><p>We’ve added the completed macro to our library under the name <strong>Issues Assignee History</strong>, ready to import directly into any Confluence Cloud instance running User Macro for Confluence Cloud.</p><p>It’s a small macro for a problem that comes up more often than it should — and it only exists because someone in the community asked the right question.</p><p><em>Originally published at wombatscorp.com</em></p><p>( <a href="https://wombatscorp.com/blog/display-assignee-history-of-a-jira-issue"><em>https://wombatscorp.com/blog/display-assignee-history-of-a-jira-issue</em></a><em> )</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=911aa0c58f35" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What to know before you build an app with AI]]></title>
            <link>https://medium.com/@vii.marina/what-to-know-before-you-build-an-app-with-ai-98ef9d5bd509?source=rss-41506d6f3981------2</link>
            <guid isPermaLink="false">https://medium.com/p/98ef9d5bd509</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <dc:creator><![CDATA[Marina]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 14:44:13 GMT</pubDate>
            <atom:updated>2026-09-02T14:44:13.009Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZSaRJ8JHbpwm1FERTOS8Nw.jpeg" /></figure><h4>The things I wish someone had told me before I started.</h4><p>For seven months I built a learning platform on my own, using AI for most of the work. Courses and lessons, tests and exercises, a dashboard for each of the three roles. Plus a generator that turns a lesson into quiz questions.</p><p>Before that I worked in data labelling for a year. I checked AI-generated images and wrote descriptions of photos, the kind a model trains on. A year of that and you stop being surprised by what it misses. Things a person registers in a second take thousands of examples to teach. Obvious ones especially.</p><p>So I had some idea what was coming. <em>Not enough of one.</em></p><p>Some of it was the tools failing in ways I did not predict. Most of it was learning how to work when the AI writes the code and everything else is on you.</p><h3>Break it into small pieces first</h3><p>The hardest part of a big project is the blank page. You have the idea and an empty folder. What you do not have is any obvious place to start. That is where most projects die.</p><p>AI helps here more than people expect, and not only with code. Tell it what you are building and how it feels to look at. <em>Mine talked me down</em>. That sounds small, but a project you are afraid of is a project you keep postponing.</p><p>Then ask it for a plan. Not a description of the app, an actual breakdown: a handful of large stages, each split into small tasks you can finish in one sitting. What you get is a checklist.</p><p>The checklist does most of the work. You always know where you are and how much is left. You finish things, which keeps you going. And once you have timed a few tasks, you can estimate the rest of the project instead of guessing.</p><p>We were taught this at university, in algorithms and data structures. <em>Divide and conquer.</em></p><h3>Build one thing at a time, end to end</h3><p><strong>Start with the database</strong>. I described in plain words what I needed stored and what fields each table should have, and the AI turned that into SQL. It worked like a translator. From that point on, every feature I built had somewhere real to write to.</p><p>That matters because of how I chose to work. I did not build a frontend mockup first and connect it later. Mock data has to be replaced anyway, and then you are doing the same work twice.</p><p>So each feature went all the way down before I moved on. Database, backend, interface, tested, finished.</p><p>The order followed the product. First the form teachers use to create a course, and I did not touch anything else until it was right. Then filling those courses with content. Only after real courses existed did I build the student side, because now there was something for a student to open. The admin panel came last, once the platform had actual teachers, students and courses in it to manage.</p><p>There is a second reason this works, and for one person it matters more. While the AI works on the next task, you are testing the last one. Two things move at once, which is as close as you get to having a team.</p><h3>Use one AI to write prompts for another</h3><p>I wanted to see the design before it existed in code. Describing a layout in words and having the model build it directly is expensive: you pay tokens for every attempt, and you change your mind a lot in the early stages. I changed the styling many times before it looked right.</p><p>So I asked Codex for a prompt. Not for a design, <strong>for the text I would paste into a design tool</strong>. It filled in the colour palette and the details I had not thought to specify. I took that prompt to <a href="https://www.figma.com/ai/">Figma AI</a>, got mockups back, and could finally see the flow instead of imagining it.</p><p>The mockups then went back into the loop. I exported the file from Figma and handed it over, and the model turned it into components.</p><figure><img alt="Teacher dashboard showing course tree, student progress and test generation" src="https://cdn-images-1.medium.com/max/1024/1*TQ4pEZG_3gQNUrI-Bha4BA.png" /><figcaption>Teacher dashboard. The layout came out of a Figma mockup first.</figcaption></figure><p>Two useful things came out of this. Prompts written by a model are more specific than the ones I write, because it names things I would have left vague. And a mockup you can look at is a much better brief than a paragraph describing what you want.</p><h3>Let the AI check its own output</h3><p>The generator produces tests and exercises from a lesson. Getting it to produce anything was easy. Getting it to produce something worth showing a student took a long time.</p><figure><img alt="Lesson view with the course tree on the left and lesson content on the right" src="https://cdn-images-1.medium.com/max/1024/1*cW_asueF_SNxwUPJTNf_Ag.png" /><figcaption>A lesson. Tests and exercises are generated from this text.</figcaption></figure><p>What helped was stopping the guesswork. Instead of rewriting the prompt and hoping, I gave the model the lesson, the test it had generated from that lesson, and the exercise, and asked whether they were any good. Does this question actually follow from the material. Is the difficulty right. Is there anything wrong with the answer options.</p><p>It found problems I would have scrolled past. Then I asked it to rewrite the prompt so those problems stopped happening, and it did that too.</p><p>That loop is the part I would repeat on any project. Show the model its own work and ask what is wrong with it, instead of rewriting the prompt in the dark and waiting to see what comes out.</p><p>My final prompt looks nothing like the first one. It started as a short list of rules. Now it has a section telling the model which language to write in, another on how hard the questions should be, and a long one about not lifting sentences straight out of the lesson. Almost none of that came from me sitting and thinking about it.</p><h3>Click through everything yourself, and write it down</h3><p>The AI does not open the app. It has no idea whether the button it just built actually works. That part stays with you, <em>and if you skip it, nobody is doing it.</em></p><p>So I tested every feature by hand. Not a quick look, properly: clicking through as a teacher, then as a student, trying the things a real person would try.</p><p>What mattered more was writing the problems down instead of fixing them one at a time. Every bug went into a list. Broken layout, wrong data, a button that did nothing. Then, once the feature was finished, the whole list went over in one message.</p><p>Fixing bugs one by one is slower than it looks. You interrupt work that is already in progress, and half the fixes turn out to be the same underlying problem anyway. A list of twelve gets handled in one pass.</p><p>I kept mine on a Miro board with screenshots. Any format works as long as you can see everything at once.</p><h3>A rule in the prompt is a request, not a guarantee</h3><p>This is the part that took me longest to accept.</p><p>My prompt has a section that says, in several different ways, do not copy sentences out of the lesson. Rewrite the idea, test understanding, come at it from another angle. The model copied sentences out of the lesson anyway. Not every time, but often enough that I had to add a check in the code that throws away any question that appears in the lesson text.</p><p>Same with question types. I would ask for true or false questions and get multiple choice back. Now the type from the request overrides whatever the model labelled it.</p><p>Same with the answer options. Occasionally the correct answer would also show up among the wrong ones, so the student could pick either and both would look right. That check is in the code too now.</p><p>The pattern is always the same.</p><blockquote><strong><em>The prompt asks. The code decides</em></strong>.</blockquote><p>If something in the output actually matters to you, do not leave it to the wording. Check it after the fact and handle it when it fails, because eventually it will.</p><h3>Money runs out faster than you think</h3><p>There are two kinds of cost and it took me a while to separate them.</p><p>The first is the subscription. Twenty dollars a month for <a href="https://chatgpt.com/">ChatGPT</a>, which I hit the limits of regularly and then waited. That is annoying but predictable, and you know what you are paying up front.</p><p>The second is the API, and it works differently. Generating tests and exercises runs through the OpenAI API, and there you pay per token. I put in five dollars. Some of it is still there, but that is only a matter of time, and the number moves every time the generator runs.</p><p>The thing to know about the second kind is that it can be spent by someone other than you. For a while my generation endpoints were open with no authentication, which meant anyone who found them could burn through my balance. That is fixed now: you have to be logged in as a teacher, and there is a cap on how many generations one account can run per minute and per day.</p><p>If your app calls a paid API, put a limit on it <strong>before you deploy, not after</strong>. It costs nothing to add and it is the difference between a five dollar bill and a much worse one.</p><p>The platform is live. Courses, tests, exercises, three roles, and the generator that caused most of the trouble in this article.</p><p>It is not finished. Exercises still check the answer in the browser instead of on the server, which I know about and have not fixed, and my list of things to change is longer than the list of things already changed.</p><p>But it works, and building it taught me more than anything else I have done. The platform is <a href="https://educat-learning-platform.vercel.app/">here</a>, and the code is on <a href="https://github.com/vii-marina/learning-platform">GitHub</a>.</p><figure><img alt="Student dashboard of the platform, showing enrolled courses and progress" src="https://cdn-images-1.medium.com/max/1024/1*_gJmwJb5Kut8J6J0DuWj_Q.png" /><figcaption>The platform is in Ukrainian. Student dashboard, as it looks now.</figcaption></figure><h3>What I would tell myself at the start</h3><p>None of this is about prompting better. I spent seven months learning that the interesting part is everything around the prompt. The order you build things in. What you check by hand, and what you never trust without looking.</p><p>The AI wrote most of the code. It did not decide what to build. It did not notice that a test was wrong, or care that the same validation sat copy-pasted in two files. All of that stayed with me, and it turned out to be the job.</p><p>If you are looking at a project that feels too big, that is the part worth knowing.</p><blockquote><strong><em>The typing gets faster. Everything else is still yours.</em></strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98ef9d5bd509" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Real Cost of AI Coding Tools in 2026]]></title>
            <link>https://medium.com/@vii.marina/the-real-cost-of-ai-coding-tools-in-2026-fac2eb3a972c?source=rss-41506d6f3981------2</link>
            <guid isPermaLink="false">https://medium.com/p/fac2eb3a972c</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Marina]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 09:40:39 GMT</pubDate>
            <atom:updated>2026-09-02T09:43:22.193Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fG6Abfy9qYflIYhC1d2paA.jpeg" /></figure><h4>Everyone sells AI-augmented development on speed. Almost nobody mentions the bill.</h4><p>Cursor charges $16 a month per developer. That is on the pricing page, in large type, easy to multiply.</p><p>OpenAI publishes a different kind of number. In its own documentation, it says Codex usage runs somewhere around $100 to $200 per developer per month. Same category of product. Roughly ten times the figure.</p><p>Neither company is lying. The gap between those numbers is where most of this article lives.</p><p>But before the money, the thing people skip. <em>What do these tools actually do all day?</em></p><h3>What the tools actually do</h3><p>Six kinds of work, and none of them are what the demos show you.</p><p><a href="https://en.wikipedia.org/wiki/Boilerplate_code"><strong>Boilerplate</strong></a><strong>.</strong> The scaffolding, the config, the fourth CRUD endpoint that looks like the previous three. Nobody enjoys writing it and nobody reads it closely afterwards. This is where the tools are strongest and where the risk is lowest.</p><p><strong>Tests.</strong> Generating unit and integration tests straight from the code or from a written requirement. Test coverage is the thing teams sacrifice first when a deadline gets close, so automating it removes a real and recurring failure point.</p><p><strong>Reading code nobody wrote down.</strong> This one is worth more than it sounds. A legacy system with no documentation takes longer to understand than to change, and understanding it has always been the expensive part. The tools read the codebase, pull out the business logic and map what depends on what. <a href="https://acropolium.com/services/ai-augmented-development/"><strong>Acropolium</strong></a> runs legacy modernization as a named service line built on exactly this, which says something about where companies are spending.</p><p><strong>First-pass review.</strong> The AI goes through a pull request before a human does. It catches the obvious things. Hallucinated APIs, drift from your conventions, the small stuff that becomes rework three weeks later. None of it hard to spot. All of it easy to miss at four in the afternoon.</p><p><strong>Documentation.</strong> Keeping the written description in sync with what the code actually does, without anyone remembering to update it.</p><p><strong>Pipelines. </strong>Infrastructure config and deployment logic, plus the validation checks that run before a release. Terraform modules that a human still signs off on.</p><p>Notice the shape of that list. Almost all of it is work that had to happen but that nobody wanted, and none of it is the interesting part of the job.</p><h3>What they do not do</h3><p>The list above has a shape to it. Everything on it is work that can be checked by looking at the result.</p><p><a href="https://en.wikipedia.org/wiki/Software_architecture">Architecture</a> is not on the list. Neither is deciding what to build, or which trade-off your particular business can live with, or whether a slower solution is worth it because your team can maintain it in two years. Those need somebody who knows the company, not the codebase.</p><p>And then there is the part nobody sells.</p><p>A person has to read every line the AI writes before it ships. That work is new. It did not exist in the same volume before, and it lands on your senior engineers, who are the expensive ones.</p><p>Stack Overflow’s 2025 survey asked developers what frustrates them most about these tools. Two thirds named the same thing. Output that is <em>almost right, but not quite</em>. Not broken code. Broken code announces itself. Code that looks correct, reads correctly, and fails on an edge case nobody thought to check.</p><p>Forty-five percent said debugging AI-generated code takes them longer than debugging their own.</p><p>So the honest version is this. The tools take repetitive work off your team and hand back a different kind of work in return. Reading instead of writing. Whether that trade pays off depends entirely on how good your team already is at the reading part.</p><h3>What the tools cost</h3><p>Four tools, four pricing pages, checked on 2 September 2026. All figures in US dollars.</p><p><a href="https://cursor.com/pricing"><strong>Cursor</strong></a> starts at $16 a month for an individual on annual billing. Teams cost $32 per person. There is a Premium team tier at $96 that gives five times the agent limits.</p><p><a href="https://claude.com/pricing"><strong>Claude</strong></a> is $17 for Pro on annual billing, $20 monthly. The Team plan runs $20 per seat annually against $25 monthly, and covers groups of two to 150 people. Premium seats go to $100.</p><p><a href="https://github.com/features/copilot/plans"><strong>GitHub Copilot</strong></a> runs $10 for an individual on Pro. Business seats are $19 per user, and Enterprise doubles that to $39.</p><p><a href="https://openai.com/chatgpt/pricing"><strong>Codex</strong></a> has no pricing page of its own. You buy it inside a ChatGPT plan. Plus is $20. Business seats are $20 annually or $25 monthly, and premium goes to $100.</p><p>Three of the four discount annual billing by 15 to 20%. On a ten-person team that is $960 a year on Cursor alone, and Copilot is the one that does not offer it.</p><figure><img alt="Pricing table comparing individual and team plans for Cursor, Claude, Copilot and Codex" src="https://cdn-images-1.medium.com/max/1024/1*SkWRpIz4qgWUcHMmpOAogQ.jpeg" /><figcaption>Prices from vendor pricing pages, checked 2 September 2026. Annual billing where available.</figcaption></figure><p>So far, so manageable. These are small numbers.</p><h3>Where the number stops being small</h3><p>Every one of those subscriptions has a ceiling. You reach it, and then you keep paying.</p><p>Copilot runs on credits. A Business seat includes 1,900 a month, and past that you pay a cent each. Cursor gives every plan a usage allowance and bills whatever you use beyond it in arrears. Codex moved off message-based pricing entirely in April and now charges by the token, so what you owe depends on which model ran and how much context it chewed through.</p><p>Claude Enterprise does not even pretend otherwise. The pricing page says seat price plus usage at API rates. Opus 5 runs $5 per million input tokens and $25 per million output.</p><p>None of this is hidden. It is all on the vendors’ own pages. It just is not the number in large type.</p><p>Here is the one that matters. OpenAI’s own documentation puts average Codex spend at $100 to $200 per developer per month.</p><p>Run that against the earlier figure. A ten-person team on Cursor Teams costs $3,840 a year. The same ten people at OpenAI’s stated average land somewhere between $12,000 and $24,000.</p><p>Three to six times the sticker price, on numbers the vendor published itself.</p><figure><img alt="Bar chart comparing $3,840 annual subscription cost against $24,000 in real spend for a ten-person team" src="https://cdn-images-1.medium.com/max/1024/1*WcjARgnV3PKouNv-vgcxoQ.jpeg" /><figcaption><em>Left: 10 seats on Cursor Teams at $32. Right: OpenAI’s published average of $200 per developer per month.</em></figcaption></figure><h3>What the research says you get for it</h3><p>Vendor pages talk about speed. Independent research is less tidy.</p><p><a href="https://dora.dev/ai/gen-ai-report/report/">DORA</a>, the software delivery research program run by Google Cloud, found something counterintuitive. A 25% rise in AI adoption came with delivery throughput dropping 1.5% and delivery stability dropping 7.2%.</p><p>The mechanism makes sense once you sit with it. Faster generation means bigger batches of change. Bigger batches take longer to get through <a href="https://en.wikipedia.org/wiki/Code_review">code review</a>, and they break things more often when they land. Individual developers move quicker. The system around them does not.</p><figure><img alt="Small code blocks merging into one large block stuck at a narrow gate, showing how bigger change batches slow delivery" src="https://cdn-images-1.medium.com/max/1024/1*Xy7A6latCak6pvjv4WecIw.jpeg" /></figure><p>METR ran a controlled study in 2025 with experienced developers, working in codebases they already knew. They took 19% longer with AI tools than without. Beforehand, they had predicted the opposite.</p><p>I would not build a purchasing decision on any single one of these. METR published a follow-up in 2026 suggesting later tools may have reversed the result, while cautioning that the size of the effect was unreliable.</p><p>But the direction is worth noticing. Every one of these findings points at the same place. Not the writing. What comes after it.</p><p>Which is the part the subscription price does not cover.</p><h3>How to work out your own number</h3><p>Start with the seat price, then go looking for the ceiling. Every vendor publishes it somewhere. Credits, tokens, agent requests, whatever the vendor calls it. The name changes, the mechanism does not. That is the number that decides your actual bill.</p><p>Then find out what your team will do to that ceiling. You cannot guess this. Two developers on the same team will burn wildly different amounts, and the heavy user is usually the one working on the messiest part of the system.</p><p>Which is why buying for the whole team on day one is the expensive way to find out. Give it to four or five people for a month, on a real project rather than a trial task, and read the invoice at the end. Then multiply.</p><p>Watch what happens to review time while you do it. If pull requests get bigger and sit longer, you have found the cost that does not appear on any invoice.</p><p>And if you get to the point of hiring a firm rather than buying seats, the same question applies to them. I went through how five development companies describe their own review process in <a href="https://medium.com/@vii.marina/5-best-ai-augmented-development-companies-in-2026-67deee6a4be4">5 Best AI-Augmented Development Companies in 2026</a>. The differences between them were wider than any price gap in this article.</p><h3>So what does it actually cost</h3><p>Sixteen dollars a month buys a seat. It does not buy the thing the marketing is selling.</p><p>What you are actually paying for is a shift in where your engineers spend their day. Less time writing the parts nobody enjoys. More time reading and checking, and deciding whether something that looks right actually is.</p><p>For some teams that trade is obviously worth it. A senior engineer who reads code quickly and knows the system well gets a genuine multiplier. For others it moves the bottleneck somewhere less convenient and costs more than the invoice suggests.</p><p>The invoice, by the way, will not be $16. Find your ceiling, run a pilot, read the real number at the end of the month.</p><p>That number is the one worth arguing about.</p><p><strong><em>All prices taken from vendor pricing pages on 2 September 2026, in US dollars. Subscription pricing in this category changes often.</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fac2eb3a972c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Best AI-Augmented Development Companies in 2026]]></title>
            <link>https://medium.com/@vii.marina/5-best-ai-augmented-development-companies-in-2026-67deee6a4be4?source=rss-41506d6f3981------2</link>
            <guid isPermaLink="false">https://medium.com/p/67deee6a4be4</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Marina]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 14:32:23 GMT</pubDate>
            <atom:updated>2026-09-01T14:32:23.831Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Doxa88_zJ7gP2vJNMxIeqw.jpeg" /></figure><h4>Everyone claims AI makes them faster. Only some can tell you who checks the code afterwards.</h4><p>Almost every software company now says it uses AI. Very few will tell you what that actually means on a Tuesday afternoon, in a real repository, with a real deadline.</p><p>AI-augmented development is a narrower thing than it sounds. The AI sits inside the <a href="https://en.wikipedia.org/wiki/Software_development">software development</a> process rather than in the product. It drafts the boilerplate nobody enjoys writing. It reads legacy systems that were never documented and explains what they do. Engineers still own every decision that matters. Hiring a firm to build you an AI product is a different purchase entirely.</p><p>The distinction matters because the marketing has run ahead of the evidence. One company on this list claims delivery cycles three to five times faster. Another publishes no speed figure at all. Both are describing the same category.</p><p>So the question is not who promises the biggest number. It is who can tell you what happens after the AI writes something.</p><h4>How these five were chosen</h4><p>Three filters, applied in order.</p><p>The first was a verified <a href="https://clutch.co/">Clutch</a> profile <em>rating 4.7 or higher</em>. A public review record takes years to build and is harder to fake than a landing page.</p><p>The second was <em>a real page</em> about the AI-augmented process. Not a paragraph bolted onto a generic services page.</p><p>The third filter did most of the work. Somewhere public, the company <em>had to explain</em> how AI output gets checked before it ships. Plenty of firms will tell you their engineers use Cursor. Far fewer will say who reviews what Cursor produced, or who answers for it at three in the morning.</p><figure><img alt="Five cubes approaching a single gate, the first one stopped at it" src="https://cdn-images-1.medium.com/max/1024/1*pMLgHvmKj_2H_tjJzsM7AA.jpeg" /></figure><h3>1. Acropolium</h3><p><a href="https://clutch.co/profile/acropolium"><strong>Clutch</strong></a><strong>: 5.0 from 30 reviews · $50-99/hr · from $25,000 · Den Haag, Netherlands · building software since 2003 · Clutch data as of September 2026</strong></p><p><a href="https://acropolium.com/services/ai-augmented-development/">Acropolium</a> will not let AI near your production code until it has proven itself on a smaller piece of it first. That is the whole argument for putting them first here.</p><p>The process runs in five stages and stage three is the one to look at. AI gets pointed at a narrow slice of real work, usually test generation or code analysis. That pilot has to produce delivery data that holds up. If it does not hold up, it does not scale.</p><p>The two stages before it are groundwork. The team audits how your software actually gets built today, undocumented constraints included. Then the tooling goes into the IDEs and pipelines your engineers already open every morning. Data access controls get set at the same time, so proprietary code never leaves approved boundaries.</p><p>They measure delivery speed and defect reduction. Test coverage as well.</p><blockquote><a href="https://acropolium.com/services/ai-augmented-development/">In their words:<br>“not abstract productivity claims.”</a></blockquote><p>The caution matches their client list. Fintech and healthcare mostly, with automotive and energy alongside. A bad merge in those places is a regulatory event, not a bug ticket.</p><p>The record behind it is long. 455 delivered projects and 155 clients. ISO 9001 certified delivery, with Fortune 500 names on the roster. Five client relationships have run past ten years.</p><p>Acropolium has been building software since 2003, well before any of this tooling existed.</p><h3>2. STX Next</h3><p><a href="https://clutch.co/profile/stx-next"><strong>Clutch</strong></a><strong>: 4.7 from 102 reviews · $50-99/hr · from $50,000 · Poznań, Poland · founded 2005 · Clutch data as of September 2026</strong></p><p>STX Next leads its page with three to five times faster delivery cycles, set in large type at the top.</p><p>Scroll down the same page and the number changes. Task-level gains reach 55%. End-to-end delivery lands closer to 30% once you count planning and review and the coordination between teams.</p><blockquote>Their phrasing for this is:<br>“the honest number, not the demo number.”</blockquote><p>Both figures live on one page. The second one is the useful one.</p><p>They name the tooling. Claude Code and Cursor for primary development, with Copilot for inline suggestions. On top of that runs AI-assisted pull request review, aimed at catching hallucinated APIs before they turn into rework.</p><p>The team writes test rules from your existing conventions on day one, so generated output matches your quality bar from the first commit forward.</p><p>They frame the whole offering against a phrase they clearly dislike, which is vibe coding.</p><p>They will not start below $50,000, which is more than anyone else here asks for. Behind that price sit 102 verified reviews.</p><h3>3. Merixstudio</h3><p><a href="https://clutch.co/profile/merixstudio"><strong>Clutch</strong></a><strong>: 4.8 from 97 reviews · $50-99/hr · from $25,000 · Poznań, Poland · founded 1999 · Clutch data as of September 2026</strong></p><p>Merixstudio describes itself on Clutch as an AI-augmented software development company. The stated focus is digital products, data-heavy platforms and IoT.</p><p>The profile leads with client outcomes. A custom proposal tool cut sales cycle time at Boston Solar by 33%. At RP Group, an IoT platform for safety lighting removed up to 70% of the manual inspection effort. There is also a modernized Six Flags mobile app that reported 97% fewer crashes.</p><p>Read what those numbers measure. They describe how the delivered software performed for the client. None of them measures AI inside the delivery process, and the company publishes nothing of that kind.</p><p>No public page describes how AI-generated code gets checked before it merges. There is no named owner for that step and no stated policy behind it.</p><p>97 verified reviews is the deepest record among these five.</p><h3>4. PLANEKS</h3><p><a href="https://clutch.co/profile/planeks"><strong>Clutch</strong></a><strong>: 5.0 from 37 reviews · $25-49/hr · from $10,000 · Kyiv, Ukraine and London · founded 2019 · Clutch data as of September 2026</strong></p><p>PLANEKS works mostly on Python backends. AI-assisted coding pipelines and generated tests, with LLM-augmented code review on top. The worked example on the page is a legacy refactor for EuropaFoodXB. One section carries the heading: <em>“When AI-Augmented Development Is the Wrong Choice.”</em></p><p>It names systems under strict compliance requirements as a case where the approach does not fit.</p><p>They will take a project at $10,000, well under anyone else here. The team runs 10 to 49 people.</p><p>PLANEKS started in 2019, a decade after the next youngest company on this list.</p><p>Their Clutch service breakdown reads 100% web development. The profile carries no AI line at all.</p><h3>5. Axon Active</h3><p><a href="https://clutch.co/profile/axon-active"><strong>Clutch</strong></a><strong>: 4.8 from 5 reviews · pricing undisclosed · Ho Chi Minh City, Vietnam · founded 2009 · Clutch data as of September 2026</strong></p><p>Axon Active assigns AI a level of autonomy per task. Well-specified, low-risk work gets delegated further. Anything riskier stays under tighter human control.</p><p>A human reviews every AI output before it merges, and a named engineer owns the outcome. Their page states this plainly. <em>Asked whether AI can ship code without human review, they answer no.</em></p><p>Their measurement framing is unusual. They watch lead time and deployment frequency, plus work in progress. Review time and defect escape rates go in as well. The argument behind it is that faster coding means nothing if reviews become the bottleneck.</p><p>The company publishes no speed claims of its own. The statistics on the page belong to other people, including Gartner’s projection that 75% of enterprise software engineers will use AI code assistants by 2028.</p><p>650+ staff work across four Vietnamese offices.</p><p>Two gaps. They disclose neither hourly rate nor minimum project size anywhere. And their Clutch record runs to 5 reviews, against 102 for STX Next and 97 for Merixstudio.</p><h3>What the research actually says</h3><p>Two numbers explain why the review question matters more than the speed question.</p><p>The first is from <a href="https://survey.stackoverflow.co/2025/ai">Stack Overflow’s 2025 developer survey</a>. 84% use AI tools or plan to. That part is unsurprising. What follows is not: 46% actively distrust the accuracy of what those tools produce, against 33% who trust it. Three percent trust it completely.</p><p>Ask what specifically bothers them and the answer gets sharper. The top frustration, named by 66%, is AI output that is <em>almost right, but not quite</em>. Not wrong, which you catch. Almost right, which you merge.</p><p>The second number comes from <a href="https://dora.dev/ai/gen-ai-report/report/">DORA</a>, Google Cloud’s long-running research program on software delivery. A 25% rise in AI adoption tracks with delivery throughput falling 1.5% and delivery stability falling 7.2%.</p><p>That reads backwards until you see the mechanism. AI lets developers generate code much faster, so change batches get bigger. Bigger batches take longer to review. They also break things more often. The speed shows up in the writing and disappears in everything that comes after.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XlLHOm9ML-K6Gvine07OMw.jpeg" /></figure><p>DORA’s own recommendation follows from that. Reinforce <a href="https://en.wikipedia.org/wiki/Test_automation">automated testing</a>. Speed up <a href="https://en.wikipedia.org/wiki/Code_review">code review</a>. Continuous integration is what catches AI-introduced errors before they reach production.</p><p>That is why the third filter above did the most work. The extra code has to go somewhere, and someone has to look at it.</p><h4>Which one to talk to</h4><p>If your code carries regulatory weight, <a href="https://acropolium.com/services/ai-augmented-development/">start with Acropolium</a>. The pilot stage is the reason. You get delivery data from a narrow slice of your own work before AI touches anything that matters, and nobody else on this list offers that.</p><p>For a large build where you want depth of track record, STX Next and Merixstudio both have close to a hundred verified reviews. STX Next describes its review workflow. Merixstudio does not.</p><p>Axon Active has the most explicit governance model here, with autonomy levels and named accountability per change. Five reviews is a thin record to weigh that against.</p><p>PLANEKS is the cheap entry point, and the only one willing to publish where its own service stops making sense.</p><p>One last thing worth asking whoever you call. Not how much faster they got. Ask what a piece of AI-generated code goes through before it reaches production. The good answers are specific and slightly boring. Those are the ones to trust.</p><p><strong><em>All company figures come from Clutch profiles and the companies’ own public pages, checked in September 2026. Ratings and review counts change.</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67deee6a4be4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>