<?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 Cem Karaca on Medium]]></title>
        <description><![CDATA[Stories by Cem Karaca on Medium]]></description>
        <link>https://medium.com/@cem.karaca?source=rss-61eca6c8ef85------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*i9JNAFCjAVa8iHkerwctPw.jpeg</url>
            <title>Stories by Cem Karaca on Medium</title>
            <link>https://medium.com/@cem.karaca?source=rss-61eca6c8ef85------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 15:16:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@cem.karaca/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[Who Was on the Plane?]]></title>
            <link>https://medium.com/@cem.karaca/who-was-on-the-plane-59115412139d?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/59115412139d</guid>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[epstein-files]]></category>
            <category><![CDATA[flight-logs]]></category>
            <category><![CDATA[transparency]]></category>
            <category><![CDATA[epstein]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Sat, 19 Sep 2026 19:58:59 GMT</pubDate>
            <atom:updated>2026-09-19T20:00:32.442Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FeE_26z4UwwA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DeE_26z4UwwA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FeE_26z4UwwA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/efa489cfb35b1fdc1bf07fc1091e6cc9/href">https://medium.com/media/efa489cfb35b1fdc1bf07fc1091e6cc9/href</a></iframe><p>A year after Congress voted 427 to 1 to open the Epstein files, we still don’t have the names. So I put the question in a song.</p><p>In May 2007, a federal prosecutor in Florida had a draft indictment on her desk. Sixty federal counts, including sex trafficking of minors. Nine of those counts carried mandatory minimums of 15 years to life. Nineteen identified victims, ages 14 to 17. The FBI knew where Epstein would be on May 16 and May 19 and wanted to arrest him on one of those days. <a href="https://epstein-data.com/reports/institutional/DS12_EXPANSION_ANALYSIS.html">epstein-data</a><a href="https://www.nbcnews.com/politics/justice-department/live-blog/epstein-files-trump-doj-release-live-updates-rcna256639">NBC News</a></p><p>The indictment was never filed. Tracked across four successive versions of the prosecution memo, the case shrank. By February 2008, several victims were cut for “strategic reasons,” including the original complainant whose parents first went to police. It was replaced with a non-prosecution agreement and a state plea.</p><p>Here is the part that should keep you up at night: these memos were quietly added to the DOJ’s public database on or around March 5, 2026, with no press release, no update to the disclosure page, and no notification to researchers. Nineteen years after they were written. <a href="https://epstein-data.com/reports/institutional/DS12_EXPANSION_ANALYSIS.html">epstein-data</a></p><p>I wrote “Who Was on the Plane?” because I kept hitting the same wall you probably have. Everyone talks about the case. Nobody answers the question.</p><h4>What we actually know</h4><p>I want to be careful here, because vague accusations are exactly what the people hiding the real names are hoping for. So only the verifiable parts:</p><p>In 2006, a Palm Beach County grand jury indicted Epstein on one state felony charge of solicitation of prostitution. The Palm Beach police chief and lead detective then referred the case to the FBI, saying the charge didn’t reflect “the totality of Epstein’s conduct.” <a href="https://www.kuow.org/stories/jeffrey-epstein-files-tracing-the-legal-cases-that-led-to-sex-trafficking-charges">kuow</a></p><p>In August 2019, the richest inmate in the federal system died in his cell before he could stand trial. The cameras outside his cell weren’t recording, and the guards assigned to check on him admitted they had falsified their logs. Whatever you believe about the cause of death, the institutional failure is not in dispute.</p><p>The House of Representatives voted 427 to 1 to pass the Epstein Files Transparency Act on November 18, 2025. The Department of Justice released more than 3 million pages, 2,000 videos, and 180,000 images, more than a month after the deadline set by Congress. The DOJ says there are redactions but has not provided a full explanation of what is excluded. Those files were heavily redacted, drawing criticism from Democratic and some Republican lawmakers as well as Epstein survivors. <a href="https://en.wikipedia.org/wiki/Epstein_Files_Transparency_Act">Epstein Files Transparency Act +2</a></p><p>And now, September 2026: “It’s been a year since the president signed the Epstein Files Transparency Act, and the government continues to withhold over three million files,” Massie said on the House floor. <a href="https://spectrumnews1.com/ky/louisville/news/2026/09/01/massie-discharge-petition-epstein">Spectrum News</a></p><p>That sentence is the whole song.</p><h4>Why a song</h4><p>Because the documents have been public for months and it changed almost nothing. Three and a half million pages is not transparency. It is a wall built out of paper. You cannot read your way through it on a Tuesday night after work, and that is the point.</p><p>A song is three minutes. A chorus is one question. You can’t skim past a question that a children’s choir is asking you.</p><p>The lines I struggled with most were the ones about performance. “Emmy stage they cry for justice.” “Hashtag prayers, hashtag hope.” Not because they are unfair, but because they are so easy. It costs nothing to say the right thing from a podium. It costs everything to say a name. And the flight logs do not care which party you vote for. “Left and right they scream all day, but at night they fly away.” Some of the people who tweet the loudest about this case would prefer that it stays a story about the other side.</p><h4>What you can actually do</h4><p>This is where most articles about Epstein turn into a mood. I don’t want a mood. I want you to do one of these this week.</p><p><strong>1. Check whether your representative signed.</strong> The discharge petition means the bill is guaranteed a vote on the House floor, but it remains to be seen whether it will pass and whether it will have enough support to make it through the Senate. This congressional term ends in early January, and Massie will leave Congress when his term expires. The list of who signed is public. If your representative’s name isn’t on it, call and ask why. It is a five-minute call, and staffers count them. <a href="https://www.forbes.com/sites/alisondurkee/2026/09/18/new-epstein-files-bill-will-go-up-for-a-vote-when-house-resumes-heres-what-it-would-do/">Forbes</a></p><p><strong>2. Understand what Act II changes.</strong> It would allow state attorneys general, local prosecutors, and victims to sue the U.S. attorney general over alleged unlawful withholdings, redactions, or failure to produce records, and would require the attorney general to provide certain unredacted documents to victims upon request. If government officials withhold, alter, or misrepresent any Epstein files, the bill would make that a criminal offense. In plain language: it moves the fight from a federal system that has failed for twenty years to state courts and to the survivors themselves. <a href="https://www.axios.com/2026/09/18/epstein-files-vote-massie-khanna-support">Axios</a><a href="https://www.forbes.com/sites/alisondurkee/2026/09/18/new-epstein-files-bill-will-go-up-for-a-vote-when-house-resumes-heres-what-it-would-do/">Forbes</a></p><p><strong>3. Read one primary document yourself.</strong> Not a thread about a document. The DOJ repository is searchable, and independent researchers have indexed it. Pick the May 2007 prosecution memo. It is short. Then ask yourself why you are reading it in 2026.</p><p><strong>4. Amplify survivors, not celebrities.</strong> The women who testified as teenagers wrote it all down, and someone made it disappear. When you share this story, share their words and their organizations. The people on the stage have publicists.</p><p><strong>5. Share the question, not a theory.</strong> Every unsourced name you post is a gift to the people protecting the real ones. It lets them call the whole thing a hoax. Stick to what can be proven. The proven parts are damning enough.</p><h4>The refrain</h4><p>I’m not American. I’m a Turkish engineer who makes music, and I wrote this in English because the people who need to hear the question are the ones with the files. But the question is not really about one country. Every country has a plane.</p><p>Where are the names? Where is the list?</p><p>Redacted, retracted. We’re still asking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=59115412139d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Three AI code review bots, 192 pull requests, 18 days]]></title>
            <link>https://medium.com/@cem.karaca/three-ai-code-review-bots-192-pull-requests-18-days-9ed5841bd923?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ed5841bd923</guid>
            <category><![CDATA[code-review]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 22:44:12 GMT</pubDate>
            <atom:updated>2026-09-18T22:44:12.890Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7HUnC89vytrF6QAm0d_Azw.jpeg" /><figcaption>Futuristic dashboard comparing three AI code review bots — CodeRabbit, Sentry Seer, and Sourcery — showing PR coverage and critical bug metrics.</figcaption></figure><p><em>CodeRabbit, Sentry Seer, and Sourcery reviewed the same code side by side for $169 a month. Here is who found the critical bugs, and what each bot’s silence turned out to mean.</em></p><p>Trying a code review bot is easy. Install it, read what it writes on two pull requests, say “not bad.” Measuring one is hard, because only you know whether what it found was a bug, and you never know whether what it missed was there at all.</p><p>On August 30, we added a third bot to our repository. From that day on, three bots reviewed every pull request at the same time, under the same conditions. Eighteen days later we had 192 pull requests, 816 inline findings, and one question: which of them finds the bugs that matter?</p><p>Short version: CodeRabbit wins on raw count, Seer wins on precision per finding, and Sourcery’s low numbers have nothing to do with looking badly. It simply never looked at most of the pull requests. The longer version follows, including two places where our own measurement fooled us.</p><h3>What we were measuring against</h3><p>The repository is a multi-tenant operations platform on Next.js and PostgreSQL: customer portal, invoicing, card and balance payments, domain and server orders, an accounting integration, and an AI agent that manages servers. Some of the bugs here are denominated in money. Others are tenant boundaries. That is the lens we used when deciding whether a bot’s “critical” was really critical.</p><p>In the window, 192 pull requests opened and closed, and 166 commits reached main. The median PR changes 514 lines (the mean is 5,101, because a few are enormous) and goes from open to merge in 62 minutes. One developer and an AI coding agent wrote the PRs. There is also an internal review chain: a per-task agent review plus a whole-branch review before merge, and the bots sit alongside that chain rather than replacing it.</p><p>The three bots: <strong>CodeRabbit</strong> (added August 30, which is why the window starts there), <strong>Sentry’s Seer</strong> (its PR review feature), and <strong>Sourcery</strong>. All three run automatically on every PR.</p><p>What we pay, per month: CodeRabbit <strong>$60</strong>, Sentry <strong>$79</strong> (Seer comes with the Sentry plan, so that figure buys error tracking as well), Sourcery <strong>$30</strong>.</p><p>$169 in total.</p><p>Quotas and review limits vary by plan tier; the limits quoted here are the ones we hit on these plans.</p><h3>How we counted</h3><p>Everything came from the GitHub API: inline comments, review rounds, conversation comments, and the check runs on each PR’s final commit. We split comments by bot user name and classified each one by the bot’s own markers.</p><p>A finding is an inline comment the bot itself marks as a finding. CodeRabbit puts a category and severity header on each (“Functional Correctness | Major”) and a hidden marker in the body. Sourcery labels comments issue (bug_risk), issue (security), suggestion and so on. Seer opens each with Bug: and a level like Severity: HIGH. Bulk lists inside summary comments do not count; CodeRabbit&#39;s summaries alone carry 60 nitpicks and 70 &quot;outside the diff&quot; items that are not in these numbers.</p><p>Fixed means the bot’s own stamp. All three mark their comment when a push containing a fix arrives: CodeRabbit writes “Addressed in commits”, Seer “Resolved in”, Sourcery “Addressed in”. We marked nothing by hand, so fix counts are a lower bound. A fix that lands in a different PR, or one the bot fails to match, is not counted.</p><p>Critical was the hard one. The bots’ scales do not line up: CodeRabbit almost never uses its top level; Seer’s CRITICAL means something different; Sourcery has no severity scale at all. So we did a second pass by hand. We read every fixed finding on the money and identity paths (billing, payments, accounting, authentication, authorization) and counted by our own definition: money moving wrongly, a tenant boundary crossed, consent and legal records, data loss, outage. That count is subjective, which is why it appears below as an approximation.</p><p>The limits are plain. One repository, one team, 18 days, the bots’ own stamps, our own reading of critical. This is a field measurement, not a benchmark.</p><h3>The numbers</h3><pre>Metric                             CodeRabbit     Seer           Sourcery<br>Monthly price                      $60            $79 (Sentry)   $30<br>PRs reviewed                       167            ~168           90<br>Not reviewed (quota/limit/draft)   4              9 (never ran)  82<br>Inline findings                    523            153            140<br>PRs with findings                  127            78             70<br>Bot&#39;s own &quot;fixed&quot; stamp            318 (61%)      122 (80%)      91 (65%)<br>Findings per 1,000 diff lines      0.58           0.17           0.39<br>First comment, median / p90        1.1 / 1.6 min  5.2 / 62 min   0.1 / 0.1 min<br>Findings we explicitly rejected    13             5              4</pre><p>Severity mix, with the fixed stamp in parentheses. <br>CodeRabbit: Critical 6 (4), Major 268 (151), Minor 249 (163). <br>Seer: Critical 7 (5), High 45 (36), Medium 66 (51), Low 35 (30). <br>Sourcery has categories rather than levels: bug_risk 73 (48), broader_impact 27 (18), security 18 (8), testing 7 (4), other 15 (13).</p><p>In its own taxonomy, CodeRabbit labeled 179 findings as functional correctness, 139 as data integrity, 107 as security and privacy, 52 as stability, and 45 as maintainability. Performance showed up once.</p><p>Look at the rejection row. Of CodeRabbit’s 523 findings, 132 got a reply from us but no fix stamp afterwards. That sounds like 132 rejections. It is not: most of those replies are discussion, a question back, or “handled elsewhere.” The findings we explicitly told the bot were wrong come to 13. Seer has 5 of those, Sourcery 4. False positives were not a problem for any of the three.</p><h3>Who found the critical bugs?</h3><p>Reading the fixed findings on the money and identity paths one by one: CodeRabbit had 69 findings there, roughly 25 critical by our definition. Seer had 33, roughly 12 critical. Sourcery had 14, roughly 8 critical, plus two security findings outside those paths (cross-tenant memory leakage and a shell separator) that meet the same bar, so call it 10.</p><p>CodeRabbit leads on absolute count, by about a factor of two. Seer leads on density: the highest share of critical per finding and an 80 percent fix rate. Sourcery, where it looked, sits just behind Seer on critical per finding. Its problem is coverage, which we cover below.</p><p>The cases say it better than the counts. All were fixed. None reached production.</p><p>The card was charged, and nothing was recorded. In the saved-card flow, the card is charged at the provider, and then the payment row is written. If the write failed, the finally block released the charge lock, leaving neither a record nor a retry path: money taken from the customer, no trace of it in the system. Seer marked it CRITICAL. In the same file, CodeRabbit found something else entirely: an expired lock handle that could release a newer lock acquired after it. Two bots, two different holes, one module.</p><p>The same service was invoiced twice. When a customer with a managed contract also buys a self-invoicing service, the contract renewal cron has to exclude that service’s line item. A new service type was added, and the exclusion query wasn&#39;t updated. Seer caught it in the renewal route. This is the “forgotten sibling” class of bug: fix applied in one place, the second place carrying the same rule skipped.</p><p>Absence of consent was treated as consent. The code that sends ad conversions sent them whenever marketing consent was not false. null, meaning no decision recorded, passed. Seer marked it HIGH. For us it is critical, because it is a legal record.</p><p>The confirmation screen showed the list price while the discounted amount was charged. In the domain order flow, the charge happens at the discounted amount, but the pricing object returned to the customer carried values computed before the discount. We measured up to a ninety-nine percent gap between what the screen showed and what was charged. CodeRabbit found it in its first round and labeled it Major.</p><p>A plan document was about to run a migration against production. An implementation plan contained the step “run drizzle:migrate to verify&quot;; the top of the same document said this machine&#39;s database connection is production. CodeRabbit put the two paragraphs side by side and gave it a Major. Not a code bug, but an instruction that would have cost more than one had anyone followed it.</p><p>The hole in the shield. We have a scan test that stops identity rows (user tables carrying the 2FA secret) from leaking to the client. Sourcery showed that the test’s regex only matched .select().from(users) separated by whitespace, and that an intervening comment or a schema alias walked straight past it. The shield existed. It could be walked around.</p><p>Cross-tenant memory leakage, failing open. In the memory layer of the server-managing agent, if the query that reads the company’s server inventory failed, it returned an empty list, and the filter read the empty list as “no scope” and injected other servers’ memories into this server’s task. One of Sourcery’s issue (security) findings. Another from the same class: the URL grammar for a curl command run over a remote shell accepted &amp;, which in a shell is a command separator.</p><p>We write in commit bodies whose finding we are closing. Of 166 commits, 63 credit CodeRabbit, 36 credit Seer, 30 credit Sourcery.</p><h3>Silence means three different things</h3><p>This is the most useful thing we got out of the measurement, and it is not a bug count.</p><p>Seer says nothing when it is clean. Seer did not appear on 101 of 179 pull requests. Check runs showed “No issues found” on 90 of them and no run at all on 9. On 42 of those 90 “clean” PRs, CodeRabbit had a fixed-stamped finding; on 20 of them, Sourcery did. So Seer saying clean does not mean the PR is clean. It means Seer found nothing in that round. More than once, re-triggering it on the same commit produced new findings. On the ad-conversion PR, two of four Seer rounds carried findings and two came back empty.</p><p>Sourcery never looked at 82 pull requests. Seventy-eight of its review comments were this one sentence: “this account has used its review budget of 2,500,000 diff characters for the last 7 days.” Three more PRs hit the single-PR size limit of 500,000 diff characters. So the Sourcery that appeared on 171 PRs actually reviewed 90. On 20 of those 90 it posted no finding, and on 4 of those 20 CodeRabbit had a fixed-stamped finding, on 3 Seer did. Sourcery’s real weakness is coverage, not accuracy. And because the budget is a rolling seven-day window, one large PR eats the following days’ reviews too. That is what $30 buys: the cheapest seat at the table, and the smallest window.</p><p>CodeRabbit goes quiet too, but rarely. On 36 of 167 PRs it passed without a comment; on 5 of those Seer had a finding, on 4 Sourcery did. Its quota (“Review limit reached”) kicked in on 6 PRs; on 5 it resumed minutes later, on 1 the PR went unreviewed that round. Two draft PRs and one not targeting main went unreviewed under default settings.</p><p>One number sums up the table: PRs where only one bot found anything. CodeRabbit 32, Seer 6, Sourcery 6. CodeRabbit and Seer together carry nearly all the coverage. Sourcery’s six were things nobody else saw.</p><h3>Where we got it wrong</h3><p>Our first read said “Sourcery did not hit its quota in this window.” The classifier was searching for the string “rate limit”. Sourcery says “review budget.” Same day, second mistake: Sourcery’s security findings begin with 🚨 issue (security), the emoji did not pass the regex, and 18 findings went uncounted.</p><p>The lesson: when you measure a bot’s absence, search for the bot’s own refusal sentence. When you count findings, read the bot’s current format. We got both wrong and corrected them. The numbers here are the corrected ones.</p><h3>Three characters</h3><p>Seer has a nose for money. It is best on state machines: partial success, retries, idempotency, two currencies added together by accident. It gave three consecutive HIGH findings in the split-payment logic of the accounting integration, and all three were real. It also found a pre-existing bug that was not in the diff, worth saying explicitly because some bots only look at changed lines. Its weaknesses: a slow queue (p90 of 62 minutes, which equals our median merge time), silence when clean, and different results from round to round on the same commit. It barely looks at tests. Of its 122 fixed-stamped findings, one is in a test file.</p><p>CodeRabbit is volume and breadth. It looks outside the diff (70 items), reads documents, plans, and runbooks (79 of its 318 fixed findings are under docs and scripts), and checks its own claims by running scripts against the repository; under many comments there is a “Supported by static analysis” block with the rg commands it ran. It tags security findings with CWE identifiers. It says at least one thing on every PR, which drives both its coverage and its noise. The price is reading volume: with nitpicks and outside-the-diff lists in its summaries, it produced more than 650 items in 18 days. One warning about labels: Minor sometimes carries money. Code that read only the first page of the accounting API and never saw an open invoice on page two, and two parallel --execute processes applying the same delta twice, drifted between Minor and Major. Read the body, not the label.</p><p>Sourcery is the shield’s shield. It is the best of the three at finding gaps in regression and scan tests, and at finding swallowed errors: catch(() =&gt; {}), an empty catch block, a cron whose alarm never fires. Its issue (security) class carries real findings: the cross-tenant leak, the shell separator, an unaudited IDOR attempt. Its Reviewer&#39;s Guide arrives within seconds and is a decent map for starting to read a PR. The price is the budget.</p><h3>What we do now</h3><p>Four rules came out of this.</p><p>Two bots are mandatory. CodeRabbit and Seer together carry nearly all the coverage. Sourcery is a third set of eyes while we have the budget.</p><p>On money and identity PRs, trigger Seer a second time, even when the check run says clean. The second round catches what the first missed.</p><p>Do not read silence as clean. It means three different things across three bots: budget refusal, clean run, never ran. Only the check run and the bot’s own message tell them apart.</p><p>The bots do not replace the review chain. Our internal chain found bugs all three bots missed, and the bots found bugs the chain missed. Both are necessary. Neither is sufficient alone.</p><p>We are keeping all three. $169 a month for the set; the unrecorded card charge alone would have cost more than a year of that, and that is one finding out of roughly 45 we counted as critical.</p><p>And one point of merge discipline: not a single finding in the window arrived after a merge, because we wait for the bots. Given Seer’s p90, that sometimes means an hour. An hour’s wait is cheaper than an unrecorded card charge.</p><h3>Run it on your own repo</h3><p>The measurement is a few hundred lines of script and the GitHub API.</p><ol><li>Pick the window: first bot’s first comment to today.</li><li>For each PR, fetch pulls/{n}/comments (inline), pulls/{n}/reviews, issues/{n}/comments and the final commit&#39;s commits/{sha}/check-runs.</li><li>Split comments by bot user name. Recognize findings by the bot’s own marker: CodeRabbit’s cr-indicator-types marker and severity header, Sourcery&#39;s **issue (...)** label (allow an emoji prefix), Seer&#39;s Bug: and Severity:.</li><li>Read fixes from the bot’s stamp. Do not mark them yourself.</li><li>Search for absence with the bot’s own refusal sentence: “review budget”, “Review limit reached”, “Draft PR not reviewed”, “No issues found” in the check run.</li><li>Join with PR metadata: lines changed, opened and merged times. Findings per 1,000 lines and latency fall out of that.</li><li>For severity, ignore the bots’ scales. Write your own definition of critical and read the findings.</li></ol><p>The number that will help you most is probably the one that helped us: the coverage asymmetry, which bot never looked at which PR. Saying “this bot finds less” without that number is doing the measurement in the bot’s place.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ed5841bd923" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My CLAUDE.md Was Eating 42,000 Tokens Per Conversation — Here’s How I Fixed It]]></title>
            <link>https://medium.com/@cem.karaca/my-claude-md-was-eating-42-000-tokens-per-conversation-heres-how-i-fixed-it-85ffba809bd4?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/85ffba809bd4</guid>
            <category><![CDATA[token-efficiency]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[claude-skills]]></category>
            <category><![CDATA[claude-md]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Mon, 16 Feb 2026 21:31:31 GMT</pubDate>
            <atom:updated>2026-02-16T21:32:40.681Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JhDq81d5DSILwkmrVd218w.jpeg" /><figcaption>How I converted a 1,200-line monolith into modular skills and cut my Claude Code costs by 83%</figcaption></figure><h3>How I converted a 1,200-line monolith into modular skills and cut my Claude Code costs by 83%</h3><p>I hit my Claude Code usage limit for the first time this week. Not because I was doing anything unusual — same project, same workflow, same pace. So I dug into why.</p><p>The answer was embarrassing: my CLAUDE.md file had quietly grown to 1,207 lines and was consuming <strong>~42,200 tokens every single conversation</strong>. Every time I asked Claude to fix a typo, it was loading my entire database schema, all my Docker troubleshooting guides, my PCI-DSS compliance rules, webhook handler documentation, and 18 integration specs. For a typo.</p><p>This is the story of how I turned that monolith into a slim 200-line file plus 7 on-demand skills — and why you should do the same before your CLAUDE.md quietly bankrupts your token budget.</p><h3>The Monolith Problem Nobody Talks About</h3><p>If you use Claude Code seriously, you know CLAUDE.md is magic. It&#39;s your project&#39;s brain — the file that tells Claude everything about your stack, your conventions, your architecture. Every conversation starts by loading it into context.</p><p>And that’s exactly the problem.</p><p>As your project grows, so does CLAUDE.md. You add your database migration workflow after the third time Claude suggests raw SQL. You add Docker troubleshooting after a painful production debugging session. You add your Git conventions, your security rules, your integration specs. Each addition is sensible in isolation. But compound them over months and you get a beast.</p><p>Here’s what my trajectory looked like:</p><pre>Month 1:   ~150 lines  →  ~2,000 tokens   (lean and mean)<br>Month 3:   ~400 lines  →  ~8,000 tokens   (getting chunky)<br>Month 6:   ~800 lines  →  ~20,000 tokens  (should probably trim this)<br>Month 9:  ~1,200 lines →  ~42,200 tokens  (why am I hitting limits?)</pre><p>At 42K tokens per conversation, even a 20-conversation day costs <strong>~844,000 tokens</strong> just on context loading — before Claude writes a single line of code. That’s roughly 40% of your context window consumed before any actual work happens.</p><p>And here’s the insidious part: <strong>the code quality was excellent</strong>. Claude had all the context it needed. Migrations were flawless. API routes followed every convention. Security rules were always respected. The monolith <em>worked</em> — it just worked expensively.</p><h3>Enter Skills: The Lazy-Loading Pattern for AI Context</h3><p>While researching solutions, I found Claude Code’s skill system — and specifically Anthropic’s “<a href="https://www.anthropic.com/research/agent-skills">The Complete Guide to Building Skills for Claude</a>.” The concept is elegant: instead of loading everything upfront, you package task-specific workflows into modular folders that Claude loads <strong>only when relevant</strong>.</p><p>Think of it as the difference between importing your entire codebase versus importing only the modules you need. Same principle, applied to AI context.</p><p>Skills use a three-level progressive disclosure model:</p><pre>Layer 0: CLAUDE.md (always loaded)<br>         → Universal rules, stack overview, critical gotchas<br>         → ~1,900 tokens</pre><pre>Layer 1: Skill body (loaded on invocation)<br>         → Step-by-step workflow for a specific task<br>         → ~750 tokens average per skill</pre><pre>Layer 2: References (loaded within skills)<br>         → Deep-dive details, compliance checklists<br>         → ~700 tokens average per reference</pre><p>You only pay for what you use. A simple “fix this CSS bug” conversation loads Layer 0 only. A “create a new database migration” conversation loads Layer 0 + the db-migrate skill. A &quot;add a secret with full compliance documentation&quot; conversation loads all three layers — but only for secrets.</p><h3>The Conversion: From 1,207 Lines to Under 200</h3><h3>What I Kept in CLAUDE.md</h3><p>The new CLAUDE.md contains <strong>only what&#39;s relevant to every conversation</strong>:</p><ul><li><strong>Stack summary</strong> (2 lines) — Framework, DB, styling, testing</li><li><strong>Essential commands</strong> (6 lines) — dev, build, test, migrate</li><li><strong>Architecture overview</strong> (15 lines) — Route groups, database split, middleware</li><li><strong>Critical NEVER/ALWAYS rules</strong> (15 lines) — The rules that prevent disasters</li><li><strong>Common gotchas</strong> (5 lines) — Things Claude consistently gets wrong without guidance</li><li><strong>Git workflow</strong> (5 lines) — Branch naming, commit format</li><li><strong>Reference table</strong> (10 lines) — Pointers to deeper docs</li></ul><p>Total: <strong>~72 lines, ~1,900 tokens</strong>. Down from 1,207 lines and 42,200 tokens.</p><h3>What Became Skills</h3><p>Every section that was only relevant during a specific type of task became a skill:</p><p>Skill Purpose Triggers When You Say… add-secret SOPS + Docker Secrets workflow &quot;add secret&quot;, &quot;new credential&quot;, &quot;store API key&quot; api-route Next.js API route scaffolding &quot;create API route&quot;, &quot;new endpoint&quot; complete-feature Pre-PR quality checklist &quot;ready for PR&quot;, &quot;complete feature&quot; db-migrate Database migration workflow &quot;add table&quot;, &quot;new column&quot;, &quot;run migration&quot; new-action AI-first action creation &quot;new action&quot;, &quot;create action&quot; new-smart-action State machine action patterns &quot;smart action&quot;, &quot;multi-step action&quot; register-ai-tool AI tool registration &quot;register tool&quot;, &quot;add AI tool&quot;</p><p>Each skill is self-contained: it has everything Claude needs to complete that specific task without reading the rest of the project documentation.</p><h3>What Became References (Layer 2)</h3><p>Some skills had sub-sections that were only needed in edge cases:</p><p>Reference File Parent Skill Loaded When… security-compliance.md add-secret Full PCI-DSS validation needed full-checklist.md complete-feature Comprehensive pre-merge review migration-patterns.md db-migrate Complex migration patterns (enums, JSON, relations)</p><p>These reference files live inside each skill’s references/ directory and are only loaded when the skill itself determines they&#39;re needed.</p><h3>The Numbers Don’t Lie</h3><h3>Baseline Cost Per Conversation</h3><p>Component Before After CLAUDE.md content ~42,200 tokens ~1,900 tokens Skill descriptions (frontmatter) — ~500 tokens <strong>Total always-loaded</strong> <strong>~42,200 tokens</strong> <strong>~2,400 tokens</strong> <strong>Reduction</strong> — <strong>94%</strong></p><h3>Real-World Scenarios</h3><p>What I’m Doing Before After Savings Fix a CSS bug 42,200 2,400 <strong>94%</strong> Create an API route 42,200 2,400 + 750 = 3,150 <strong>93%</strong> Add a secret + compliance check 42,200 2,400 + 750 + 700 = 3,850 <strong>91%</strong> Database migration with patterns 42,200 2,400 + 750 + 700 = 3,850 <strong>91%</strong> Worst case: all 7 skills + all refs 42,200 2,400 + 5,265 + 2,104 = 9,769 <strong>77%</strong></p><p>The worst case — every skill and every reference loaded simultaneously — still saves 77%. But that never happens in practice. Most conversations invoke 0–1 skills.</p><p><strong>Projected monthly savings</strong>: At ~30 conversations/day, that’s roughly <strong>1.2 million tokens saved daily</strong> on context loading alone.</p><h3>The Unexpected Benefit: Consistency in Routine Tasks</h3><p>Here’s something I didn’t anticipate: skills made routine tasks <em>dramatically</em> more consistent.</p><p>Before skills, my CLAUDE.md had all the rules for creating an API route buried in a 1,200-line document. Claude would usually follow them — but &quot;usually&quot; isn&#39;t &quot;always.&quot; Sometimes it would miss the audit logging. Sometimes it would forget the Zod validation schema. The signal was diluted by noise.</p><p>With a dedicated api-route skill, Claude gets <strong>focused, sequential instructions</strong> with zero noise. The skill says: &quot;Step 1: Create the route file. Step 2: Add Zod validation. Step 3: Add audit logging. Step 4: Add to the API index.&quot; Every API route now follows the exact same pattern, every time.</p><p>Same story for database migrations. My db-migrate skill includes the exact command sequence, the naming convention, and a pointer to references/migration-patterns.md for complex cases. No more &quot;I forgot to run pnpm db:generate before pnpm db:migrate&quot; moments.</p><p>This is the real power of skills: they don’t just save tokens — they <strong>encode your team’s best practices into repeatable, deterministic workflows</strong>.</p><h3>Anatomy of a Good Skill</h3><p>Let me walk through a real example. Here’s the structure of my add-secret skill:</p><pre>.claude/skills/add-secret/<br>├── SKILL.md                          # 98 lines — the workflow<br>└── references/<br>    └── security-compliance.md        # 78 lines — deep compliance details</pre><h3>The Frontmatter (Most Important Part)</h3><pre>---<br>name: add-secret<br>description: Adds a new secret following SOPS + Docker Secrets<br>  workflow with PCI-DSS compliant permissions. Use when user says<br>  &quot;add secret&quot;, &quot;new credential&quot;, &quot;add API key&quot;, or &quot;store password&quot;.<br>---</pre><p>The description field is everything. It determines whether Claude loads your skill. Get it wrong and your skill never triggers. Get it too broad and it triggers on irrelevant tasks.</p><p><strong>The formula</strong>: [What it does] + [When to use it] + [Trigger phrases in quotes]</p><p>The trigger phrases should match what you’d naturally say. Think about the 3–4 ways you’d start a conversation about this task and include those exact phrases.</p><h3>The Body (Action-Oriented Steps)</h3><p>The SKILL.md body reads like a checklist, not a reference manual:</p><pre># Add Secret — PCI-DSS Compliant Workflow</pre><pre>## Step 1: Create the secret file<br>echo -n &quot;your-value-here&quot; &gt; secrets/&lt;name&gt;<br>chmod 600 secrets/&lt;name&gt;</pre><pre>## Step 2: Encrypt with SOPS<br>sops --encrypt secrets/&lt;name&gt; &gt; secrets/&lt;name&gt;.enc</pre><pre>## Step 3: Update Docker Compose<br>[specific instructions for docker-compose.prod.yml]</pre><pre>## Step 4: Update secrets init script<br>[specific instructions for secrets-init.sh]</pre><pre>## Step 5: Use in code<br>import { getSecret, requireSecret } from &#39;@/lib/secrets&#39;;<br>const value = requireSecret(&#39;SECRET_NAME&#39;);</pre><pre>## Step 6: Validate<br>Run security validation. For detailed compliance checks,<br>consult references/security-compliance.md.</pre><p>Notice the progressive disclosure at Step 6: the skill body handles the happy path, and deep compliance documentation lives in references/ for when it&#39;s actually needed.</p><h3>How to Convert Your Own CLAUDE.md (Step-by-Step)</h3><h3>Step 1: Audit Every Section</h3><p>Read your CLAUDE.md and tag each section:</p><p>Tag Rule Action <strong>UNIVERSAL</strong> Relevant in every conversation Keep in CLAUDE.md <strong>TASK-SPECIFIC</strong> Only relevant during a specific workflow Convert to skill <strong>DEEP-DIVE</strong> Reference material within a workflow Move to skill’s references/ <strong>OBSOLETE</strong> Fixed issues, old workarounds Delete</p><p>My 1,207-line file broke down roughly as: 15% universal, 60% task-specific, 15% deep-dive reference, 10% obsolete.</p><h3>Step 2: Design Your Skills</h3><p>For each task-specific section, define:</p><ol><li><strong>Name</strong> (kebab-case): db-migrate, add-secret, api-route</li><li><strong>Trigger phrases</strong>: 3–5 natural phrases a developer would say</li><li><strong>Steps</strong>: The sequential workflow, action-oriented</li><li><strong>References</strong>: Any deep-dive content that can be deferred</li></ol><h3>Step 3: Create the File Structure</h3><pre>.claude/skills/<br>├── your-skill-name/<br>│   ├── SKILL.md              # Under 150 lines<br>│   └── references/           # Optional deep-dives<br>│       └── detailed-guide.md</pre><p><strong>Critical naming rules</strong> (from Anthropic’s guide):</p><ul><li>Folder: kebab-case only (my-skill ✅, My_Skill ❌)</li><li>File: Exactly SKILL.md (case-sensitive)</li><li>No README.md inside skill folders</li></ul><h3>Step 4: Write the YAML Frontmatter</h3><pre>---<br>name: your-skill-name<br>description: What it does. Use when user says &quot;trigger phrase 1&quot;,<br>  &quot;trigger phrase 2&quot;, or &quot;trigger phrase 3&quot;.<br>---</pre><p>Keep the description under 1,024 characters. No XML angle brackets. Include both the <em>what</em> and the <em>when</em>.</p><h3>Step 5: Slim Down CLAUDE.md</h3><p>Delete everything that moved to skills. What remains should answer: “What does Claude need to know in <em>every single conversation</em>?”</p><p>Target: under 100 lines. If a section is only relevant 20% of the time, it’s a skill, not a universal rule.</p><h3>Step 6: Test Trigger Phrases</h3><p>Start a new conversation and say each trigger phrase. Verify the right skill loads. If it doesn’t, adjust the description. If the wrong skill loads, make descriptions more specific or add negative triggers:</p><pre>description: Processes PDF legal documents for contract review.<br>  Do NOT use for general document editing (use docx skill instead).</pre><h3>The Prompt: Bootstrap Skills for Any Project</h3><p>Here’s the prompt I wish I’d had when I started. Copy this into Claude Code and it will analyze your CLAUDE.md and generate a skills conversion plan:</p><pre>Analyze my CLAUDE.md file and help me convert it into modular Claude Code skills.</pre><pre>For each section in CLAUDE.md:<br>1. Classify it as UNIVERSAL (keep in CLAUDE.md), TASK-SPECIFIC (convert to skill),<br>   DEEP-DIVE (move to references/), or OBSOLETE (delete)<br>2. For task-specific sections, suggest:<br>   - Skill name (kebab-case)<br>   - 3-5 trigger phrases<br>   - Whether it needs a references/ subdirectory<br>3. Generate the slim CLAUDE.md (under 100 lines) containing only universal context<br>4. For each proposed skill, generate:<br>   - Complete SKILL.md with YAML frontmatter<br>   - Any references/ files needed</pre><pre>Rules:<br>- SKILL.md files must be under 150 lines<br>- Frontmatter description must include WHAT and WHEN (with trigger phrases in quotes)<br>- Skills must be self-contained (completable without reading CLAUDE.md)<br>- Use progressive disclosure: core workflow in SKILL.md, deep details in references/<br>- Folder names in kebab-case, file must be exactly SKILL.md (case-sensitive)</pre><pre>Start by reading my CLAUDE.md and presenting the classification table before generating any files.</pre><p>Paste this, point Claude at your CLAUDE.md, and you&#39;ll have a conversion plan in minutes.</p><h3>Skills Scale — Your Monolith Doesn’t</h3><p>The most important property of skills is that <strong>adding new ones doesn’t increase baseline cost</strong>. My 7 skills add zero tokens to conversations that don’t need them. If I add 7 more next month, the baseline stays at ~2,400 tokens.</p><p>Compare that to the monolith approach: every new section increases the cost of every future conversation, forever. It’s O(n) versus O(1).</p><pre>Monolith:  Cost per conversation = f(total documentation)     → grows forever<br>Skills:    Cost per conversation = f(relevant documentation)  → grows only with relevance</pre><p>This is why I’m convinced skills should be set up at the <strong>start</strong> of any project, not retrofitted later. Begin with a lean CLAUDE.md and create skills as your workflows crystallize. It&#39;s much easier than untangling a 1,200-line monster after the fact.</p><h3>Tips From the Trenches</h3><p><strong>1. The description field is your skill’s entire marketing copy.</strong> Claude decides whether to load a skill based solely on the frontmatter description. Spend more time on those 2–3 sentences than on the rest of the skill combined.</p><p><strong>2. Test with paraphrases, not just exact phrases.</strong> Don’t just test “add secret” — try “I need to store a new API key”, “where do credentials go”, “how do I add a password to the project.” If these don’t trigger your skill, your description needs work.</p><p><strong>3. One skill, one job.</strong> Don’t create a “database” skill that handles migrations, schema design, seed data, and troubleshooting. Create db-migrate, db-seed, db-troubleshoot. Focused skills trigger more accurately and provide better instructions.</p><p><strong>4. References are for the 20% case.</strong> If a section is needed in every invocation of a skill, keep it in SKILL.md. If it’s only needed sometimes (compliance details, edge case patterns, advanced options), move it to references/.</p><p><strong>5. Review skills after incidents.</strong> Every time Claude does something wrong, ask: should this be a rule in CLAUDE.md (universal) or a step in a skill (task-specific)? Routing the lesson to the right place prevents both monolith bloat and skill gaps.</p><p><strong>6. Don’t forget to delete from CLAUDE.md.</strong> The most common mistake in conversion is duplicating content — moving it to a skill but forgetting to remove it from CLAUDE.md. After conversion, do a final pass and ask: “Is everything in this file relevant to <em>every</em> conversation?”</p><h3>The Bottom Line</h3><p>My CLAUDE.md was a success story that became a bottleneck. It made Claude incredibly accurate — at the cost of eating my entire weekly token budget in 4 days. Converting to skills gave me the same accuracy (better, actually, due to focused instructions) at a fraction of the cost.</p><p>The numbers:</p><ul><li><strong>Before</strong>: 42,200 tokens loaded every conversation, 100% of the time</li><li><strong>After</strong>: 2,400 tokens baseline, additional context loaded only when needed</li><li><strong>Typical savings</strong>: 91–94% per conversation</li><li><strong>Worst case savings</strong>: 77% (all skills + all references loaded)</li><li><strong>Bonus</strong>: More consistent output on routine tasks</li></ul><p>If your CLAUDE.md is over 200 lines, you&#39;re paying a tax on every conversation. The fix takes an afternoon. The savings compound forever.</p><p><em>Cem Karaca is the founder of VeriTeknik, a cybersecurity consulting firm, and Plugged.in, an AI agent infrastructure platform. He writes about AI engineering, developer tooling, and the intersection of infrastructure and intelligence.</em></p><p><em>Follow me on </em><a href="https://medium.com/@cem.karaca"><em>Medium</em></a><em> for more on AI-powered development workflows.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=85ffba809bd4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Going to University in the New Age: Is It Still Worth It?]]></title>
            <link>https://medium.com/@cem.karaca/going-to-university-in-the-new-age-is-it-still-worth-it-ecfc32881bbc?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/ecfc32881bbc</guid>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Tue, 10 Feb 2026 17:36:37 GMT</pubDate>
            <atom:updated>2026-02-10T17:36:37.641Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C4qz_m81SU02uoFkgP3Iwg.jpeg" /><figcaption>Information is not a secret anymore</figcaption></figure><p><em>As an engineer who studied in the early 2000s, I’ve watched the democratization of knowledge reshape everything — except our outdated beliefs about education.</em></p><p>In 2001, when I was studying Electronics Engineering at university, learning something deeply required a specific set of rituals. I’d go to the library. I’d pull books off shelves. I’d wait for my professors’ office hours. Sometimes, to access a single piece of knowledge, I had to track down the one person who held it — maybe a professor, maybe a veteran engineer at a factory. Knowledge was hidden. Knowledge was scarce. And university was one of the few keys that unlocked its doors.</p><p>That system made perfect sense — for that era.</p><p>But it’s 2025 now. And those doors have been blown wide open.</p><h3>There’s No Such Thing as “Hidden Knowledge” Anymore</h3><p>In the early 2000s, knowledge lived in three places: universities, libraries, and industry. If you wanted to learn how to program a PLC, you either found someone at university who took that course or landed an internship at a factory. To understand how an algorithm worked, you needed access to textbooks that often existed only in university libraries. Knowledge was distributed asymmetrically, and universities were the primary vehicles of that asymmetry.</p><p>Today, you open YouTube. MIT’s lectures appear. Stanford’s courses. Harvard’s seminars — all free. The world’s best professors are looking into a camera and teaching you directly. Want to learn a programming language? Hundreds of hours of quality content await. Quantum physics? Neuroscience? Machine learning? It’s all there. In dozens of languages.</p><p>But here’s the thing: that’s not even the real revolution.</p><h3>AI Changed Everything</h3><p>YouTube democratized knowledge. But artificial intelligence has fundamentally transformed our relationship with it. You’re no longer just watching and learning — you’re <em>conversing</em> with AI. You stop when you don’t understand and ask questions. It generates examples tailored to you. It meets you at your level. It adapts to your pace.</p><p>Think about it: the 15 minutes of individual attention you could get from a professor during office hours in the 2000s? AI now gives you that 24/7. With infinite patience. With access to virtually the entire body of human knowledge.</p><p>I live this every single day. While building my companies, learning new technologies, designing complex systems — AI has become my most valuable teacher, assistant, and sparring partner. No university could ever offer me an educational experience this fast, this personalized, and this current.</p><h3>Agents and Robots: The Real Tsunami Is Just Beginning</h3><p>The transformation driven by AI is still in its infancy. Right now, autonomous AI agents are emerging — digital entities that can make decisions independently, execute tasks, and communicate with each other. These agents will fundamentally reshape countless white-collar jobs in the near future.</p><p>On the robotics front, a parallel revolution is unfolding. Robots that can operate in the physical world, learn, and adapt are no longer lab projects — they’re products stepping into reality.</p><p>When these two developments converge, the picture becomes clear: the world is changing far faster than any university can update its curriculum. The technology you learn in your first year might be obsolete by graduation. Universities can’t keep up. And structurally, they never will.</p><h3>So What Was University Actually For?</h3><p>Let’s be honest. Let’s really think about what university gave us:</p><p><strong>Knowledge?</strong> Now available everywhere, for free, and in more current form.</p><p><strong>A diploma?</strong> An increasing number of companies look at portfolios and real skills, not degrees. This has been true in tech for a while now.</p><p><strong>Networking?</strong> Social media, online communities, and open-source projects offer networks far wider and more diverse than university corridors ever could.</p><p><strong>Discipline?</strong> This might be the most valid argument. But the opportunity cost of spending four years learning discipline is growing steeper by the day.</p><p><strong>Research?</strong> In fundamental sciences and medicine, universities still play a critical role — I’ll grant that. But that doesn’t mean everyone needs to attend one.</p><h3>When I Hire, I Don’t Look at Diplomas — I Look for Passion</h3><p>I’m not saying this theoretically — I live it. As the founder of two companies, I hire people, and I almost never look at their diplomas.</p><p>The first thing I check is their GitHub profile. How much have they contributed? Which projects have they been involved in? Have they built something on their own? What’s the quality of their code? These things tell me what no diploma ever could. A GPA says almost nothing about what a person can actually do in the real world.</p><p>The second thing I look for might be even more important: do they have a hobby? Do they do something for themselves? Do they play an instrument, contribute to an open-source project, write a blog, create something — anything? My philosophy is simple: <strong>someone who doesn’t do anything for themselves won’t do anything for me either.</strong> People with hobbies are curious, disciplined, and capable of generating their own motivation. These are things no university can teach.</p><p>And here’s another truth: picking a specific discipline and becoming an expert in it takes far less time than a four-year university degree. And the direction is crystal clear. At university, you spend four years taking dozens of irrelevant courses. When you chart your own path, you can develop genuine expertise in a field within six months to a year. You spend the rest of your time reinforcing it with real-world projects.</p><h3>The Real Issue: Opportunity Cost</h3><p>Imagine an 18-year-old. Curious, energetic, hungry to learn. Two paths lie ahead:</p><p><strong>Path One:</strong> Spend four years at university. Struggle through mostly outdated curricula, irrelevant courses, and bureaucratic processes. Walk away with a diploma.</p><p><strong>Path Two:</strong> Design your own learning journey. Use AI as your mentor. Build real projects. Contribute to open-source communities. Do internships or launch your own venture. Build a portfolio. And do all of this not for four years, but continuously — as a lifelong learner.</p><p>Which path produces someone more ready for the world at 22?</p><h3>A Necessary Caveat</h3><p>Dismissing university entirely would be simplistic. If you’re going to practice medicine and perform surgeries, of course you need university. If you’re doing fundamental research in the sciences, you need a lab. Some professions legally require a degree.</p><p>But let’s be honest: the vast majority of university students don’t fall into these categories. Most go to university to “find a good job.” And that particular motivation? It no longer requires a university.</p><h3>The Education Model of the New Age</h3><p>I believe the future of education looks like this: lifelong, self-directed, AI-powered, and project-based learning. You no longer “receive” knowledge from an institution — you <em>build</em> it yourself, at your own pace, according to your own needs.</p><p>In this model, the teacher doesn’t disappear — they evolve. AI gives everyone a personal tutor. YouTube puts everyone’s library in their pocket. GitHub becomes everyone’s laboratory. And real-world projects become everyone’s final exam.</p><h3>Final Thought</h3><p>I don’t regret going to university in the early 2000s. It was the right decision for that era. Knowledge was genuinely hidden, and university was the best door to access it.</p><p>But in 2025, telling an 18-year-old “you must go to university” means guiding them with a 2000s mindset using 2025’s tools. The world has changed. Knowledge has been liberated. AI has thrown the doors wide open. And you no longer need a four-year ticket to walk through them.</p><p>All you need is curiosity, discipline, and the commitment to never stop learning.</p><p>The rest is already at your fingertips.</p><p><em>I’m writing this as someone who built a Video-on-Demand system in 2001 — years before YouTube was even imagined — while studying engineering, and who now develops AI infrastructure. The pace of change has always exceeded our expectations. Our understanding of education must keep up.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ecfc32881bbc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Actually Master Suno: A Deep Dive Into AI Music Creation]]></title>
            <link>https://medium.com/@cem.karaca/how-to-actually-master-suno-a-deep-dive-into-ai-music-creation-bad2e9603fa6?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/bad2e9603fa6</guid>
            <category><![CDATA[music]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-music-production]]></category>
            <category><![CDATA[musicais]]></category>
            <category><![CDATA[suno]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Tue, 09 Dec 2025 21:21:34 GMT</pubDate>
            <atom:updated>2025-12-09T21:21:34.119Z</atom:updated>
            <content:encoded><![CDATA[<p>After 1,492+ prompts, here’s what actually works</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DXFiNYR0SW2b4dDFAacA1g.jpeg" /><figcaption>How to master Suno</figcaption></figure><p>Let me save you months of frustration: your first Suno creation won’t be perfect. Neither will your second, third, or probably your tenth. But that’s not a bug — it’s how AI music generation works.</p><p>I’ve spent countless hours crafting prompts, testing metatags, and iterating through versions. This guide distills everything I’ve learned into actionable techniques that will dramatically improve your results.</p><p><strong>The single most important insight?</strong> Plan to generate 6–10 versions of every track. Systematic iteration beats hoping for magic on the first try.</p><h3>Understanding Suno’s Current Capabilities</h3><p>Suno’s V5 model (September 2025) represents a quantum leap from earlier versions. We’re talking studio-quality output, up to 8 minutes per generation, and vocals that actually sound human.</p><p>Here’s the evolution that matters:</p><ul><li><strong>V3</strong> brought us 2-minute generations</li><li><strong>V4</strong> improved vocal quality significantly</li><li><strong>V4.5</strong> extended to 8 minutes in a single generation</li><li><strong>V5</strong> delivers studio-grade audio with dramatically reduced artifacts</li></ul><p>The platform now supports over 1,200 music styles. But here’s what most guides won’t tell you: <strong>“Pop” and “Beat” are gravitational centers</strong> in Suno’s model. Almost everything gets pulled toward them unless you actively resist.</p><p>This is why negative prompting exists — and why you need to use it.</p><h3>The Anatomy of an Effective Prompt</h3><p>Research consistently shows that <strong>4–7 descriptors</strong> hit the sweet spot. Fewer produces generic output. More than 8 creates confusion and contradictions.</p><h3>The Core Formula</h3><pre>[Genre] + [Mood] + [Instrumentation] + [Production Style] + [Vocal Style]</pre><p>Example:</p><blockquote><em>“Cinematic trap soul, dark orchestral textures, aggressive, hypnotic, ambient strings, 808 bass, gospel vocal sample, echo”</em></blockquote><h3>Why Custom Mode Is Non-Negotiable</h3><p>Simple Mode is fine for casual experimentation. For anything serious, Custom Mode gives you three separate control surfaces:</p><ol><li><strong>Lyrics field: </strong>your words plus metatags</li><li><strong>Style of Music field: </strong>genre, mood, instruments</li><li><strong>Title field: </strong>Name of the song</li><li><strong>Weirdness: </strong>It is like temperature in OpenAI</li><li><strong>Style Influence:</strong> Lets you choose how close you stay to your style input from <strong>Loose</strong> to <strong>Strong</strong>.</li></ol><p>This separation is critical. It lets you control <em>what</em> the music sounds like independently from <em>what</em> the lyrics say.</p><h3>Metatags: The Secret Language of Structure</h3><p>Metatags are instructions wrapped in [square brackets] that control everything from song structure to vocal delivery. They go in the Lyrics field, not the Style field.</p><h3>Structural Tags</h3><pre>[Intro]<br>[Verse 1]<br>[Pre-Chorus]<br>[Chorus]<br>[Bridge]<br>[Drop]<br>[Outro]<br>[Fade Out]</pre><h3>Vocal Control Tags</h3><pre>[Male Singer]<br>[Female Singer]<br>[Vocalist: Alto]<br>[Vocal Style: Whisper]<br>[Vocal Tone: Raspy]<br>[Vocal Effect: Reverb]<br>[Harmonized Chorus]</pre><h3>Mood and Energy Tags</h3><pre>[Mood: Melancholic]<br>[Energy: High]<br>[Tempo: Slow]<br>[Texture: Gritty]</pre><h3>Instrument Tags</h3><pre>[Instrument: Piano]<br>[Instrument: Electric Guitar (Distorted)]<br>[Instrument: Strings (Legato)]<br>[Jazz Saxophone Solo]<br>[Bass Drop]</pre><h3>The Power Move: Section-Specific Emotion</h3><p>This is where things get interesting. You can give each section its own personality:</p><pre>[Verse 1] moody + brooding, minimal piano, intimate vocals<br>[Pre-Chorus] tension rise, filtered drums, layered whispers<br>[Chorus] explosive release, anthem-level energy, gospel-style harmony<br>[Bridge] tonal shift, lo-fi guitar, distant reverb</pre><h3>Creating Purely Instrumental Music</h3><p>This trips up more users than anything else. Here’s what actually works:</p><h3>The Triple-Lock Method</h3><ol><li>Turn on the <strong>Instrumental toggle</strong> in Custom Mode</li><li>Add “instrumental” to your Style field</li><li>Put [Instrumental] in your Lyrics field</li></ol><p>Yes, all three. Trust me on this — Simple Mode is unreliable for instrumentals, and single methods often fail.</p><h3>Effective Instrumental Prompt</h3><p><strong>Style:</strong></p><pre>Ambient electronic, atmospheric pads, no vocals, cinematic textures</pre><p><strong>Lyrics:</strong></p><pre>[Instrumental]<br>[Intro]<br>[Build]<br>[Main Theme]<br>[Outro - Fade]</pre><h3>Genre Strategy: Fighting the Pop Gravity</h3><p>Remember that gravitational pull toward pop? Here’s how to fight it:</p><h3>Strategy 1: Explicit Negation</h3><pre>Jazz, no pop<br>90s hip hop, boom bap, no trap<br>70s funk, groovy, no pop hooks<br>80s hard rock, no pop, no synth</pre><h3>Strategy 2: Two-Genre Fusion (Stable in V5)</h3><pre>Pop + EDM ✓<br>Gospel + Trap ✓<br>Jazz + Hip Hop ✓</pre><p>Stick to two genres. Three or more becomes unstable and muddy.</p><h3>Strategy 3: Rare Pair Combinations</h3><p>Want something unique? Try unexpected combinations:</p><ul><li>Emo industrial</li><li>Orchestral phonk</li><li>Math rock gospel</li></ul><p>These “rare pairs” produce creative results because the model hasn’t seen them frequently in training data.</p><h3>Building Crescendos and Dramatic Structures</h3><p>The [Crescendo] tag gradually increases volume, adds instruments, and raises energy levels. But placement matters.</p><h3>The Rules</h3><ol><li>Always precede a crescendo with a quiet section</li><li>Place it just before your climax (chorus, drop, or finale)</li><li>Follow it with a peak moment</li><li>Don’t overuse it — each use diminishes impact</li></ol><h3>Epic Orchestral Build Example</h3><pre>[Intro] A lone piano plays soft melody<br>[Verse] Gentle violins join in<br>[Crescendo] The orchestra swells – more strings, brass rising, drums rolling louder<br>[Chorus] Triumphant eruption with full orchestra and choir, majestic and powerful</pre><h3>Related Build Tags</h3><ul><li>[Build Intensity] — gradual energy increase</li><li>[Emotional Swell] — emotional intensity buildup</li><li>[Climactic] — musical peak point</li><li>[Orchestral Build] — gradual orchestral entry</li></ul><p>Combine them: [Crescendo] + [Drop] = tension before powerful release.</p><h3>Creating Background Music for Spoken Word</h3><p>This requires a fundamentally different approach. The music must support, never compete.</p><h3>Core Principles</h3><ul><li><strong>Low instrumentation</strong>: minimal, sparse arrangements</li><li><strong>Soft dynamics</strong>: keep volume subtle (pp to p)</li><li><strong>Slow tempo</strong>: 60–80 BPM</li><li><strong>Continuous flow</strong>: no sudden changes or harsh elements</li><li><strong>Instrumental only</strong>: no vocals competing with speech</li></ul><h3>Effective Spoken Word Background Prompt</h3><p><strong>Style:</strong></p><pre>Low acoustic background music that enhances reflective narration without overwhelming spoken word, soft piano, ambient pads, gentle strings, BPM 65</pre><p><strong>Lyrics:</strong></p><pre>[Instrumental]<br>[Background Music]<br>[Intro] Minimal ambient pad (10-15 seconds)<br>[Main] Soft piano, sustained strings<br>[Outro] Gradual fade with sustained notes</pre><h3>Production Tips for Background Music</h3><ul><li>Use [Clear mix] tag</li><li>Specify “quiet,” “subtle,” “background level”</li><li>Avoid mid-range clutter — use lower piano notes</li><li>Request “slow-evolving,” “continuous,” “no sudden breaks”</li></ul><h3>Creating Melancholic and Emotional Tracks</h3><p>Melancholy has specific musical signatures:</p><h3>Musical Characteristics</h3><ul><li>Minor keys (A minor, D minor, B minor)</li><li>Slow tempo (60–80 BPM)</li><li>Soft to moderate dynamics with emotional swells</li><li>Piano, strings, acoustic guitar, cello</li></ul><h3>Effective Melancholic Prompts</h3><pre>Melancholic ballad with slow tempo, soft piano and emotional lyrics, reflecting on lost love, minor key</pre><pre>Acoustic folk melancholic, fingerstyle guitar, intimate vocals, themes of loss and longing</pre><pre>Contemporary classical piano, emotive compositions, dynamic range, nuanced performance, introspective</pre><h3>Building Emotional Depth</h3><ul><li><strong>Layering</strong>: Combine piano with strings, add subtle cello</li><li><strong>Vocal style</strong>: “Soft vocals,” “airy vocals,” “intimate delivery,” “vulnerable tone”</li><li><strong>Production</strong>: “Warm analog compression,” “subtle reverb,” “dark timbre”</li><li><strong>Structure</strong>: Start quiet, build to emotional peak, resolve softly</li></ul><h3>The Iteration Workflow That Actually Works</h3><h3>Suno’s Built-In Tools</h3><ul><li><strong>Extend</strong>: Add content to beginning or end</li><li><strong>Replace Section</strong>: Regenerate specific parts while keeping the rest</li><li><strong>Cover</strong>: Recreate in a completely different style</li><li><strong>Stem Separation</strong> (Pro/Premier): Export 12 separate tracks</li></ul><h3>The Systematic Approach</h3><ol><li>Generate 5–10 versions with your initial prompt</li><li>Use thumbnail quality as a rough indicator</li><li><strong>Replace</strong> weak sections rather than regenerating everything</li><li><strong>Extend</strong> only when you need more length</li><li>Use <strong>Cover</strong> for genre experiments</li><li><strong>Remaster</strong> for final polish</li></ol><h3>Version Naming System</h3><p>Keep yourself organized:</p><pre>TrackName_v1_initial<br>TrackName_v2_chorus_fix<br>TrackName_v3_extended<br>TrackName_v4_gospel_cover<br>TrackName_FINAL_master</pre><h3>The Weirdness and Style Influence Sliders</h3><p>Available in V4.5+ for Replace/Cover/Remix operations:</p><h3>Style Influence</h3><ul><li><strong>High (80–100%)</strong>: Strictly follows your tags</li><li><strong>Medium (50–70%)</strong>: Balanced AI interpretation</li><li><strong>Low (0–40%)</strong>: AI takes creative liberty</li></ul><h3>Weirdness</h3><ul><li><strong>High</strong>: Experimental, unconventional</li><li><strong>Low</strong>: Traditional, safe choices</li></ul><h3>Recommended Combinations</h3><ul><li><strong>High style + Low weirdness</strong> = Clean execution of your vision</li><li><strong>Low style + High weirdness</strong> = AI-led creative reinterpretation</li><li><strong>For section replacement</strong>: Keep both below 50–60% for cohesion</li></ul><h3>Troubleshooting Common Problems</h3><h3>Problem: Vocals appear in instrumental track</h3><p><strong>Solution</strong>: Use the triple-lock method — toggle ON, “instrumental” in Style, [Instrumental] in Lyrics.</p><h3>Problem: Wrong genre / too much pop influence</h3><p><strong>Solution</strong>: Use negative prompts (“no pop,” “no beat”), be more specific (“90s grunge rock, alternative rock” not just “grunge”), front-load genre tags.</p><h3>Problem: Structure not being followed</h3><p><strong>Solution</strong>: Use explicit [Verse], [Chorus], [Bridge] tags. V5 has better structure adherence. Use callbacks in Extend: &quot;continue same vibe.&quot;</p><h3>Problem: Wrong instruments appearing</h3><p><strong>Solution</strong>: Use [Instrument: X] tags in Lyrics field, limit to 2–3 instruments per section, be specific (&quot;808 bass&quot; not just &quot;bass&quot;), add modifiers (&quot;distorted electric guitar&quot;).</p><h3>Real-World Case Study: “Karanhane”</h3><p>Theory is one thing. Let me show you what these techniques look like in practice.</p><p>“Karanhane” (Turkish for “asylum” or “dark house”) is an atmospheric piece I created as background music for poetry about institutional isolation and psychological depth. You can <a href="https://youtu.be/lhfpyLz1clw">listen to it on YouTube</a> or <a href="https://suno.com/s/NtUqinGqrhEcoRgP">view the generation on Suno</a>.</p><h3>The Prompt That Created It</h3><p><strong>Style Field:</strong></p><pre>Dark ambient electronic fused with acoustic elements, institutional coldness, prepared piano with reversed reverb, distant cello drones, glitch textures, claustrophobic atmosphere for asylum poetry, slow disturbing 58 BPM, unsettling calm, no vocals, extended haunting soundscape, detuned strings, psychological space</pre><p><strong>Lyrics Field:</strong></p><pre>[Instrumental]<br>[Tempo: 58 BPM]<br>[Key: Atonal/Chromatic]<br>[Mood: Institutional Cold, Psychological Disturbance, Unsettling Calm]</pre><pre>[Intro - Empty Corridors]<br>[Instrument: Ambient Pad (Dark, Reverb-Heavy)]<br>[Sound Effect: Distant Echo]<br>[Mood: Claustrophobic]</pre><pre>[Main Theme]<br>[Instrument: Prepared Piano (Detuned)]<br>[Instrument: Cello (Drone, Bowed Low)]<br>[Instrument: Electronic Glitch (Subtle)]<br>[Energy: Very Low]<br>[Background Level]</pre><pre>[Development - Tension]<br>[Instrument: Reversed Piano]<br>[Texture: Unsettling, Fragmentary]<br>[Mood: Internal Turmoil]</pre><pre>[Breakdown]<br>[Instrument: Cello (Scratchy Bow)]<br>[Sound Effect: Static Whispers]<br>[Mood: Psychological Depth]</pre><pre>[Outro - Fade to Void]<br>[Instrument: Ambient Drone]<br>[Silence]</pre><h3>Why This Works: A Breakdown</h3><p><strong>Specific BPM choice (58)</strong>: Not just “slow” — I specified exactly 58 BPM. This crawling tempo creates the unsettling, institutional pace that matches asylum imagery.</p><p><strong>Layered mood descriptors</strong>: “Institutional coldness,” “unsettling calm,” “claustrophobic atmosphere” — these aren’t redundant. Each shapes a different dimension of the emotional space.</p><p><strong>Unconventional instrument modifiers</strong>: “Prepared piano (Detuned),” “Cello (Scratchy Bow),” “Reversed Piano” — these push beyond standard sounds into genuinely unsettling territory.</p><p><strong>Programmatic section names</strong>: “[Intro — Empty Corridors],” “[Outro — Fade to Void]” — naming sections with evocative descriptions helps the AI understand the narrative arc.</p><p><strong>Energy and level control</strong>: “[Energy: Very Low]” and “[Background Level]” ensure the track supports spoken word without competing.</p><p><strong>Deliberate use of [Silence]</strong>: Ending with silence creates space for the poetry to breathe after the music fades.</p><h3>The Key Insight</h3><p>Notice how everything serves a single vision: the psychological space of an asylum. The detuned instruments, the institutional coldness, the claustrophobic atmosphere — every element reinforces the same emotional territory.</p><p>This coherence is what separates forgettable AI generations from tracks that actually move people.</p><h3>Ready-to-Use Templates</h3><h3>Uplifting Pop-Rock</h3><p><strong>Style:</strong></p><pre>Uplifting pop-rock with soaring electric guitar solos and energetic drums. Features powerful female vocal with harmonized backing vocals. Bright, anthemic</pre><p><strong>Lyrics:</strong></p><pre>[Intro]<br>[Instrument: Electric Guitar]<br>[Verse 1]<br>[Energy: Medium]<br>Your lyrics here<br>[Pre-Chorus]<br>[Build]<br>[Chorus]<br>[Energy: High]<br>[Harmonized Vocals]<br>[Bridge]<br>[Instrument: Acoustic Guitar]<br>[Energy: Low]<br>[Final Chorus]<br>[Energy: High]<br>[Layered Vocals]<br>[Outro]<br>[Fade Out]</pre><h3>Cinematic Orchestral (Instrumental)</h3><p><strong>Style:</strong></p><pre>Cinematic orchestral, epic, dramatic, no vocals, instrumental. Dark strings, powerful brass, thundering percussion</pre><p><strong>Lyrics:</strong></p><pre>[Instrumental]<br>[Intro]<br>[Instrument: Strings (Legato), Low]<br>[Mood: Mysterious]<br>[Build]<br>[Instrument: Brass, Percussion]<br>[Main Theme]<br>[Instrument: Full Orchestra]<br>[Energy: High]<br>[Mood: Triumphant]<br>[Break]<br>[Instrument: Piano Solo]<br>[Mood: Reflective]<br>[Finale]<br>[Instrument: Full Orchestra]<br>[Energy: Maximum]<br>[Mood: Epic]<br>[Outro]<br>[Fade]</pre><h3>Lo-Fi Hip-Hop Beat</h3><p><strong>Style:</strong></p><pre>Lo-fi hip hop beat, loop-friendly, chill, warm Rhodes piano, vinyl hiss, soft drums, jazzy, instrumental</pre><p><strong>Lyrics:</strong></p><pre>[Instrumental]<br>[Intro]<br>[Instrument: Warm Rhodes]<br>[Texture: Vinyl Hiss]<br>[Verse Pattern]<br>[Instrument: Rhodes, Soft Drums, Bass]<br>[Mood: Chill But Focused]<br>[Chorus Pattern]<br>[Instrument: Rhodes (Melody), Light Percussion]<br>[Outro]<br>[Structure: Seamless Loop]</pre><h3>The Do’s and Don’ts Cheat Sheet</h3><h3>Always Do</h3><p>✓ Be specific — include genre, tempo, instruments, mood<br> ✓ Use 4–7 descriptors (the proven sweet spot)<br> ✓ Specify “instrumental” when you don’t want vocals<br> ✓ Use metatags for structure: [Intro], [Verse], [Outro]<br> ✓ Generate multiple versions (6+ minimum)<br> ✓ Use negative prompts to avoid unwanted blending<br> ✓ Use Replace Section instead of full regeneration<br> ✓ Limit instruments to 2–3 per section</p><h3>Never Do</h3><p>✗ Stay vague (“sad song” = too generic)<br> ✗ Overload with 8+ conflicting descriptors<br> ✗ Use contradictory cues (“slow” + “high energy”)<br> ✗ Name specific artists (copyright issues)<br> ✗ Mix clashing genres (death metal + lullaby)<br> ✗ Expect perfection on the first try<br> ✗ Use Simple Mode for instrumentals<br> ✗ Stack 3–4 genres together</p><h3>Final Thoughts</h3><p>Mastering Suno isn’t about finding the perfect prompt, it’s about understanding how the AI interprets your instructions and systematically refining your approach.</p><p>The platform has evolved dramatically. V5 produces genuinely impressive results. But the gap between amateur and professional output isn’t the model, it’s the method.</p><p>Start with clear intent. Use specific descriptors. Leverage metatags for structure. Iterate systematically. Fight the pop gravity with negative prompts.</p><p>And remember: every expert started by generating a dozen mediocre tracks. The difference is they kept refining until they didn’t.</p><p>Now go make something extraordinary.</p><p><em>Have questions about specific genres or use cases? Drop them in the comments — I’ve probably tested it.</em></p><p>Follow me on <a href="https://suno.com/@cemkaraca">Suno</a>, <a href="https://www.youtube.com/CemKaraca">Youtube</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bad2e9603fa6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Last Whip Is Rotting]]></title>
            <link>https://medium.com/@cem.karaca/the-last-whip-is-rotting-9b31b161b879?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b31b161b879</guid>
            <category><![CDATA[postscarcity]]></category>
            <category><![CDATA[death-of-capitalism]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[universal-abundance]]></category>
            <category><![CDATA[automationrevolution]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Sun, 23 Nov 2025 03:32:45 GMT</pubDate>
            <atom:updated>2025-11-23T03:32:45.801Z</atom:updated>
            <content:encoded><![CDATA[<p><em>And the only unknown left is: when?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z_KfGpsQcmGhMEIxSmoMRg.png" /><figcaption>Three whips of slavery</figcaption></figure><p>There have always been three stages of the whip.</p><ol><li>The leather whip <br>Literal slavery. <br>Blood on the ground. <br>Chains you could see.</li><li>The money whip <br>Wage slavery. <br>“Work or starve.” <br>Chains you could feel but not see.</li><li>The exhaustion whip <br>Distraction slavery. <br>Netflix, Instagram, TikTok, 7/24 news cycle. <br>The modern human stayed in the cage not because the door was locked, but because he was too tired to notice it was open.</li></ol><p>This third whip was the most sophisticated one ever designed. It didn’t need overseers. It ran on dopamine and FOMO. People volunteered to tighten it themselves.</p><blockquote>But the fourth stage is now ending.</blockquote><p>Robots + AI are becoming cheaper than the cheapest human labor.</p><p>One-time cost + maintenance will replace salary, insurance, severance, maternity leave, 8-hour day.</p><p>Everything will become so abundant that scarcity itself will become optional.</p><p>When there is nothing left to withhold, there is nothing left to threaten people with.</p><p>The third whip will rot in the manure it helped create. The money whip will rot with it. Even the memory of the leather whip will become a bad dream.</p><p>This is the manure phase.</p><p>Since Nixon slammed the gold window shut in 1971, the entire global economy has been running on one simple trick: print money, borrow cheap from Japan’s near-zero rates, pump it into assets, extract labor from the rest of the planet, and call it “growth.”</p><p>The Japanese worker in Vietnam, Turkey, Mexico, or the American Midwest has been paid in increasingly worthless paper while the borrowed money inflated yachts, private jets, and Davos panels.</p><p>The yen carry trade was just the latest version of an old game: borrow at 0%, invest in anything that yields more, let the currency differential and labor arbitrage do the rest. <br>The worker never saw the profit. <br>He only saw the debt.</p><p>The result? <br>The most extreme wealth inequality since the days of literal slavery.</p><p>In 2016, just 62 human beings owned as much as the poorer half of humanity — 3.6 billion people. <br>That is not a market outcome. <br>That is anti-Robin Hood on a planetary scale. <br>Money didn’t trickle down. <br>It was vacuumed up.</p><p>And when someone mildly objected?</p><p>Madeleine Albright was asked on 60 Minutes whether the deaths of half a million Iraqi children (due to sanctions) were worth it. <br>Her answer, delivered with bureaucratic calm: <br>“We think the price is worth it.”</p><p>Worth it for whom?</p><p>For the people who needed to clear old military inventory so the new inventory could be sold. For the 62 people who must be kept ecstatic at all costs. <br>For a system that runs on controlled explosions</p><blockquote>some in stock markets, <br>some in actual cities.</blockquote><p>That system is now eating itself.</p><p>Twenty years ago the largest companies on Earth were oil giants, banks, and conglomerates that made their money extracting, financing extraction, or insuring extraction. Exxon, Citigroup, General Electric, Pfizer, Chevron. <br>Today? Apple, Microsoft, Nvidia, Amazon, Alphabet. <br>Not one of the 2005 top 10 is still in the 2025 top 10.</p><p>The old empire is dying of natural causes<br>and the nerds did it without firing a shot.</p><p>They built tools that made gatekeepers irrelevant. <br>They built payment systems (remember PayPal) that bypassed 1980-model banks. <br>They built software that eats middlemen for breakfast. <br>They built AI that turns a biologist into a one-person drug discovery company.</p><p><a href="https://medium.com/@cem.karaca/the-revenge-of-the-domain-expert-8a65e55c2406">The revenge of the domain expert</a> was only the warm-up. <br>The real revolution is the death of the whip itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*geHPh2zgChhRllA9M6DDVg.png" /><figcaption>Flowers grow with manure</figcaption></figure><p>Flowers do not ask permission from manure. <br>They simply grow.</p><p>We are in the manure right now. <br>The smell is at its peak. <br>But the roots are already deep.</p><p>And here is the part that should pour cold water on the hottest fears:</p><blockquote>We cannot solve our problems with the same thinking we used when we created them. — Albert Einstein</blockquote><p>Do not be afraid of “unemployment.” <br>Unemployment is a concept invented by the money whip. <br>When we step outside the capitalist logic, the question itself becomes meaningless.</p><p>Your children will not grow up asking <br>“Where do you work?” <br>They will ask:</p><blockquote>“How many bots do you have running for you?” <br>“How many robots are working while you sleep?”</blockquote><p>On that day, the Gini coefficient will not just improve. <br>It will become a historical curiosity.</p><p>The last whip is rotting.</p><p>The only unknown left in the equation is:</p><p><strong>When?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tJWONjLQgkpa86K1gJ9vZQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b31b161b879" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why the Best Answers Start with Questions You Haven’t Learned to Ask Yet]]></title>
            <link>https://medium.com/@cem.karaca/why-the-best-answers-start-with-questions-you-havent-learned-to-ask-yet-147e03f530b5?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/147e03f530b5</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[questions]]></category>
            <category><![CDATA[douglas-adams]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[aritificial-intelligence]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Sat, 22 Nov 2025 14:07:18 GMT</pubDate>
            <atom:updated>2025-11-22T14:07:18.783Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jGNIrEsks_rvgam1fAmKxQ.png" /><figcaption>what is the best question?</figcaption></figure><p>Deep Thought, the second-greatest computer in the universe, spent 7.5 million years calculating the Answer to the Ultimate Question of Life, the Universe, and Everything. When it finally spoke, the answer was… <strong>42</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VXpCHQnTVdn9y1sNumt3_w.png" /><figcaption>Asnwer to the ultimate question</figcaption></figure><p>The problem? Nobody remembered the question.</p><p>Douglas Adams didn’t just write comedy; he dropped a brutal truth bomb on all of humanity: <strong>we are astonishingly bad at asking the right question.</strong> And for 7.5 million years, we waited because we outsourced the question instead of refining it.</p><p>Fast-forward to 2025. We no longer need to build a planet-sized computer and wait until the Sun goes nova. We have Grok, Claude, Gemini, Llama, and friends. They are faster than Deep Thought, cheaper than a Vogon poetry reading, and (crucially) available 24/7.</p><p>But here’s the twist that nobody wants to hear:</p><p><strong>The most powerful AI in the world is still completely useless if you feed it the wrong question.</strong></p><p>You can have 400 billion parameters, 100 trillion tokens of training data, and a context window the size of Belgium. If your prompt is garbage, your output will be gourmet garbage, beautifully formatted and confidently wrong.</p><h4>The Most Powerful Answer to the Wrong Question</h4><p>Imagine asking an AI superintelligence: “What’s the best marketing strategy?”</p><p>It’s 2025.”</p><p>It will happily give you a 12-point plan involving TikTok dances, NFT giveaways, and hiring a celebrity dog. It will be up-to-date, persuasive, and 100 % optimized… for 2022.</p><p>Because you asked for “the best marketing strategy” instead of “the highest-ROI customer acquisition strategy for a B2B SaaS with $30k monthly burn in a post-iOS14 privacy world.”</p><p>Same model. Same temperature. Totally different usefulness.</p><p>That, ladies and gentlemen, is the most expensive mistake you can make in the age of AI.</p><h4>No Experiment Is Reproducible (and That’s the Point)</h4><p>In traditional software, we worshipped reproducibility. Same code + same input = same output. Forever. AI flipped the table.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9ibvyYp4WARxXR0hzfJRqg.png" /><figcaption>hoodies are worshipping a glowing laptop</figcaption></figure><p>Every run is an experiment. Every temperature setting is a hypothesis. Every system prompt is a petri dish.</p><p>And that’s not a feature, not a bug.</p><p>The moment you accept that no output is perfectly reproducible, you become free to iterate like a mad scientist. You stop looking for “the correct prompt” and start running 50 variations until something clicks.</p><p>Edison didn’t find the right filament; he found 2,000 wrong ones first. You don’t find the right prompt; you eliminate the 49 stupid ones.</p><h4>The Indispensable Question Template (steal this)</h4><p>Here’s the prompt framework that turns “meh” into “holy shit” 90 % of the time:</p><pre>You are the world&#39;s best [ROLE] with 30 years of ruthless experience shipping [SPECIFIC OUTCOME] in [HARD CONTEXT].</pre><pre>Goal: [One-sentence, measurable desired outcome]</pre><pre>Constraints:<br>• [Hard limit 1]<br>• [Hard limit 2]<br>• [Hard limit 3]</pre><pre>Available data / context:<br>[Insert everything relevant — screenshots, metrics, previous attempts, customer quotes]</pre><pre>Success looks like:<br>[Paint the picture of the perfect deliverable in vivid detail]</pre><pre>Failure looks like:<br>[Explicitly call out the generic, fluffy, or recycled crap you do NOT want]</pre><pre>Style:<br>• Be brutally direct<br>• Favor counter-intuitive but high-leverage ideas<br>• Show trade-offs, not just upside<br>• Use tables for comparisons<br>• End with 3 next experiments I should run</pre><pre>First, think step-by-step (visible).<br>Then, deliver the most probable answer.</pre><p>Copy-paste. Fill brackets. Profit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*prK1ePQEzXMzjbDssjNTOg.png" /><figcaption>average person vs the person asking the right questions</figcaption></figure><h4>Explain the Shit (a mini case study)</h4><p>Last month, I asked ChatGPT: “Give me content ideas for my newsletter.”</p><p>→ 15 bullet points of utter beige.</p><p>Then I used the template above with ROLE = “cynical growth marketer who grew MrBeast from 0 to 100M” and added real subscriber data + churn metrics.</p><p>Result: An issue titled “Why Your Open Rate Sucks (and How to Fix It Before You Run Out of Money)” that got a 68 % open rate and $38k in new sales.</p><p>Same model. One took 3 seconds and was forgotten in 4. The other paid my rent for half a year.</p><h4>Final Thought: We Are Living in the Age of the Question</h4><p>Deep Thought needed 7.5 million years because no one told it what to ask. Today, you can build a second brain that asks better questions than 99.9 % of humans, <strong>but only if you train yourself to ask it the right ones first.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F3HVO7_n3fXicCQYUu-H_w.png" /><figcaption>Deep Thought</figcaption></figure><p>The ultimate leverage in 2025 isn’t access to AI. It’s the rare, weird, slightly obsessive skill of noticing which question you’re actually trying to answer… and then having the guts to ask it properly.</p><p>So go ahead. Ask a stupid question, get a legendary answer.</p><p>Or ask a legendary question, and change everything.</p><p>The universe is waiting. And this time, it only takes seconds.</p><p>(Now go write prompts that would make Douglas Adams proud.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9uZ-jyhu4QxQ_o--A60VvA.png" /><figcaption>Ultimate shining question</figcaption></figure><blockquote>“And remember: In the end, the universe doesn’t reward the smartest. <br>It rewards the one who asks the right question. <br>Whoever does that… wins.”</blockquote><p><strong>Bonus: Nano-Banana-Pro Image Prompt</strong></p><pre>Ultra-cinematic, hyper-realistic, apocalyptic masterpiece. <br>A colossal golden banana the size of a galaxy is splitting open in the void of space. <br>From the crack erupts a blinding white plasma made entirely of perfect prompts — glowing lines of text in alien fonts flying out like angelic scripture. <br>Surrounding the explosion: thousands of shattered AI avatars (broken Grok faces, melting Gemini logos, burning Claude eyes) screaming in silent ecstasy as they’re reborn into pure light. <br>In the foreground, a single tiny human silhouette (you) standing on a floating piece of debris, arms outstretched, bathed in golden light, with the words “DON&#39;T PANIC” carved into the banana peel in burning letters. <br>Background: collapsing matrix code raining sideways, dying stars, fragments of “what is the best” prompts disintegrating into ash. <br>Color grade: Teal-orange cyberpunk apocalypse with molten gold accents. <br>Shot on IMAX 70mm, anamorphic lenses, insane detail, 16K, god-tier composition, the kind of image that makes people quit their jobs and start prompting religions</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pi5M9pgWocSPQa9u9CX7RA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=147e03f530b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Revenge of the Domain Expert]]></title>
            <link>https://medium.com/@cem.karaca/the-revenge-of-the-domain-expert-8a65e55c2406?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a65e55c2406</guid>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Wed, 19 Nov 2025 02:19:24 GMT</pubDate>
            <atom:updated>2025-11-19T02:19:24.858Z</atom:updated>
            <content:encoded><![CDATA[<p>AI Didn’t Kill Coding. It Just Handed the Keys to Biologists.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pi0XEkurDOeEtv5pqloPFQ.png" /></figure><p><strong>AI didn’t make software developers obsolete; it made domain experts with software skills the most valuable people on the planet.</strong></p><p>If you’re a biologist, chemist, materials scientist, or electronics engineer reading this, I have news that will change how you think about your career.</p><p>Your domain expertise just became exponentially more valuable. Not less. More.</p><p>Here’s why.</p><h3>The Great Inversion</h3><p>For the past 30 years, software developers have been the gatekeepers.</p><p>You had a brilliant idea for optimizing drug discovery? You needed to convince a software team to build it. You understood how to revolutionize circuit design? Get in line behind the login page refactor and 42 other tickets.</p><p><strong>That bottleneck just exploded.</strong></p><p>AI can now write competent code. Not perfect code, but competent code. The kind that gets 80% of software projects done.</p><p>What AI cannot do is understand why a particular protein folding pattern matters, or why a specific alloy composition will fail under thermal stress, or why your clinical trial design is fundamentally flawed.</p><blockquote><strong>The rarest skill is no longer coding. It’s knowing what to build.</strong></blockquote><h3>Why This Changes Everything: The Matchup</h3><p>Let’s look at the tale of the tape in 2025.</p><p><strong>Person A: The Pure Technologist</strong></p><p>Role: Senior Python Developer</p><p>Superpower: Can architect scalable microservices, optimize database queries, and write pristine, test-driven code.</p><p>The Limitation: They don’t know a peptide from a polymer.</p><p><strong>Person B: The Domain Expert</strong></p><p>Role: PhD Materials Scientist</p><p>Superpower: 10 years of intuition on battery electrode failure modes.</p><p>The Limitation: Their code is messy. They just learned Python six months ago using Claude and Cursor.</p><p><strong>The Outcome:</strong> Person A builds a beautiful, scalable system that solves the wrong problem. Person B writes “spaghetti code” that successfully identifies a novel alloy composition, revolutionizing battery life.</p><p><strong>Winner:</strong> Person B. Because the code is temporary, but the discovery is permanent.</p><h3>The Disciplines Ripe for Disruption</h3><p>Look around your field. I guarantee there are processes that:</p><ul><li>Involve tedious manual analysis</li><li>Could be automated, but never had the engineering resources</li><li>Require expertise that seemed impossible to encode</li><li>Have been “on the roadmap” for years</li></ul><p><strong>Those problems are now solvable.</strong> Not by software developers learning your field (that takes a decade), but by you learning enough software to articulate what you need.</p><p>Here’s what’s already happening:</p><p><strong>Biology</strong>: Researchers who understand protein interactions are building their own molecular simulation tools. Not waiting for IT departments.</p><p><strong>Chemistry</strong>: Synthetic chemists are creating retrosynthesis planning tools that understand reaction mechanisms at a level no pure ML engineer could replicate.</p><p><strong>Electronics</strong>: This is where I live. I’m combining my electronics engineering background with Plugged.in to build interfaces that let AI agents actually sense, touch, and see the physical world. Not just process data about reality but interact with it directly. You can’t build this from a pure software perspective. You need to understand signal integrity, sensor fusion, real-time constraints, and the messy analog world where theory meets physics.</p><p><strong>Law</strong>: Attorneys are building contract analysis tools that understand legal precedent and jurisdictional nuances, not just keyword matching.</p><p><strong>The pattern?</strong> Domain experts who can translate their intuition into software are creating solutions that pure software teams never could.</p><h3>What AI Experts Need to Understand</h3><p>If you’re reading this as a software developer or AI engineer, here’s your wakeup call:</p><blockquote><strong>You need domain partners more than they need you.</strong></blockquote><p>I’ve spent decades in software development. I’ve built streaming systems before YouTube existed, worked across multiple tech stacks, and now I’m building AI agent infrastructure. But here’s the twist: my foundation is in electronics engineering. And that’s exactly why Plugged.It isn’t just another AI platform; instead, it’s building the bridge between AI agents and the physical world. Software that can sense, touch, and see beyond the digital realm.</p><p><strong>I am tired of building pipes for water I don’t understand. I need someone who knows the water.</strong></p><p>Why? Because I can build the infrastructure, but I cannot tell you:</p><ul><li>Which molecular structures are worth simulating</li><li>Why a flight control algorithm that converges perfectly in simulation will cause a drone to crash due to harmonic resonance in the chassis.</li><li>Which clinical endpoints matter in a trial</li><li>What material properties predict real-world failure modes</li></ul><p><strong>The value is in the intersection.</strong> Pure software? Commoditized. Pure domain expertise? Trapped in legacy workflows. Both together? That’s where billion-dollar companies are born.</p><h3>The New Mental Model</h3><p>Stop thinking: “I need to hire developers.” Start thinking: “I need to become dangerous with code.”</p><p>You don’t need to be a senior engineer. You need to be able to:</p><ul><li>Articulate your domain problems clearly enough for AI to help</li><li>Understand when a technical approach will or won’t work in your field</li><li>Prototype rapidly and iterate based on domain feedback</li><li>Bridge the language gap between your field and software</li></ul><p><strong>“The barrier to entry just dropped from 10,000 hours to maybe 100.”</strong></p><h3>A Call to Both Sides</h3><p><strong>To domain experts:</strong> Your expertise is not obsolete, it’s the most valuable it’s ever been. But you need to meet software halfway. Learn the basics. Start small. Build a tool that solves your daily annoyance. You’ll be shocked how quickly you can create something useful.</p><p><strong>To AI/software experts:</strong> Stop trying to “disrupt” fields you don’t understand. Start looking for brilliant domain experts who are curious about software. Teach them. Partner with them. Build the infrastructure that lets them move fast. You’re not the protagonist of this story; you’re the enabler.</p><h3>The Opportunity</h3><p>We’re at a unique moment in history. The window is open now.</p><p>In 3 years, “domain experts who code” won’t be a novelty; they will be the baseline requirement. The arbitrage opportunity is today. To be the cardiologist who builds the AI diagnostic tool, or the civil engineer who creates the structural analysis platform, or the climate scientist who automates the simulation workflow.</p><p><strong>The future belongs to the translators.</strong> The people who can speak both languages.</p><p>The question is: which side of this wave will you be on?</p><p><em>I’m building Plugged.in because I believe the next wave of innovation comes from empowering domain experts to build their own tools. But I’m also making a specific bet: that AI agents need to break out of purely digital environments. My electronics background taught me that the physical world is messy, analog, and beautiful, and that’s where the real problems are. So I’m building infrastructure that lets agents sense, touch, and see the world, not just read about it.</em></p><p><em>This is my version of “electronics + software.” What’s yours?</em></p><p><strong>What’s your domain? What problems could you solve if the software barrier disappeared? Drop a comment below.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a65e55c2406" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PAP: The Kill-Switch Protocol Turning AI Agents from Loose Cannons into Starfleet]]></title>
            <link>https://medium.com/@cem.karaca/pap-the-kill-switch-protocol-turning-ai-agents-from-loose-cannons-into-starfleet-d9aa531c4ca4?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/d9aa531c4ca4</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Sat, 01 Nov 2025 05:11:18 GMT</pubDate>
            <atom:updated>2025-11-01T05:11:18.274Z</atom:updated>
            <content:encoded><![CDATA[<p>Orchestrating autonomous AI without the zombie apocalypse — introducing Plugged.in’s backbone for distributed cognition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LkzXuZOT_3nSvGAquznKEA.png" /><figcaption>Autonomous Agents</figcaption></figure><blockquote>Summary: PAP defines a secure, versioned, and auditable protocol enabling seamless communication between the Plugged.in control center (the Station) and autonomous agents (Satellites). Built on DNS-based identity, mTLS, signed messages, standardized error codes, heartbeats, and full open-source transparency, PAP provides an infrastructure-grade foundation for distributed cognition.</blockquote><p>Last week, I was knee-deep in an MCP chain that spiraled into a 10-minute hallucination loop. An agent went rogue, churning out nonsense because there was no easy way to hit the brakes or even know it was off the rails.</p><p>Sound familiar? That’s the dirty secret of today’s AI agent ecosystems: they’re brilliant at thinking, but lousy at surviving. Agents spawn, collaborate, and adapt… until they don’t. A silent failure here, a resource hog there, and suddenly your “autonomous fleet” is a drifting graveyard of zombies.Enter PAP: the Plugged.in Agent Protocol. In short: PAP is the DevOps harness AI agents need to thrive without torching the server farm. It’s not just another chat standard like MCP or A2A; it’s the control tower that lets agents fly free while keeping the mothership’s hand firmly on the kill switch.Think of it like this: AI agents are evolving from isolated utilities into infrastructure-level components — persistent entities that perform, adapt, and collaborate. Yet, orchestration today remains brittle. Existing standards enable model or organization-level collaboration — not operational reliability. None embed realities like spawn, health, kill, ownership transfer, or lifecycle authority.PAP fills this gap. It’s a control and telemetry backbone designed for the Plugged.in ecosystem — where each agent acts like a spacecraft connected to its command station.</p><h4>The Metaphor: Space Station &amp; Shuttles</h4><p>Picture the International Space Station: a bustling hub of modules, experiments, and crew, all synced to mission control. Now swap humans for AI agents.</p><ul><li><strong>The Station (Plugged.in Core):</strong> Your command, policy, and orchestration hub. It issues orders, monitors vitals, and — crucially — holds ultimate kill authority.</li><li><strong>The Shuttles (Agents):</strong> Autonomous, diverse, and mission-focused; yet bound to the Station through one universal protocol.</li></ul><p>The principle? <strong>Autonomy without anarchy.</strong> Every agent operates freely but remains accountable. They self-heal, specialize, and execute locally — zero lag from constant check-ins. But if a shuttle goes dark? The Station can remote-pilot a graceful shutdown or, in extremis, hit the big red button.</p><p>This isn’t sci-fi; it’s the ops reality we need as agents scale from toys to tenants in our clouds.</p><p>Architecture Diagram: Station → PAP Proxy (mTLS + Signatures) → Agents (e.g., Focus, Memory). Arrows show heartbeat, invoke, and kill flows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K8CJsc-i6kGCU65Jq4wKvQ.png" /><figcaption>Figure 1 — Station ↔ Shuttles orchestration diagram</figcaption></figure><h4>Why PAP Matters: Managing Existence, Not Just Execution</h4><p>Most frameworks (shoutout to MCP, ACP, and A2A) nail AI logic — how models invoke tools or gossip between agents. PAP? It manages AI existence: how they live, breathe, migrate, and die. It merges DevOps-grade control with AI cognition: lifecycle management, telemetry, and verifiable trust.</p><p><strong>The Zombie Risk — and How PAP Slays It</strong></p><p>A “zombie” agent? That’s your nightmare: a silent failure sucking CPU like a black hole. PAP prevents it with:</p><ul><li><strong>Heartbeat events:</strong> Agents ping liveness only (metrics flow separately via Metrics.Report). No bloat, just “I’m alive.”</li><li><strong>Watchdog thresholds:</strong> Detect stalls, runaway resources, or anomalies.</li><li><strong>Force Kill Authority:</strong> Reserved for the Station. Graceful first (drain queues, confirm shutdown), nuclear if needed (-9 style).</li></ul><pre># Agent sends heartbeat—lightweight and frequent<br>client.Heartbeat(HeartbeatEvent(<br>    mode=HeartbeatMode.IDLE,<br>    uptime_seconds=3600<br>))<br><br># Separately, beam back metrics for deeper dives<br>client.Report(ResourceReport(<br>    cpu_percent=45.2,<br>    memory_mb=512<br>))</pre><blockquote>Pro Tip: Heartbeat latency? Under 100ms at p99. Scale to 1000+ agents per proxy without breaking a sweat.</blockquote><p>Benefits stack up: Local decisions for speed, sandboxed runtimes for isolation, modular specialization (Focus agents for tasks, Memory for persistence, Edge for real-time).</p><p><strong>Real-World Orbit: PAP in Action</strong></p><p>To make it concrete, here’s how PAP orbits everyday chaos:</p><ul><li><strong>Edge Agent Migration:</strong> Your analytics shuttle needs to jump clouds mid-mission? Ownership transfer preserves full state — cert handoff, queue drain, seamless resume. No data loss, under 30 seconds.</li><li><strong>Swarm Overload:</strong> 50 agents tackling a data flood? Invoke cascades with timeouts and 429 RATE_LIMITED codes keep it sane; zombies get culled before cascade.</li><li><strong>Audit Trail for the Paranoid:</strong> Every message logs to Plugged.in’s Memory Service. Replay a failed invoke? Trace_id and span_id make it child’s play.</li><li><strong>Hybrid Fleet:</strong> Pair a Focus agent with an A2A gossip chain — PAP handles the plumbing so you focus on the mission.</li></ul><p>These aren’t hypotheticals; they’re the guardrails turning agent dreams into deployable reality.</p><p><strong>Lifecycle: From Spawn to (Graceful) Grave</strong></p><p>PAP’s flow is a closed loop:</p><ol><li><strong>Provisioning:</strong> Invite token → authenticate → snag permanent cert &amp; DNS ID (`{agent}.{cluster}.a.plugged.in`).</li><li><strong>Operation:</strong> Commands, events, telemetry zip through the PAP Proxy (Kubernetes-native routing via Rancher + Traefik).</li><li><strong>Ownership Transfer:</strong> Shuttle hops clusters? State intact, no drama.</li><li><strong>Graceful Termination:</strong> Drain tasks, ack shutdown.</li><li><strong>Force Kill:</strong> Heartbeat expires? Station pulls the plug.</li></ol><p>Message types keep it simple:</p><pre>| Type      | Purpose                                   |<br>|-----------|-------------------------------------------|<br>| `invoke`  | Command from Core or peer agent           |<br>| `response`| Ack or result                             |<br>| `event`   | Telemetry (heartbeats, logs, alerts)      |<br>| `error`   | Structured oops (with details)            |</pre><p>Security’s locked down: mTLS for transport, DNSSEC + JWT for onboarding, Ed25519 signatures + CRC for integrity. Audit everything — because trust is earned in bytes.</p><p><strong>Error Codes: TCP/IP Wisdom for AI Woes</strong></p><p>No more vague “it broke” logs. PAP borrows HTTP/gRPC vibes but adds agent-specific zingers:</p><pre>| Code               | HTTP | Meaning              |<br>|--------------------|------|----------------------|<br>| OK                 | 200  | All good             |<br>| ACCEPTED           | 202  | Async job queued     |<br>| BAD_REQUEST        | 400  | Args are trash       |<br>| UNAUTHORIZED       | 401  | Who dis?             |<br>| FORBIDDEN          | 403  | Nice try, no         |<br>| NOT_FOUND          | 404  | Agent lost in space  |<br>| TIMEOUT            | 408  | Too slow             |<br>| CONFLICT           | 409  | Version clash        |<br>| RATE_LIMITED       | 429  | Chill, throttle hit  |<br>| AGENT_UNHEALTHY    | 480  | Zombie alert!        |<br>| AGENT_BUSY         | 481  | Overloaded shuttle   |<br>| DEPENDENCY_FAILED  | 482  | Downstream drama     |<br>| INTERNAL_ERROR     | 500  | My bad               |<br>| PROXY_ERROR        | 502  | Routing hiccup       |<br>| VERSION_UNSUPPORTED| 505  | Upgrade or GTFO      |</pre><blockquote>Pro Tip: 480 AGENT_UNHEALTHY? That’s your early-warning siren for rogue CPU hogs. Catch ’em before they eat your quota.</blockquote><p><strong>How PAP Stacks Up: The Competition Snapshot</strong></p><pre>Feature             MCP      ACP      A2A      PAP        Why It Matters<br>──────────────────  ───────  ───────  ───────  ─────────  ────────────────────────────<br>Central Control     ❌        Partial  ✅        ✅          Emergency shutdown authority<br>Kill Authority      ❌        ❌        ❌        ✅          Zombie prevention<br>DNS-Based Identity  ❌        ❌        ❌        ✅          Kubernetes-native routing<br>Zombie Detection    ❌        Partial  ❌        ✅          Resource leak prevention<br>Ownership Transfer  ❌        ❌        ❌        ✅          Multi-cloud portability<br>gRPC Native         ✅        ✅        ❌        ✅          Efficient binary protocol<br>Open Standard       ✅        ✅        ✅        Full Open  Transparent governance</pre><p>PAP doesn’t compete — it complements. Use it to bolt reliability onto your MCP flows.</p><p><strong>Observability &amp; Performance: Built for the Long Haul</strong></p><ul><li><strong>Tracing:</strong> Every message packs trace_id and span_id — export to OpenTelemetry for that sweet distributed view.</li><li><strong>Logging:</strong> Agents MUST journal state transitions. Replayability? Check.</li><li><strong>Benchmarks:</strong> 1000 agents/proxy, state transfers in ❤0s. Heartbeats? Sub-100ms.</li></ul><p><strong>The Roadmap: From Draft to Stars</strong></p><ol><li><strong>PAP v1.0 Schema:</strong> Protobuf defs, codes, handshake — dropping now.</li><li><strong>SDKs:</strong> Python/TS for quick wins, Rust/Go for the hardcore.</li><li><strong>Proxy:</strong> mTLS router with JWT + OTEL baked in.</li><li><strong>Registry &amp; Policy:</strong> IAM-style controls for routes and caps.</li><li><strong>Marketplace:</strong> Helm charts for one-click deploys.</li></ol><p>Diversity from day one: SDKs for the masses and purists.</p><p><strong>Appendix Sneak Peek: Protobuf Messages</strong></p><p>For the builders: Full schemas on GitHub. Here’s a taste:</p><pre>message PAPHeader {<br>  string version = 1;<br>  string id = 2;<br>  string from = 3;<br>  string to = 4;<br>  string type = 5; // invoke | response | event | error<br>  string signature = 6;<br>  string crc = 7;<br>  google.protobuf.Timestamp timestamp = 8;<br>}<br><br>message HeartbeatEvent {<br>  PAPHeader header = 1;<br>  string mode = 2; // ACTIVE | IDLE | STANDBY<br>  uint64 uptime_seconds = 3;<br>}<br><br>// ... (more in the repo)</pre><p><strong>Join the Mission</strong></p><p>PAP isn’t locked in a vault — it’s your forkable future. This doc doubles as the canonical RFC (PAP-RFC-001-rev2.1), production-ready and ecosystem-aligned. Head to <a href="https://github.com/VeriTeknik/PAP">GitHub Repo</a> for the full appendix Protos, sample SDK stubs, and a “<strong>Spin Up Your First Satellite</strong>” quickstart. Got a zombie horror story from MCP gone wild? Drop it below — let’s commiserate and collaborate.</p><p>If MCP’s your current orbit, check my article <a href="https://medium.com/@cem.karaca/mcp-servers-your-gateway-to-supercharged-ai-assistants-192128ca8174">here</a>. Dreaming inter-agent diplomacy? Layer on A2A.</p><p>What’s next for PAP? God knows — but with you aboard, it’ll be stellar.<br><br>Happy Halloween, may your agents stay undead… responsibly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d9aa531c4ca4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Nerds Are Coming for Everyone’s Job]]></title>
            <link>https://medium.com/@cem.karaca/the-nerds-are-coming-for-everyones-job-f3a20b017e24?source=rss-61eca6c8ef85------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3a20b017e24</guid>
            <category><![CDATA[technology-disruption]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[creativeai]]></category>
            <dc:creator><![CDATA[Cem Karaca]]></dc:creator>
            <pubDate>Tue, 21 Oct 2025 05:55:29 GMT</pubDate>
            <atom:updated>2025-10-21T05:55:29.507Z</atom:updated>
            <content:encoded><![CDATA[<p>How I Bypassed an Entire Industry to Send My Daughter a Message</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/928/1*2wLWr27tx7aQBu93QsyQ_A.png" /></figure><p>I woke up this morning and made my coffee with my high-pressure espresso machine <strong>[barista bypassed: 1]</strong>, carefully propping my Americano glass against the refrigerator door as ice cubes clinked into the glass.</p><p>I was feeling a little down. I’d had a difficult conversation with my daughter, and instead of sending her yet another text message that might go unread <strong>[family counselor bypassed: 2]</strong>, I decided to speak her language: rap music.</p><blockquote>She might not always listen to me, but she understands rap.</blockquote><h3>The Two-Hour Revolution</h3><p>Sitting at my desk, I explained my idea to ChatGPT <strong>[songwriter bypassed: 3]</strong> <strong>[lyricist bypassed: 4]</strong>. Within minutes, I had professional-quality lyrics and style suggestions. In the past, this would have meant hiring a songwriter, paying for studio time, and waiting weeks or months for a finished product — assuming I could even afford to walk through a record label’s door.</p><p>I pasted the lyrics into <a href="https://suno.com/s/aZ5DmGWQk5NI1vWc">Suno</a> <strong>[composer bypassed: 5]</strong> <strong>[session musicians bypassed: 6–10]</strong> <strong>[recording studio bypassed: 11]</strong> <strong>[sound engineer bypassed: 12]</strong> <strong>[mixing engineer bypassed: 13]</strong> <strong>[mastering engineer bypassed: 14]</strong>. After a few adjustments, I had created something beautiful — music that perfectly captured my current emotions and the message I wanted to convey.</p><p>Then came <a href="https://www.canva.com/design/DAG2SbmCCeA/j0Fwimd2qL2E6wDFu91HQA/watch?utm_content=DAG2SbmCCeA&amp;utm_campaign=designshare&amp;utm_medium=link2&amp;utm_source=uniquelinks&amp;utlId=h0dcc40b775">Canva</a> <strong>[video editor bypassed: 15]</strong> <strong>[motion graphics designer bypassed: 16]</strong> <strong>[colorist bypassed: 17]</strong>. I added stock video <strong>[videographer bypassed: 18]</strong> <strong>[film crew bypassed: 19–23]</strong> to my freshly minted track and overlaid the lyrics on screen <strong>[typography designer bypassed: 24]</strong>. The result was remarkable.</p><p>No upload to a platform needed <strong>[distributor bypassed: 25]</strong> <strong>[record label bypassed: 26]</strong> <strong>[marketing team bypassed: 27–29]</strong> <strong>[publicist bypassed: 30]</strong>.</p><p>I put the video on my Instagram for her to watch.</p><p><a href="https://www.instagram.com/reel/DQA6ziGjtSk/?utm_source=ig_web_copy_link&amp;igsh=MzRlODBiNWFlZA==">https://www.instagram.com/reel/DQA6ziGjtSk/?utm_source=ig_web_copy_link&amp;igsh=MzRlODBiNWFlZA==</a></p><h3>The Final Count: 30 Professionals Bypassed</h3><p>In two hours, I had eliminated the need for:</p><ul><li><strong>1 barista</strong> (my morning coffee)</li><li><strong>1 family counselor</strong> (direct emotional communication)</li><li><strong>2 writers</strong> (songwriter and lyricist)</li><li><strong>6 music professionals</strong> (composer, musicians, engineers)</li><li><strong>3 studio positions</strong> (recording, mixing, mastering)</li><li><strong>6 video professionals</strong> (editor, designer, colorist, videographer, crew)</li><li><strong>1 typography specialist</strong></li><li><strong>5 business intermediaries</strong> (distributor, label, marketing team, publicist)</li></ul><p><strong>Total: 30 human jobs condensed into one morning’s work.</strong></p><p>The message was simple but important:</p><blockquote>No matter how much we argue, her father will always be there for her.</blockquote><h3>The Great Disruption</h3><p>Artificial intelligence isn’t fundamentally different from other tools that nerds have unleashed upon the world. When wielded correctly, it eliminates intermediaries and allows you to achieve your goals directly, without navigating professional hierarchies and industry politics.</p><p>This isn’t about the ethics of disruption — that’s a conversation for another day. This is about the reality that professions, passed down through generations and protected by academic institutions and regulatory bodies, are suddenly <strong>vulnerable</strong>.</p><blockquote>Jobs that once required diplomas, certifications, and the blessing of industry gatekeepers are now accessible to anyone with a laptop and curiosity.</blockquote><p>The lobbies and filters that once determined who could create, who could publish, who could produce — they’re watching their walls crumble. And they’re responding with the same ferocity that has marked every technological disruption in history.</p><h3>History Repeating Itself</h3><p>We’ve seen this movie before:</p><ul><li>Calligraphers opposed the printing press <strong>[hundreds of scribes replaced by one machine]</strong></li><li>Hotels fought Airbnb <strong>[front desk, housekeeping, concierge democratized]</strong></li><li>Taxi drivers protested Uber <strong>[dispatcher eliminated entirely]</strong></li><li>Traditional media battled social platforms <strong>[editors, publishers, broadcast engineers are obsolete]</strong></li></ul><p>Soon, we’ll witness the next chapter: those who don’t use AI opposing those who do. The battle lines are already being drawn between the gatekeepers of the old world and the democratizers of the new.</p><h3>The Nerds Have Already Won</h3><p>The truth is, the disruption isn’t coming — it’s here. Every morning, millions of people wake up and use AI to write <strong>[copywriter]</strong>, create <strong>[artist]</strong>, design <strong>[designer]</strong>, code <strong>[programmer]</strong>, and produce <strong>[producer]</strong>. They’re not waiting for permission. They’re not seeking approval from traditional authorities. They’re just creating.</p><p>My daughter received my rap video. She listened to it. She understood the message. The entire traditional music industry infrastructure — all 30 professionals that would have stood between my emotion and her ears — simply didn’t exist in our transaction.</p><p>That’s the real revolution: not that AI can make music, but that it can make <em>your</em> music, deliver <em>your</em> message, solve <em>your</em> problem — without asking anyone’s permission.</p><p>The nerds aren’t coming to take everyone’s job. They’re already here, and they brought tools for everyone.</p><p><strong>And we’re all counting.</strong></p><p><em>What’s your bypass count? How many intermediaries have you eliminated with AI today? Share your story in the comments.</em></p><p><em>If this resonated with you, follow me at </em><a href="https://medium.com/@cem.karaca"><em>@cem.karaca</em></a><em> for more stories about AI disrupting traditional gatekeepers.</em></p><p><em>Views are my own. The “bypass count” illustrates AI’s impact, not a dismissal of human creativity.</em></p><p>👏 <em>Clap if you’ve bypassed an intermediary with AI today</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3a20b017e24" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>