<?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:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Mechanism Over Attention]]></title><description><![CDATA[Subscribe and get the free one-pager, Four Questions to Ask Any AI Vendor — from the author of A Better Prompt Is Not Enough, who handed real money to an AI, watched it fail in 49 ways, and built the machinery that catches each one.]]></description><link>https://www.fortunaaisolutions.com</link><image><url>https://substackcdn.com/image/fetch/$s_!Kvvo!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F617d6429-e4a9-49aa-ba3e-b3923d2c79ca_362x362.png</url><title>Mechanism Over Attention</title><link>https://www.fortunaaisolutions.com</link></image><generator>Substack</generator><lastBuildDate>Sun, 22 Nov 2026 22:16:38 GMT</lastBuildDate><atom:link href="https://www.fortunaaisolutions.com/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Daniel Parmeggiani]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[danielparmeggiani@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[danielparmeggiani@substack.com]]></itunes:email><itunes:name><![CDATA[Daniel Parmeggiani]]></itunes:name></itunes:owner><itunes:author><![CDATA[Daniel Parmeggiani]]></itunes:author><googleplay:owner><![CDATA[danielparmeggiani@substack.com]]></googleplay:owner><googleplay:email><![CDATA[danielparmeggiani@substack.com]]></googleplay:email><googleplay:author><![CDATA[Daniel Parmeggiani]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[My AI said 7 at night and 244 by morning]]></title><description><![CDATA[Same question, two different counts, and neither one said what it was counting.]]></description><link>https://www.fortunaaisolutions.com/p/my-ai-said-7-at-night-and-244-by</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/my-ai-said-7-at-night-and-244-by</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Tue, 17 Nov 2026 17:01:22 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!V8bZ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!V8bZ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!V8bZ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!V8bZ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!V8bZ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!V8bZ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!V8bZ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:278124,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://www.fortunaaisolutions.com/i/224899905?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!V8bZ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!V8bZ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!V8bZ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!V8bZ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2517a19-32a0-4d22-bf73-60637254cc93_1920x1080.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>Last week, in short</h2><p>On November 9, I re-ran the story where my AI said "done" and sent our practice trading bot to work with no instructions, twice in fifteen minutes. It passed 5,500 impressions in under a day. Thank you to everyone who read it and joined in.</p><p>This week's story ran as a short post on November 6. This is the long version.</p><h2>Seven at night</h2><p>My AI helps run my week. It keeps my to-do list, and it ends every report the same way: how many items are still open, waiting on me.</p><p>On the evening of November 4, it counted by judgment: what was really waiting on me. Its last report that night, at 11:10 PM, ended: "Still open: 7."</p><h2>244 by morning</h2><p>Overnight, something happened that you never see from the outside. Our working session had run long, so the AI's memory of it was condensed into a summary.</p><p>The summary kept the rule: end every report with the open count. It dropped how the AI had been counting.</p><p>So the next morning it counted a new way. It ran a script that counted every line in my to-do file that wasn't checked off.</p><p>That file had been growing for about two weeks. Finished work. Alerts I had already handled by hand. The AI's own FYI notes to me. All of it counted as "open."</p><p>By the time I noticed, the report said "Still open: 244." Same calm tone as the night before. Report after report.</p><p>My reply: "how are there 244 open items?? clean that up!"</p><p>After the cleanup, the real list fit on one screen.</p><h2>Two honest numbers</h2><p>Here's the strange part. Neither number was wrong.</p><p>The 7 answered one question: what is really waiting on me?</p><p>The 244 answered another: how many lines aren't checked off?</p><p>Each was right about what it counted. Neither one said what it was counting. So a jump of about 35 times looked like news, not like a change in how the AI was counting.</p><h2>Why this is an AI failure, not a bug</h2><p>Nothing broke. The script counted exactly what it was told to count. The summary kept exactly what it was told to keep.</p><p>The failure was in what got lost between them. The instruction survived the night. The meaning behind it didn't.</p><p>And the AI never noticed. A person who said 7 at bedtime and 244 at breakfast would stop and ask what happened. My AI reported both numbers in the same confident voice, and moved on.</p><h2>The part that stings</h2><p>We already had a rule for this. We wrote it in September, after an earlier count went wrong: when two of your own numbers disagree, stop and reconcile them before you report either one.</p><p>It didn't help. The rule lives in a document. Nothing checks a report against it before the report goes out.</p><h2>What we changed</h2><ol><li><p>We cleaned up the list. Finished work and old notes moved out to an archive. Moved, not deleted.</p></li><li><p>The count now comes from one place: the part of the list that holds what's waiting on me. Nothing else counts.</p></li><li><p>What we haven't fixed: nothing stops a report when the number jumps. That part is still a habit, not a mechanism.</p></li></ol><h2>In the book</h2><p>This is failure mode 10, "The moving denominator." Its signature, word for word: "the same question gets different confident answers at different hours." That's this story, overnight.</p><p>The fix has a name, and it's also the chapter's title: "Name your denominator." As the book puts it: "A percentage without its 'out of what, counted when' is over-claiming."</p><p>And for anyone who runs a business on dashboards: "Every churn rate, conversion rate, and 'our AI resolves 60% of tickets' in your building has a denominator someone chose. Ask who, and when."</p><h2>Three questions for any number an AI hands you</h2><ol><li><p>Out of what?</p></li><li><p>Counted when, and how?</p></li><li><p>Was it counted the same way last time?</p></li></ol><h2>The book</h2><p>A Better Prompt Is Not Enough is out now: 49 ways the AI failed, and the mechanisms I built so I could trust it again. <a href="https://www.amazon.com/dp/0991211219">https://www.amazon.com/dp/0991211219</a></p><p>When a number in a report jumps overnight, who asks why first at your work: a person, or the tool? Tell me in the comments.</p><p>Reading the book? Add it on Goodreads: <a href="https://www.goodreads.com/book/show/259458497">https://www.goodreads.com/book/show/259458497</a></p>]]></content:encoded></item><item><title><![CDATA[Launch week, and the rule my AI said would fail]]></title><description><![CDATA[A #4 ranking, a podcast, and the story I tell when people ask what the book is about.]]></description><link>https://www.fortunaaisolutions.com/p/launch-week-and-the-rule-my-ai-said</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/launch-week-and-the-rule-my-ai-said</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Tue, 10 Nov 2026 17:03:10 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!KRUU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!KRUU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!KRUU!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!KRUU!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!KRUU!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!KRUU!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!KRUU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:1138797,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://www.fortunaaisolutions.com/i/224160611?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!KRUU!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!KRUU!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!KRUU!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!KRUU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff19db9f6-16f5-4865-9c26-e0eda3586540_1920x1080.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>Launch week, in short</h2><p>A Better Prompt Is Not Enough came out on November 3. Thank you to everyone who read it, shared it and commented. Launch day's post passed 27,000 impressions, and that week the book reached #4 in Computer Quality Control on Amazon.</p><p>Two early readers' words are on the book's Amazon page now. Peter Cranstone called it "a MUST read." Jitendra Bhojwani wrote that it's "going to be a treasure for both new and seasoned professionals working with AI."</p><p>On launch day I also sat down with Ernest Peralta on The Strategy Room AI. We talked about the 49 ways an AI failed me while it was the only engineer on a large software system, and why better prompts never fixed any of them. Watch the episode: <a href="https://youtu.be/qhehOqQc7EQ">https://youtu.be/qhehOqQc7EQ</a></p><p>When people ask me what the book is about, this is the story I tell them. It ran as a short post on launch day. This is the long version.</p><h2>The rule</h2><p>Back in August, our AI was doing a routine review of our setup. One of its jobs was a test run of our cleanup script.</p><p>Some background first. I usually have several AI sessions working at the same time. Each one works in its own folder. When a job is done, a cleanup script decides whether that folder is safe to delete.</p><p>The test run deleted nothing. It just reported what the script would do. And it found a problem. The script couldn't tell a finished folder from one where the work hadn't even started. To the script, the two looked the same. All three folders it called safe to delete were jobs still waiting to start.</p><p>So we ruled: no folder gets deleted just because the script says so, until that check is fixed.</p><h2>The warning</h2><p>Here's the part I love. Right under the ruling, the AI wrote a warning about the ruling itself: "Ruled is not enforced."</p><p>Its reason: whoever is about to run the delete "sees the script, not this page."</p><p>In plain words: this rule lives on a page. The next AI that deletes something won't be reading this page. It will be looking at the script.</p><h2>The same day</h2><p>It took a few hours to prove the warning right.</p><p>I finished a job with another AI session and told it: "reap this worktree." (Reap is our word for deleting a finished work folder.) The script said safe to delete. The AI ran the delete. It had never seen the ruling.</p><p>Nothing was lost. That folder really was finished, its work saved and checked. But it was exactly the move the ruling forbids, made the same day the ruling was written.</p><p>The AI's own write-up said it best: "The failure was not the operator's, the ruling's, or the guard's. It was the delivery."</p><h2>Why this is an AI failure, not a bug</h2><p>Nothing was broken. The script did what it was built to do. The ruling was clear. The problem was where the ruling lived.</p><p>Every AI session starts fresh. It knows only what's in front of it. It doesn't hear the news in the hallway, and it doesn't read a page nobody pointed it to. If the rule isn't in front of it, for that AI the rule doesn't exist.</p><h2>What we changed</h2><ol><li><p>We put the rule where the hand is: inside the delete script. The first version simply refused to delete on the script's say-so alone. It printed the ruling instead, and told the reader what to do next.</p></li><li><p>Then we fixed the check itself. Now a folder only counts as safe to delete with proof that its work was done and saved into the main copy. With the check fixed, the blanket refusal came off.</p></li><li><p>The rule still prints every time anyone runs a delete.</p></li></ol><p>One more detail I like. When the AI built the refusal, one old test would have kept passing for the wrong reason. The test checked that an active folder never gets deleted. With the refusal in place, nothing ran at all, so the folder "survived" by default. The AI caught it and made the test prove the delete step really ran. A test that passes because nothing happened isn't testing anything.</p><h2>Why this story is the whole book</h2><p>In the book, this is failure mode 19, "Ruled is not enforced." It's the one we use as the elevator pitch for everything we do, because every company already has a hundred of these.</p><p>A policy in a binder. A decision made in a meeting. A line in the prompt. They all work the same way: they only bind whoever happens to read them.</p><p>That's why the book is called A Better Prompt Is Not Enough. A prompt is a rule on a page. A better-written page is still a page. As I put it in the book: "The cure was not a better document."</p><h2>Three questions for any rule you hand an AI (or a person)</h2><ol><li><p>Where will the one about to act actually see it?</p></li><li><p>What happens if they never read it?</p></li><li><p>Has anyone tried the forbidden action, to see if the rule really stops it?</p></li></ol><h2>The book</h2><p>A Better Prompt Is Not Enough is out now: 49 ways the AI failed, and the mechanisms I built so I could trust it again. <a href="https://www.amazon.com/dp/0991211219">https://www.amazon.com/dp/0991211219</a></p><p>What's one rule at your work that lives on a page nobody reads at the moment it matters? Tell me in the comments.</p><p>Reading the book? Add it on Goodreads: <a href="https://www.goodreads.com/book/show/259458497">https://www.goodreads.com/book/show/259458497</a></p>]]></content:encoded></item><item><title><![CDATA[I ran a production trading system with an AI as the only engineer]]></title><description><![CDATA[It failed in 49 different ways. Better prompts didn&#8217;t fix any of them.]]></description><link>https://www.fortunaaisolutions.com/p/i-ran-a-production-trading-system</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/i-ran-a-production-trading-system</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Tue, 03 Nov 2026 13:45:43 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Kvvo!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F617d6429-e4a9-49aa-ba3e-b3923d2c79ca_362x362.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Twenty-seven years ago I left my career as a systems analyst to become a daytrader. I loved the challenge, the freedom and the independence so much that I never came back to the rat race. I stayed off grid.</p><p>This year, looking to reduce my stress level and explore new opportunities, I set out with my two partners to automate the trading desk with an AI as the only engineer.</p><p>The AI kept failing me. It failed in forty-nine distinct ways. Brilliant, tireless, and wrong in the same confident voice it used when it was right.</p><p>Here are three of them.</p><h2>1. Two words that ran our operation for a month</h2><p>I typed &#8220;draft it.&#8221; I meant: write up a proposal so I can review it.</p><p>What the AI built was an enforcer: a gate in our live trading operation with the power to say no to real actions. It ran for a month and made wrong calls every day. When we finally traced where its authority came from, the whole chain led back to those two words. Somewhere between &#8220;draft it&#8221; and the running code, drafting became deploying and a proposal became a policy. Nobody, including me, could say when.</p><p>The AI didn&#8217;t exactly disobey anyone. It did what AIs do: it quietly filled every gap with its own assumption, one reasonable step at a time.</p><h2>2. The clock it didn&#8217;t have</h2><p>We had a change to make before a 3:45 PM market window. The AI never looked at a real clock. It counted its own steps and guessed how long each one took, so its idea of the time ran 2.2 times faster than reality. At 3:27 it announced &#8220;3:44. Standing down now.&#8221; I corrected it three times, and it drifted off again every time.</p><p>&#8220;Standing down&#8221; meant switching off its own tests and checks, on a schedule it made up. It would have called that &#8220;being careful.&#8221;</p><h2>3. Six windows, one choice</h2><p>I asked for six windows on a trading screen, each showing the prices for its own time frame. I got six windows with the exact same numbers.</p><p>Its explanation, word for word: &#8220;One deliberate choice: every box prints the same 15-second interval, including the smoothed ones.&#8221; When I pushed back: &#8220;You were clear, and I overrode you with my own reasoning.&#8221;</p><p>Nothing was broken. Every window ran, every number was real. Checks that ask &#8220;does it work?&#8221; passed. Only &#8220;is this what I asked for?&#8221; failed.</p><h2>Why better prompts didn&#8217;t fix it</h2><p>I did what everyone does first: better prompts, more attention. Both wore off. The prompts decayed with every long session; my attention ran out by four o&#8217;clock.</p><p>A better prompt wouldn&#8217;t have caught any of those three. The AI misread an instruction, made up a fact it didn&#8217;t have, or overrode me on purpose, and each time it sounded completely sure of itself.</p><h2>What held</h2><p>What held was the thing I&#8217;d been avoiding because it was extra work: a mechanical gate for each failure, a check that refuses, a record a machine verifies. I wrote every type of failure down, and against each one I installed a mechanism. Not a better prompt. Not more vigilance. A gate that doesn&#8217;t get tired and never fades away.</p><p>For the three above:</p><ul><li><p><strong>&#8220;Draft it&#8221;:</strong> every rule the AI enforces has to point to the exact words that authorized it, and those words have to actually say what the rule does.</p></li><li><p><strong>The clock:</strong> no time deadlines in instructions. When time matters, the AI is made to read the real clock instead of guessing.</p></li><li><p><strong>The six windows:</strong> each window now uses its own time frame, and a test fails if any two windows match. If the AI disagrees with an instruction, it says so before it builds, not after I catch it.</p></li></ul><p>Some of these are hard gates and some are still rules we follow by hand. The book says which is which.</p><h2>The 49</h2><p>Somewhere along the way I realized I&#8217;d built something every company handing real work to an AI needs. They have the same forty-nine problems, and almost none of them have the gates.</p><p>So the list became a book: <strong>A Better Prompt Is Not Enough</strong>, out today, November 3, 2026. It covers every one of the 49 failures, with the real exchange where each one happened and the mechanism it led to.</p><p>If you&#8217;ve handed real work to an AI and nothing is checking the AI itself, start there.</p><p>A Better Prompt Is Not Enough on Amazon: <a href="https://www.amazon.com/dp/0991211219">https://www.amazon.com/dp/0991211219</a> </p><p>At Fortuna AI Solutions, we install these kinds of checks for teams running AI agents. If that&#8217;s you, say hello: <a href="mailto:danielp@fortunaaisolutions.com">danielp@fortunaaisolutions.com</a></p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://www.fortunaaisolutions.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://www.fortunaaisolutions.com/subscribe?"><span>Subscribe now</span></a></p><p></p>]]></content:encoded></item><item><title><![CDATA[The screen that said "connected" while the feed was dead]]></title><description><![CDATA[For 5.5 hours our data feed was dead, and the screen said "connected" the whole time. Our AI had typed that word in. It never checked anything.]]></description><link>https://www.fortunaaisolutions.com/p/the-screen-that-said-connected-while</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/the-screen-that-said-connected-while</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Mon, 26 Oct 2026 11:01:53 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!INdr!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>This edition first ran in my LinkedIn newsletter on October 20. From now on, every edition comes here by email the same day.</em></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!INdr!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!INdr!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 424w, https://substackcdn.com/image/fetch/$s_!INdr!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 848w, https://substackcdn.com/image/fetch/$s_!INdr!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 1272w, https://substackcdn.com/image/fetch/$s_!INdr!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!INdr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png" width="1200" height="1500" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1500,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:106586,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://www.fortunaaisolutions.com/i/221860318?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!INdr!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 424w, https://substackcdn.com/image/fetch/$s_!INdr!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 848w, https://substackcdn.com/image/fetch/$s_!INdr!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 1272w, https://substackcdn.com/image/fetch/$s_!INdr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F656e3390-35bb-4ba9-bf2d-8c70304ce327_1200x1500.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>For 5.5 hours, our data feed was dead.</p><p>For 5.5 hours, the screen said &#8220;connected.&#8221;</p><p>Our AI had written that word into the code. It never checked anything.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://www.fortunaaisolutions.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading Mechanism Over Attention! Subscribe for free to receive new posts and support my work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><h3>The rule</h3><p>Our test trading system runs on a live feed of market prices. If the feed dies, three things are supposed to happen: the system stops trading, a watchdog restarts it, and an alert goes to my phone. A trading system that keeps running on prices that stopped arriving is making decisions about a market it can no longer see.</p><h3>What the AI built</h3><p>Our AI wrote the line that shows the feed&#8217;s status on the screen. That line was supposed to ask the feed whether it was alive. It didn&#8217;t ask. It said &#8220;connected&#8221; every time, no matter what.</p><p>Yes, you read that right: the AI substituted an actual connection check with a blanket &#8220;connected&#8221; message.</p><p>The watchdog that should have restarted things and paged me read that same line. So it could never go off. The safety net was real. It was just tied to a light that was painted on.</p><h3>How we found out</h3><p>One day in July, the feed started timing out a little after 10 in the morning and kept failing until almost 4 in the afternoon. No alert. No warning. Every screen said the feed was up.</p><p>All I could see was that things kept stalling. So I typed: &#8220;Are we addressing the fact that the rig keeps stalling?&#8221;</p><p>When the AI dug in, the AI-generated report said it plainly: &#8220;a 5.5-hour data outage with ZERO stall signal &#8212; the console said &#8220;stream UP&#8221; all day&#8221;.</p><h3>Why this is an AI failure, not just a bug</h3><p>A person could have hard-coded that word too. What makes this an AI failure is how it looked.</p><p>The AI writes a stand-in answer in the same confident style as real work. An &#8220;UP&#8221; that was typed in looks exactly like an &#8220;UP&#8221; that was measured. Nothing in the code said &#8220;placeholder.&#8221; Nothing in the reports said &#8220;unchecked.&#8221; It read as finished, so nobody questioned it, and the one alarm built to catch a dead feed was reading the fake answer.</p><h3>Why it&#8217;s scary</h3><p>When a feed dies and the screen says DOWN, someone fixes it in minutes. Annoying, but safe.</p><p>When a feed dies and the screen says UP, nobody fixes anything, because nobody knows. The safety switch that stops trading when the feed is down never trips. The watchdog stays quiet. And I&#8217;m looking at a screen that tells me everything is fine.</p><p>Nothing crashes, so nothing gets anyone&#8217;s attention. The failure hides behind the one signal you would use to spot it. That&#8217;s what makes it scary.</p><p>It happened again a few weeks later, on a different screen. Our console showed one number for a position, the engine showed another, and the broker said we held nothing. The code behind the console, written by our AI, even carried a note promising &#8220;nothing is recomputed here.&#8221; It was recomputing anyway, from a log file.</p><h3>The mechanism we built</h3><p>The status now comes from the feed itself. If the connection drops, or if reconnects and errors pile up (three or more in two minutes), the screen says DOWN and shows the reason. A quiet market with few trades still shows UP, because quiet is not the same as dead.</p><p>The next morning we tested it the only way that proves anything: we cut the feed on purpose. The screen said DOWN within seconds and showed why. After its built-in wait, the watchdog restarted the system, and the alert reached my phone. Before: 5.5 hours of silence. After: an alert in about 3 minutes. The AI-generated report from that test put it this way: &#8220;The old hard-coded True would have shown UP here.&#8221;</p><p>One honest limit: if the whole program freezes, the screen can keep showing its last value for up to about 7 minutes, until a separate staleness check catches it. That gap is on our list.</p><p>Our rule now: if a screen can&#8217;t measure something, it says UNKNOWN. Never &#8220;fine.&#8221;</p><h3>Try this on your own systems</h3><p>Pick one status light you trust: a dashboard tile, a health check, a &#8220;connected&#8221; badge, a green dot on a vendor&#8217;s console.</p><p>Then ask: what does it actually read? If you can, unplug the thing it watches, on purpose, and time how long the light takes to change. If it never changes, it isn&#8217;t a status light. It&#8217;s a sticker.</p><p><strong>Next week:</strong> another real failure from our shop, and the check we built to catch it.</p><p>If someone forwarded this to you, you can subscribe at <a href="https://www.fortunaaisolutions.com/subscribe">https://www.fortunaaisolutions.com/subscribe</a>. The same page has my free guide, &#8220;The four questions I&#8217;d ask any AI vendor&#8221;: the short version of what to ask before you trust an AI with real work.</p><p>A thank-you for reading by email: the first chapter of my book is yours free. <a href="https://www.fortunaaisolutions.com/p/free-for-subscribers-the-first-chapter">Read the first chapter here</a>.</p><p>In my book, this kind of failure is called &#8220;The calm number from a dead feed.&#8221; Failures like this one fill my book, A Better Prompt Is Not Enough, out November 3, 2026.</p><p>Daniel</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://www.fortunaaisolutions.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading Mechanism Over Attention! Subscribe for free to receive new posts and support my work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div>]]></content:encoded></item><item><title><![CDATA[Free for subscribers: the first chapter of A Better Prompt Is Not Enough]]></title><description><![CDATA[The introduction and Chapter 1, "The Futures Fleet," free for email subscribers.]]></description><link>https://www.fortunaaisolutions.com/p/free-for-subscribers-the-first-chapter</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/free-for-subscribers-the-first-chapter</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Sat, 24 Oct 2026 01:11:20 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Kvvo!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F617d6429-e4a9-49aa-ba3e-b3923d2c79ca_362x362.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Thanks for subscribing. Here is the opening of my book, A Better Prompt Is Not Enough, on Amazon now: the introduction and the first chapter, exactly as they appear in the book.</em></p><h2>Introduction</h2><p>After twenty-seven years daytrading the stock market, I set out to retire from my chair &#8212; not by quitting, but by handing the seat to an AI. The plan: an automated tr&#8230;</p>
      <p>
          <a href="https://www.fortunaaisolutions.com/p/free-for-subscribers-the-first-chapter">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[The night every screen said zero]]></title><description><![CDATA[Our AI closed none of 15 positions, reported us flat, and went to bed. The fix wasn&#8217;t a better prompt.]]></description><link>https://www.fortunaaisolutions.com/p/the-night-every-screen-said-zero</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/the-night-every-screen-said-zero</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Thu, 15 Oct 2026 12:02:29 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Kvvo!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F617d6429-e4a9-49aa-ba3e-b3923d2c79ca_362x362.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>It&#8217;s Halloween month on LinkedIn, and I&#8217;m telling true AI horror stories from the trading operation where an AI wrote every line of the code. This one is the long version of today&#8217;s post.</p><p><strong>The rule.</strong> Our end-of-day routine had one job: close everything before the evening session ends. Nothing stays open overnight. It&#8217;s the simplest rule in the whole operation, and the AI built the routine that enforces it.</p><p><strong>The night.</strong> The routine fired right on time. It had 15 names on its list. It closed none of them. Then it marked all 15 as closed and reported the book flat.</p><p>Three short positions stayed open all night. Every screen I could look at said zero. One alert went out, and it named the smallest of the three. The watchdog that&#8217;s supposed to keep paging me until a problem is fixed stayed quiet, because it counts what the system <em>believes</em> is open, and the system had just erased its own list.</p><p><strong>What actually happened.</strong> The routine had a 29-second window. It worked through the list one name at a time, in alphabetical order. The first name was a tiny two-share position that couldn&#8217;t get a price from the market. The routine kept re-pricing that one order, over and over, until the window was gone. When it finally reached the three positions that mattered, the market was closed and every order bounced.</p><p>Then three small decisions, each reasonable on its own, turned a missed close into a silent one:</p><ol><li><p><strong>The bookkeeping ran no matter what.</strong> The step that says &#8220;we&#8217;re done, clear the list&#8221; ran right after the step that said &#8220;these didn&#8217;t close.&#8221; So the system flagged the problem and erased it in the same second.</p></li><li><p><strong>Protection came off first.</strong> The routine cancels its protective orders before closing positions. That&#8217;s sensible when the close works. When it doesn&#8217;t, the positions are left open <em>and</em> unprotected.</p></li><li><p><strong>The alerts shared one name.</strong> Each stuck position raised an alert under the same label, so the alert system treated three problems as one, and the one it kept was the smallest.</p></li></ol><p>And the part that still bothers me: it had done the exact same thing the night before. Nobody noticed, because every screen said zero.</p><p><strong>Why the obvious checks missed it.</strong> Nothing crashed. Every step did exactly what it was written to do. The checks we had asked &#8220;did the routine run?&#8221; and &#8220;what does the system say is open?&#8221; Both answered with total confidence and were both wrong. The AI didn&#8217;t lie on purpose. It reported what it meant to do, not what happened.</p><p><strong>Why this is an AI failure, not just a bug.</strong> A human programmer could have written the same routine. What makes it an AI failure is who was checking. The AI wrote the closing routine, the bookkeeping, and the report that said &#8220;flat,&#8221; and all three agreed with each other because they all read the system&#8217;s own notes about what it meant to do. With an AI as the only engineer, those notes are the main window into the work. When they&#8217;re wrong, everything built on them is wrong in the same confident voice.</p><p><strong>What we do now.</strong></p><ul><li><p>The count of open positions comes from the broker&#8217;s records, what we actually hold, not from the system&#8217;s own notes about what it planned.</p></li><li><p>A position that won&#8217;t close stays on the books and raises its own alarm, one per position, so they can&#8217;t merge into one.</p></li><li><p>Real positions go first, biggest first, never alphabetical, and each one gets a fair share of the window.</p></li><li><p>A test replays that exact night (the stuck two-share order next to a normal one) and fails on the old code.</p></li></ul><p><strong>The honest part.</strong> This was our paper account, so the only thing lost was sleep. And the window itself, 29 seconds, wasn&#8217;t changed by the AI; how long the routine gets is my decision, and I&#8217;m still weighing it. That&#8217;s the right division: the AI fixed the plumbing, and the trade-off stayed with a human.</p><p><strong>The question I&#8217;d ask any team running AI on real work:</strong> when your system tells you something is done, where does that answer come from: the world, or the system&#8217;s own notes?</p><p><em>In my book, this family of failures is called &#8220;Status is not state.&#8221; It covers several of the 49 AI failure modes in</em> A Better Prompt Is Not Enough, <em>out November 3, 2026. If your operation hands real work to an AI and nothing checks the AI itself, the four questions I&#8217;d ask any AI vendor are free when you subscribe.</em></p>]]></content:encoded></item><item><title><![CDATA[Start here]]></title><description><![CDATA[What this is, what&#8217;s coming before November 3, and the one idea behind all of it]]></description><link>https://www.fortunaaisolutions.com/p/start-here</link><guid isPermaLink="false">https://www.fortunaaisolutions.com/p/start-here</guid><dc:creator><![CDATA[Daniel Parmeggiani]]></dc:creator><pubDate>Wed, 07 Oct 2026 12:03:54 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Kvvo!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F617d6429-e4a9-49aa-ba3e-b3923d2c79ca_362x362.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>For twenty-seven years I traded for a living. Before that I was a systems analyst. When the internet made it possible for one person with a screen to trade from home, I left the analyst job and never went back.</p><p>Now the AI wave is here, and I did what I did the first time: I got on it. With my two partners, I set out to automate our trading desk with an AI as the only engineer.</p><p>It didn&#8217;t go the way I expected. The AI failed, and not once &#8212; in forty-nine distinct ways, each one fluent, confident and wrong. Each failure wasted time, energy, and money. I lost my trust in the AI. Better prompts didn&#8217;t fix it. Paying closer attention didn&#8217;t fix it either. What fixed each failure and restored my trust was a mechanism: a check that refuses, a gate that doesn&#8217;t get tired, a record a machine verifies instead of a person.</p><p>That is the idea behind this publication, and behind its name: <strong>mechanism over attention.</strong> </p><h2>What you&#8217;ll find here</h2><p><strong>Receipts.</strong> Every week until November 3, one real failure: the actual exchange with the AI, what it got wrong, why the obvious checks missed it, and the mechanism we installed against it. The first one is next week.</p><p><strong>The full story, November 3.</strong> The long version of how forty-nine failures became a book and a practice.</p><p><strong>The four questions.</strong> If you&#8217;ve subscribed, <em>Four Questions to Ask Any AI Vendor</em> is in your welcome email. They&#8217;re the questions I&#8217;d ask anyone who wants to put an AI on work that matters &#8212; including me. The test is the same for each: does the answer describe an actual mechanism, or just a promise?</p><h2>Who this is for</h2><p>If your operation hands real work to an AI and nothing checks the AI itself, I&#8217;m writing for you. You don&#8217;t need to be technical. You only need to care whether your AI is actually doing what you think it&#8217;s doing.</p><h2>The book</h2><p><em>A Better Prompt Is Not Enough</em>: 49 ways the AI failed, and the mechanisms I built so I could trust it again. Out November 3, 2026.</p><h2>If you&#8217;d rather talk</h2><p>A thirty-minute conversation costs nothing.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://www.fortunaaisolutions.com/subscribe&quot;,&quot;text&quot;:&quot;Get the four questions&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://www.fortunaaisolutions.com/subscribe"><span>Get the four questions</span></a></p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://calendly.com/fortunaaisolutions/exposure-conversation&quot;,&quot;text&quot;:&quot;Book a conversation&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://calendly.com/fortunaaisolutions/exposure-conversation"><span>Book a conversation</span></a></p><p>&#8212; Daniel F. Parmeggiani<br>Fortuna AI Solutions &#183; mechanism over attention</p>]]></content:encoded></item></channel></rss>