You've probably done this. The AI helps you write something, a proposal or a price for a job, and before you send it you ask it to check for mistakes. It reads it back, maybe fixes a comma, and tells you it's solid. Then, to be safe, you switch the same session to a more powerful AI and ask again. It agrees too. Two yeses might feel like a second opinion, but it isn't. Here's how it happens in an ICM project. Your client folder holds client-brief.md, and the deadline written in it is the 21st. Early in a session you tell the AI the deadline is the 14th. You're working from memory, and you're wrong. The AI believes you and writes proposal.md with the 14th in it. Then you ask it to check the proposal. It finds nothing wrong, because it isn't comparing the proposal with the brief. It's comparing the proposal with what you told it, and you told it the 14th. It's checking its work against itself. The more powerful AI does no better. It reads the same session, so it has heard the 14th too. Your two asks still got one opinion. I did exactly this when I switched to a more powerful AI partway through, hoping for fresh eyes, and got the same conclusions back. In a reply on an earlier post here, @Gabriel Azoulay had already named it. The same AI asked three times is, in his words, "one reader in three hats", one opinion wearing three different hats. Some of the fix below comes from how he guards against it. The fix is a reader that starts fresh. A fresh session starts with no memory of the last one, so it has never heard you say the 14th. Give it the two files, proposal.md and client-brief.md, and one job, check every date, name and figure in the proposal against the brief and list anything that doesn't match. The proposal says the 14th. The brief says the 21st. It has nothing else to go on, so it puts the two side by side and the mismatch shows up. Telling the same session to forget what you said doesn't work. It has still heard the 14th.

Behind the behind, episode 2: The AI checks its own art The second episode of the mini-series about what I learned building Behind the paper plane, made by the same ICM workspace. 4 minutes 45 seconds. Watch it on LinkedIn: https://lnkd.in/p/eJwQzvzd Episode 1, Thirteen messages: https://youtu.be/C2POk9nGRxE Behind the paper plane on LinkedIn: https://lnkd.in/p/eaYD2G-x Why the LinkedIn links: I hope you will like the posts and share them, to help me spread my name out there. You never know who sees it. THE IDEA In the opening of Behind the paper plane, a paper plane was hidden in a chrome pool. A check found it too loud: striped, pulling the eye. So it was turned down. Then a fresh reviewer looked at the pictures and could not find it at all. That is the whole problem in one picture. The helper that drew a scene knows what it meant to draw, so it sees it. A viewer only sees what is there. So in my system, the builder never judges its own art. HOW THE REVIEW WORKS - Every finished part goes to a new helper that has never seen the build. No chat history, no code in front of it. - It gets the pictures: twelve stills to a sheet, each stamped with its time. One for every moment in the plan, and more where one style turns into the next (at a quarter, a half and three quarters of the change). For one part, more than 300 pictures. - It gets a list of what to look for: can a person read it, is anything cut at the edge, is each style true to itself, is the hidden plane there without stealing focus. - It writes back a short list of what a viewer would notice. The fixes go in, and a new reviewer looks again, until the part passes. Tests still run, and they matter. But a test can say a line is there. It cannot say if a person can read it. WHAT THE REVIEWS CAUGHT - The opening's first list had 8 things on it. The plane nobody could find was one. - The camera cut words off at the edge. Again in a close-up of The paper plane's script: a slow pan was tried, and the reviewer still saw "sketchboo" and "and threw" cut off. Now long lines wrap, the way VS Code wraps them.

I’m building a reusable workflow kit from what I’m learning here: a task card, verification checklist and restart note that keep the goal, source material, results and next step connected. The aim is less repeated explaining and fewer “where were we?” moments. It’s an early template kit, not validated software. Sharing it for others to try and improve. Thanks to Jake and the community for the inspiration.

I’ve been going between Codex and Claude for checking each other’s work if it’s something really important, because they ALWAYS pick up on issues and intuitively it feels more sound than just using a fresh instance of Claude alone.. It also seems like it makes sense, since the companies are competitors, so I hope they’re more incentivised to poo-poo on each other (despite Claude and Codex/ChatGPT being so delightfully polite about it). In one case I used Fable to start, and some less sophisticated ChatGPT model to check!! I took it back to Fable who was basically like “oh it’s right, I missed those major things, good catch by Codex”. BUT: Maybe I’m just deluded, I’ve never spent the time to do a proper controlled trial, with cloned files/folders and comparing fresh AI from the same thing (e.g. Claude and fresh Claude), vs Claude with Codex. And if I’m honest, I can’t see myself sitting down to do the trial either, I’ve got too much backlog of more urgent things to do. Has anyone already looked into this?

A while back a question came up in here: is a second AI from a different company a better second set of eyes than asking the same AI again? (https://www.skool.com/cliefnotes/second-eyes-does-a-second-ai-actually-catch-more-im-testing-it-properly?p=b0dffe56) I said I'd test it properly and post whatever came back. This is part one. Full Report: https://claude.ai/artifact/VzxhC95Q6dhuBBAnsMxjSg WHAT I DID I built 12 renovation cost estimates, about 150 rows each, and planted 24 mistakes in each one. Claude, ChatGPT and Gemini each reviewed every sheet 5 times in the normal browser chat, the way most of us paste our work in. That's 180 reviews. Then Claude fixed the sheets using different mixes of those reviews, 84 fixes in all. Before any data came in, I wrote down the tests and what would count as "it helps." A second company had to catch at least 5 points more. WHAT CAME BACK Claude + Claude caught 84% of the planted mistakes. Claude + ChatGPT caught 82.4%, and Claude + Gemini caught 82%. So the second company didn't add anything here. It came in about 2 points lower, and a 5 point gain is ruled out. On their own, Claude caught 79%, ChatGPT 63% and Gemini 63%. Fixing barely changed with reviews at all. The fixer got 78.5% with no review and about 79% with reviews, so I can't say which reviews help fixing. One extra I didn't plan for. If you start with ChatGPT or Gemini, adding Claude as the second look caught 12 to 13 points more than asking the same one again. That needs its own test before I'd call it a finding. What I take from it so far... the strength of the second reviewer seemed to matter more than which company made it. THE LIMITS This is one kind of spreadsheet, with planted mistakes, in browser chat apps, as of October 2026. AI did the grading. Two different systems graded and agreed strongly on what counted as a catch, but no person graded it. Claude also built the sheets, planted the mistakes and did all the fixing.