I was retyping the same reminders every week to Claude. For months.
“ground it.” “no wall of text.” “check schema.” “wrong, stop.”
Every session, I had to force-inject these into the conversation flow with my AI.
What’s the problem with this?
I kept treating each reminder as a one-off annoyance, the kind that working with AI comes with. The problem was deeper and broader than that.
My first attempt at improving things: reading all the tips about AI setups
- “Best practices for your AI setup.”
- “10 rules every CLAUDE.md needs.”
- “The one prompt that will make your AI 99% better instantly.”
There’s no shortage of them. I read and watched a lot, and early on, I genuinely learned from them.
Then I stalled. I was spending more time disqualifying saved tutorials than learning from them, and FOMO kept me consuming anyway. Worse: the abundance started making me question concepts I had already internalized.
Not because those tutorials are wrong, but because, TO ME, they looked either generic or sold as the ultimate hack. And they talked about what breaks for other people, not my specific use case.
A moment of realization
My AI isn’t failing me the way these tutorials describe.
My AI fails in my own stack, on my recurring blind spots, on my weaknesses.
What actually works: chat transcripts (aka the evidence was already on my disk)
Every reminder I typed, every time I lost patience, every drift I caught: it’s all stored in transcripts. Months of AI sessions, stored in full.
So I fired up a new session, picked Fable 5, and went meta:
“go through all sessions and conversations you and I had for all projects (including Cowork). Analyze the way we talk, work, and surface trends and opportunities for creating skills and hooks that would improve greatly not only our communication… but also improving the quality of your outputs.”
It dispatched 8 agents over 5.4MB of my own messages: 160+ sessions across all my projects, months of work.
What came back was like time travelling through the way I had worked up until that moment.
- “ground it, don’t infer”: retyped 40+ times.
- The same deploy instruction: 30+ times.
- One 500-word instruction block I’d been hand-pasting into sessions: 7+ times.
- Reminders to keep answers short: in every single batch it analyzed.
Each one a thing I’d been treating as a one-off annoyance, now written out with a count next to it.
Once you see it as a list, you can build against it. This is the pattern everything useful in my setup comes from:
"did agents review this?" ──► became an agent-review-gate hook.
"no wall of text" ──► became the terse output style I currently use.The first warns the model when it heads toward a commit without a review.
The second caps every answer to what I actually need to read.
The most painful find? Two guardrails I’d built months ago were there, switched off. Nothing surfaced it until I time-travelled through my transcripts.
So what, Matteo?
- Best-practice lists and tutorials tell you what fails for other people. Good as a starting point, but potentially weaker as a foundation.
- Any AI model is a doer, planner, fixer, but also a silent “witness” to its own failures and limitations. Tell it to look at its work more frequently.
- The ways to improve are hidden in plain sight: in your session transcripts.
I spent months looking outward for the fix: the right tutorial, the perfect CLAUDE.md, the one prompt that finally makes it work. The whole time, the answer was sitting in my own history, written in my own frustration. Nobody else’s best-practice list knows which mistake I make on a Tuesday afternoon, on my stack, for the tenth time. My transcripts do.
P.S.: On the AI-hype fatigue and, more broadly, the AI-is-the-solution-to-all-problems narrative many seem to be piling on: read Elena Verna’s “Please stop the AI Confidence Theater”. It perfectly describes this dystopian-y and lun-AI-tic world we live in.