Planes keep a black box so that when something goes wrong there is a full recording of what happened and when. Build the same thing into your ICM system: a logbook it writes itself, every job and every journal, that you can hand back to Claude at any point.
The setup is simple. Every Claude Code job writes a short record when it finishes - what it did, how it went, a grade, and any errors. Every design chat writes a journal when it is done drafting your ideas out. Every handover is stored in a set space.
All of the above - ORGANISED
Nothing is thrown away. The record just builds on its own in the background.
Then, any time you have spare usage, point Claude at the whole logbook: pull the bits that ran slow, the bits that failed, and the improvements worth making.
Once the logbook exists, here is what it gives you:
- Query anything - ask in plain language what happened, when, and why.
- Find the mistakes - surface exactly where and when something went wrong.
- Improve the whole system - point at the record and tell it to fix the system as a whole
- Security audits - every change that was made, timestamped and ready to review.
- Client work - hard evidence and the answer already sitting there if a client complains.
- Investor conversations - the same record holds up when you are showing what you have built.
- Older systems - go back and read what a system used to do long after you have moved on from it.
- Across projects - copy edits and systems from one project directly into another with ease
And this is where it goes next. Fairly soon you will be able to hand a model every logbook across every project you run and let it treat the whole thing as training data - not to answer questions about the past, but to learn how you actually operate and carry that forward on its own. The record stops being something you query and becomes the material a system clones itself from: your decisions, your patterns, the way you work, all already written down. That is the real reason to start now. Every entry you log today is training data for the autonomy you hand off tomorrow.
If you have got extra Fable usage right now, this is a strong thing to point it at over the next couple of days, before it turns back into credits. The setup is mostly plumbing - a template each job fills in and a place to put it - and the analysis pass on top gives you genuinely useful output for very little. Build the recorder now while the usage is effectively free, and the black box is ready the first time something breaks.
I have attached below my LOG_PROMPT.md, which I give to Claude Chats, and the CC_TOUCHDOWN.md, which is hard coded as the last step in every Claude Code prompt. Open for feedback as always!
Are you logging what your system actually does, or reconstructing it after the fact when something goes wrong? And anybody currently with a logbook system, how is yours designed and are there any other use cases that you would be willing to share with the community?