A new agent recovered the state of the project. Then a second model did the same
In the first pilot I delivered the System session on a client’s real project. It was planned for 90 minutes and took about 108. Here is what happened, which part of the criterion we closed and which part is still open. The client stays anonymous: I don’t have their consent to use their name, so I don’t use it.
- First session
- ~108 min
- Same recovery
- 2 models
- Full criterion
- 1/2
90 planned, about 108 actual. On the project, with the client at the keyboard.
The same folder held up even after switching vendors.
One question closed. The second one opens the next block.
The project was there. The agent didn’t know where to look.
A client who uses AI in their work. The problem wasn’t using the model: it was making the state of the project recoverable even from a new session, without pointing it to the right file by hand.
The fear was concrete: switching tools and losing the operating context built up over time. We worked within what could be proven: the project’s facts in its own files, explicit routing and the same recovery run again on a second model.
About 108 minutes on the part that matters.
The client had set up the skeleton on their own before the session, with the pre-session kit. Interview done, CLAUDE.md and project already in place when the call started. So the time didn’t go into setup: we spent it seeing where their file starts to bloat, where the routing lies and how to fix it.
In the session, with their hands on the keyboard, we built four things.
- A lean CLAUDE.md. Kept short on purpose, because it’s injected into every session: if it bloats, it eats your context window.
- A routing table built on their real facts. Every row points to a file with real content in it, not an empty folder.
- Their real project inside the system. Not a made-up example: the actual work, the work they really care about finding again.
- All of it under git, locally. So every step is versioned and recoverable, and tomorrow they can move the system wherever they want.
It was their hands on the keyboard almost the whole time. I set the criterion and showed them where they would go wrong, before they did.
The criterion, stated up front. Verified in front of the client.
I stated the full criterion at the start, as the very first thing: Block 1 closes when a freshly opened agent, one that wasn’t there while we worked, answers two operational questions by going to the right files, in front of the client, and says where it got the answers. Nobody tells it where to look.
Then it was time to put it to the test. New session, zero context from previous sessions, the client’s operational question written right there on the spot.
And that’s how it went. The agent went straight to the right files (the saved decisions, the project summary and the README) and answered citing the source. Right before the client’s eyes, not on my say-so.
Then the part that really sells it: the same folder, a model from another vendor. Pointed at the same system through the AGENTS.md convention, which refers back to CLAUDE.md, the second model made the same recovery. Portability demonstrated, not just claimed. The system doesn’t die when you switch tools, and the client saw it rather than hearing it from me.
The full criterion is two questions. In this first session we tested and passed one: a fresh session plus a replication on a second model. The second one closes at the start of the next block, on the task the client took home. One out of two, written as it is, without inflating it. If I didn’t write it this way, the criterion would be worth nothing, and the criterion is the product.
The system is theirs. They run it.
I didn’t sell them my CLAUDE.md or my scripts: they’re hardwired to my own setup. We built theirs, which they understand line by line. That’s also why tomorrow they can migrate it wherever they want.
The system, in their own files
Written in their repo, versioned with git, portable to any model. No vendor lock-in.
The method for delegating
The three-part structure [YOU ARE] [TASK] [ACCEPTANCE CRITERION], which they reuse on their own for every task.
The prompt template
The prompts already written, saved in the repo, ready to run again without starting from scratch every time.
The session record, with assignments
Within 24 hours: what they did, the outcome of the criterion, what to prepare for the next block.
The same method, on your real project.
Block 1 is planned for 90 minutes. In the first pilot it took about 108. Your hands stay on the keyboard and you verify the criterion yourself. From €249. The session runs in English or Italian, remotely or on site.