Lessons that should transfer
The rest of this page argues from the record. This page collects the claims, with the evidence and a note on adapting each one, so someone with a different project can use the shape without the specifics. The specifics were a farming fantasy written in ten days by someone comfortable with a terminal. The shape should survive being moved.
Voice is the decision with the most effect on everything else, and it comes first. The project spent its opening hours on register with no story in sight, and the project’s own notes say that register is still the book’s voice today. Evidence: the first session, whose pre-premise stretch is shown in full. To adapt: whatever your medium, spend the first session on how the output should sound or look, using throwaway subject matter, before you say what it is about.
Examples outrank rules, and the model must read the examples first. Drafting from rules alone produced compliant, generic prose, but drafting from exemplars produced the voice. Evidence: the voice memo page and the voice-drift incident on 24 and 25 August. To adapt: keep a small set of outputs you actually like, protect them, and make reading them the first instruction in every session.
Feedback should name what the reader felt, and then become a rule. The dialogue blind tests, readers judging passages without knowing which were AI-written, came from “this feels artificial” and ended as countable rules for each character’s speech. Evidence: 27 August. To adapt: report effects, ask for causes, and end every round by having the general finding written into a file.
Anything not in a file did not happen, and the model keeps the files. The handoff document, the pointer file and the dated rulings exist because of this, and every one of them was written and maintained by the model on David’s instruction rather than by David. Evidence: the handoff page and the fact that the last session of the project started by reading the queue the previous one left. To adapt: have the model keep a short root instructions file and a living next-steps file from the first day, and make updating the second part of ending every session.
The model will fabricate facts about its own earlier work, and only a tool will notice. The citation checker found 206 mismatches between notes and manuscript, and 110 of those were quotations that never existed, in a set of notes that read perfectly well. Evidence: 28 August. To adapt: any claim the model makes about what it previously produced, quotations, summaries, “as established in chapter three”, should be mechanically checkable, and periodically checked.
Check the whole corpus, not just the current piece. Both major discards were caused by properties of the whole that no per-scene check could see: drift across a book and echo across drafts. Evidence: the overlap checker and the 30 August discard. To adapt: build at least one tool that reads everything you have produced and asks about repetition, drift and similarity to things you rejected.
Use many small, cheap checks first, then one larger, careful check. The recurring structure in the project’s checks is a fan-out of small model calls each asking one question, followed by one larger model judging the results. It was cheaper than one big pass, and on the project’s own seeded tests it was not worse at finding problems. Evidence: the linting page. To adapt: never ask one call to find everything. Give each reviewer one question and give one reviewer the job of deciding.
Expect to discard, and make discarding cheap. Two drafts of book one were thrown away. In each case rules had already been written out of what the draft showed. Evidence: the throwing away page. To adapt: version control, exemplars stored apart from drafts, and the habit of extracting the rules a bad draft reveals before deciding its fate.
Keep creative work in one long conversation and delegate the bounded work. Splitting drafting across agents lost coherence, but splitting checks across agents gained speed and fresh eyes. Evidence: the delegation page. To adapt: the test for delegating a task is whether it can be fully described in a brief and its result fully described in a report. Drafting fails that test. Checking passes it.
Loosen supervision gradually. Supervision went from scene by scene to batches to unattended over four days, and the final unattended run had a written stopping rule in the handoff file. Evidence: 31 August and 2 September. To adapt: loosen one notch at a time, on evidence, and never leave the model alone without telling it when to stop.
Ask the model to catch itself, and act on it when it does. The model stopped mid-draft to flag a risk that its own draft had been contaminated by material it shouldn’t have used, and it retracted a set of results when its scoring method proved unreliable. Both followed clear instruction that surfaced problems were wanted. Evidence: the loop page. To adapt: put it in the standing instructions, and when the model stops to tell you something is wrong, treat that as the process working rather than as a delay.
One last thing. It is more a warning than a lesson, and more an impression from reading the record than a measured finding. What was planned up front, the first draft and the first set of voice rules, changed the most. What lasted came out of a specific failure and was written down with the failure attached. If you are starting something like this, plan lightly, read closely, and keep the files.