The thing that clicked today: the chat is not the memory. The notes are the memory, and a conversation is one bounded unit of work that reads them at the start and leaves them ready for the next one at the end.
Nearly everything on this site was built with AI assistants writing most of the code while I set direction, reviewed, and verified. Over two months that produced a specific kind of mess. A conversation that went long got expensive and confused. A conversation that started fresh did not know what the last one had decided. Notes existed, but nobody had said which note was the one you read first.
Built / shipped
The method, adopted by mapping rather than by bootstrapping. The fleet already ran most of what the method needs: deep per-section session notes, a thin dated journal, a project status note, a handoff. The one real gap was a roadmap. The risk was spinning up a parallel system, which the method itself forbids, so the move was to map the roles onto the machinery that existed and fill only the gap.
Three continuity documents per project. A contract: the stable rules and safety boundaries of that project, rarely edited. A roadmap: the open plan, a small registry of locked decisions, a short progress log. A handoff: exactly one current section, its stamps, bounded context, the model tier it wants, watch items, resource posture, and a stop condition. The handoff is self-sufficient on purpose: a fresh conversation reads it and can start without reading anything else.
One conversation, one section. Resolve which project owns the request. Orient from that project's contract and handoff. Diagnose before touching anything. Execute only the named section, keeping a few bounded commands on the fast path and delegating understood multi-step work to the cheapest worker that can do it. Verify the material results at the source, not from a summary. Write the session note in two voices (mine for goal, decisions, and what is next; the worker's for execution and blockers) and a thin journal line. Close: tick the roadmap item, add its progress line, rewrite the handoff to the next self-sufficient section, and leave a kickoff prompt of at most five lines. Stop. New scope goes into the roadmap and a fresh conversation, never "while we're here."
A definition of done that includes continuity. A section is complete only when implementation, verification, the session record, and the handoff rewrite are all done. A worker's response, a local commit, or a passing test without the continuity close is incomplete. That rule is what makes the next conversation cheap.
Section sizing. One focused deliverable plus its verification and logging, with headroom for one test-and-fix cycle. Anything that needs independent approval, touches unrelated repositories, carries more than one irreversible decision, or needs a handoff longer than about forty lines gets split.
Problems & fixes
The first thing the method found was that the method's own document was not indexed yet, so the tool that searches my notes could not find it. Linked at its canonical path and flagged for me to confirm, which is the right shape for a tool telling you its own limitation.
A status note that duplicated the handoff and the roadmap drifted against both within days, and was retired fleet-wide later that month. Two documents that can disagree will.
The orientation itself turned out to be a cost. Reading the roadmap every turn was the largest single token expense in a long conversation; the fix, a few days later, was tiered orientation (handoff every time, reference documents only when their stamps change, roadmap only at replan or close) and archiving closed work the moment it closes.
The session schema got stricter in August: a fixed filename shape, frontmatter that must agree with the filename and the owning project, and six required sections with no placeholders, enforced prospectively. Older notes stay as evidence of how the work was done at the time, and are not rewritten or backfilled with invented voice.
Decisions
Continuity lives in documents, not in chat. A conversation is disposable; the handoff is not.
One section per conversation, and the handoff names exactly one next section.
Locked decisions are a curated registry that points at the journal entries where they were made, not a second ledger.
Done includes the close. Without it, the next conversation pays for the shortcut.
Learned
The expensive failure of long AI work is not a wrong answer, it is a lost thread. A method that makes every conversation start from a small, current document fixes more than any prompt.
Mapping onto what exists beats bootstrapping. The gap was one file, not a system.
Every rule here came from a failure: the retired status note, the per-turn roadmap tax, the notes with missing voice. The method is a list of things that went wrong once.
Still open / next
The method is written down and every project now carries the three documents. Whether every conversation honours the close is visible in the handoffs themselves, and the nightly lint that checks note shape is how drift will show up before it costs anything.