CRAFT for Claude Cowork · Lessons That Compound
Most of what a project learns, it learns by getting something wrong. This post is about the machinery that decides whether that lesson is worth anything on Thursday — or has to be bought again.
Start with the shape of the problem. A fresh AI session begins fresh — that is an architectural choice, not a flaw, and there are good reasons for it. But it has a consequence worth naming: unless something outside the conversation is keeping score, a mistake stays available to be made again. You correct a wrong assumption on Monday. Whether anything on Thursday has a record of Monday is not something you control — and “probably” is a poor foundation for a process. The same correction, the same explanation, the same stretch of the morning — a second time, and a fifth. CRAFT’s general answer to that — keeping your project’s state in files that outlive the conversation — has been written up already. The machinery below is narrower: what specifically happens to a correction, so that it holds.
Link one · Recorded
A lesson is a rule, not a note
When something goes wrong in a CRAFT project — a bad assumption, a step done in the wrong order, an output that passed review and shouldn’t have — the fix is only half the response. The other half is writing the reason down, in the project’s own lessons file, in a form a future session can actually act on.
That form matters more than it sounds. “Be careful with headings” is a note; nobody can execute it. A lesson in a CRAFT project names what happened and states what to do about it. This project’s file goes one step further and scores each entry for confidence — how strongly the project stands behind the rule. Here is one, condensed for publication:
Lessons file — one entry
LL-089 · CONTENT/PROCESS
WHAT HAPPENED: A piece’s headline was replaced late, after the body was finished. A closing line still paid off the old phrasing — a callback to a setup the reader would never see.
RULE: After any title, thesis, or section-heading change, search the full piece for the replaced wording and every echo of it. Now a standing check on every review pass.
CONFIDENCE: 88 out of 100
One entry from a real project’s lessons file, condensed for publication. The entry number, the category, and the confidence score are as recorded.
Read it as a small piece of machinery rather than a diary entry. The category files it where later work will trip over it. What happened is a pattern a later session can recognize. The rule is an instruction it can follow without knowing the story. And the confidence score tells it how much weight to give the whole thing — recorded as a number, so a later session can weigh the rule instead of guessing at how firmly it was held.
Entries like that one get drafted at the close of the session that hit the problem, while the details are still in hand — what went wrong, the rule it suggests, the confidence. What it cannot do is put the entry in the file on its own. Every proposed entry is shown to you first, one at a time, to approve, edit, or throw out; only what you approve gets written. The session supplies the account. You decide what the project keeps.
A rule in the file is the difference between a project that has been burned and a project that knows not to touch the pan.
Link two · Loaded
A rule nobody reads is still just a diary
Writing it down is necessary and nowhere near sufficient — plenty of teams keep a document exactly like this, and nobody opens it. What changes the economics is when it gets read.
In a CRAFT project the lessons file is not reference material you consult when you remember to. It is part of what a session reads on the way in, alongside the rest of the project’s records — the same move that lets a session pick a project back up where you stopped, reading the project’s own memory file on the way in, is what pulls the rules in, because that memory file points at the lessons file — the project’s ledger of corrections.
So they arrive before the work does: a session beginning late in a long project starts with the recent lessons already in front of it — not a summary someone wrote about them but the entries themselves, with any that have gone stale flagged as such.
The other end of the loop is the close. Lessons don’t accumulate by good intentions; they get captured when a session is ended properly — the close is where what was learned gets written down, along with everything else the next session will need. Skip the close and the day’s work mostly survives, but the day’s learning is the first thing to evaporate.
Link three · Checked
And a rule nobody enforces is a suggestion
Which leaves the last gap. A rule that is written down and loaded can still go unapplied — and a rule an AI has read is not thereby a rule anyone has checked the work against.
So the rules get checked, not just carried. CRAFT ships a verification layer — work whose entire job is inspecting other work, with fact-check gates and confidence scoring — and part of what that layer checks is whether what was just produced contradicts a lesson the project has already recorded. A recorded rule becomes something a review pass scores you against — a claim that contradicts one loses confidence, and a claim that loses enough of it comes back marked unverified rather than passing quietly.
Look again at the entry above — the rule line does not stop at “search the full piece.” It ends by saying that search is now a standing check on every review pass. That clause is the whole difference between a lesson and a rule.
What actually compounds
Put the three links together — recorded as a rule, loaded on the way in, checked by the review — and you get the property this post is named for. The aim is that each mistake is paid for once. The payment converts into a standing rule. The rule is there for the sessions that follow, whether or not anyone remembers the incident that produced it.
The practical effect shows up over months: a project where the same failure gets rarer the longer it runs. Early sessions are the expensive ones — that is where the discoveries happen. Late sessions arrive to a shelf of closed doors. The work doesn’t get easier because the work got simpler; it gets easier because a whole category of failure has already been retired, one entry at a time.
Schematic, not measured data. It shows one project over time: a mistake is paid for once, and the rule it produced is in the ledger from then on. The number of sessions is arbitrary, and nothing here counts mistakes — new ones, like the second marker, stay available.
Be clear about what that does and doesn’t buy. It does not make a project mistake-proof, and nothing here claims it does — novel mistakes remain fully available, and some of them will be good ones. What a lessons ledger removes is the repeat: the failure you already paid for, arriving again in a session that never heard about it. That is a smaller promise than perfection, and a narrower one. It is also, as the case for building this way argues, the kind of thing that comes from keeping your project’s knowledge in plain files you own — a rule in a file can be loaded, quoted, and checked against.
What CRAFT asks of you
Two things, both small, both real. The first is closing sessions rather than abandoning them — the ledger is written at the end, so a session you walk away from teaches the project nothing. The second is a habit more than a task: reading the proposed entries and being willing to keep the unflattering version of what went wrong. Nothing is recorded without your approval, which also means nothing is recorded if you wave the whole list through. A lesson phrased to protect someone’s pride is a lesson no future session can use. The file is yours, and it works in proportion to how plainly it is written.
Start paying for mistakes once
The download includes the whole loop this post described: the lessons file, the session close that writes to it, and the verification pass that reads it. Run one project for a few weeks and the ledger stops being an idea in a blog post and becomes a list of doors you have already closed.
No account needed. Free during beta. The download is ungated. One honest prerequisite: CRAFT runs on Claude Cowork, so you need Cowork available in your Claude plan — CRAFT adds no cost, account, or gate on top of it.
The honest bounds. CRAFT for Claude Cowork is in beta, and we say so plainly. A lessons ledger reduces repeated mistakes; it does not prevent new ones, and it is only as good as what gets written into it — a rule nobody records is a rule nobody follows.
And yes, the file grows: nothing is ever removed and entries are never renumbered, so a rule you relied on last month is still there to be checked. What doesn’t grow without limit is what a session has to carry — sessions read the recent end of the ledger, newest first, with anything aging flagged rather than silently trusted. That is also the answer to a rule that was right once and isn’t anymore: it surfaces the question instead of deciding for you, and because this is a plain text file in your own folder, you can retire the entry yourself when it has stopped earning its keep.
Claude’s own memory features work at a different layer and are worth keeping; this is project-level record-keeping in files you own, not a replacement for anything. AI can make mistakes — review anything consequential yourself, ledger or no ledger. This post reflects the framework as it is today and gets updated as CRAFT evolves.
