Your first run is never the recipe’s first run

Worked, or only written?

A recipe, in CRAFT for Claude Cowork, is a written, reusable procedure — the steps, the checks, and what “done” looks like, saved as a file you can read before you run it. Here is the shape of one, so you know what you’d actually be reading:

From the outside, every method layer sounds alike: structure, discipline, repeatability. The question that actually matters is one a feature list can’t answer: has this way of working been used — repeatedly, recently, on work that had to ship — or does it just look organized? It’s a question few feature lists even try to answer. Here, the answer is a place you can visit.

A track record you can go read

The discipline behind CRAFT for Cowork doesn’t only describe itself; it runs a publication. Ketelsen.ai — the founder’s prompt-craft site — is updated every Monday, and its own pages name the engine behind it: “It’s called CRAFT … and it’s the engine underneath everything you read on this site.” Each weekly edition — prompt sets written in tiers from beginner to advanced — comes off a structured recipe run, and the archive is public: the framework’s own site describes a builder with “100+ weeks building AI methods in public,” and the archive behind that line now holds over 1,000 published prompts. Those counts are snapshots of a growing archive, so they’ll have moved by the time you look — the point is that you can look. The full write-up of how CRAFT builds Ketelsen.ai is published there too.

The receipts at a glance — counts are a snapshot of a growing archive, so they’ll have moved by the time you look.

To be clear about what that archive is: it’s the founder’s own site, and it isn’t offered as a testimonial. It’s something rarer — the builder using his own method in public, on a schedule, week after week, where anyone can watch it hold up or slip. A dated trail like that isn’t the kind of thing you assemble retroactively.

And prompt posts aren’t the only kind of work on the trail. The framework’s own build ran as tracked project work, and the about page counts it plainly: “Every session — 119 of them across the original project, hundreds more across the Beta — is reconstructible. The audit trail is the product.” Different kind of work — long, multi-session, file-heavy — same discipline, same paper.

Publishing a framework takes a weekend. A dated, public archive of the method doing real work only accumulates one week at a time, in front of anyone who cared to look — and that’s the asset behind CRAFT for Cowork that no feature list captures. There’s no homework in any of this, either: the links are simply there when you want them. The origin story is written down. The archive is browsable. The about page traces who builds it and why.

Your first run is the recipe’s first run at your problem. It isn’t the method’s first day at work.

What that buys your first run

Back to the quarter-end report. You open the recipe, read the steps, and run it. What you’re relying on in that moment is not that this recipe has seen your report before — it hasn’t. You’re relying on the shape underneath it: written steps, defined checks, a named output — a shape that has been exercised in public, week after week, on work with a publish date attached.

Without a written procedure, the report still gets done — Claude Cowork is doing the work either way. What changes is what you spend: you re-decide the steps every quarter, and next quarter you re-decide them again. The recipe’s contribution is that the deciding is already done, in a shape that’s been worked before. Your first run is the recipe’s first run at your problem. It isn’t the method’s first day at work — and that’s why it starts with a read-through, not a leap of faith.

If that sounds like a well-written prompt you could keep in a note — the difference is the parts a note doesn’t carry. A note holds the asking. A recipe also holds the checking: it isn’t done until its gate passes — the totals reconcile, or the exec note doesn’t get written.

None of this claims an inch beyond what Claude Cowork itself provides. Claude Cowork does the work — the reading, the writing, the running of your project. The track record belongs to the discipline layered on top, and that layer is what you’re evaluating when you evaluate CRAFT.

What it asks of you

A few minutes of reading before your first run. A recipe is a file, and the point of a file is that you can read it — so read the one you’re about to trust, the way you’d skim a checklist before using it. And you can watch a first run before you commit to anything: the beginner series walks one end to end — there’s “nothing to install or open — you ask for it by name, in plain language” — and its next page covers what comes in the box and how to find any recipe. Past that, CRAFT for Claude Cowork stays what it is: a methodology layer for Claude Cowork. The platform does the work; the working discipline is what you’re adding.