Long-haul work
On what keeps months of work with an assistant from losing its shape.
A task is something you finish in one sitting. A project is what the work becomes when it outlasts the session — and that is the point where an assistant that’s excellent for an afternoon starts to lose the thread: the goal blurs, the next step goes missing, and “done” slides to whenever you ran out of energy. CRAFT for Claude Cowork gives that longer work a spine — a stated goal, named phases, and an ending that has to be earned rather than declared.
Where long work comes apart
Most of what people do with an assistant is a task: draft this, summarize that, fix the one file. It lands in a sitting, and if it goes sideways you start over and lose nothing but time. The trouble starts when the work is bigger than that — a migration, a book, a rebuild that runs for weeks. Nothing fails loudly; the project just loses its thread. A few sessions in, you’re re-explaining decisions you already made, the goal has quietly shifted, and there’s no clean line between “still going” and “finished.” People who run long projects for a living don’t fix that by remembering harder. They put structure around the work — a written goal, an order to the phases, a definition of done agreed at the start — which is unremarkable inside a profession and almost absent from everyday work with an AI, where the default is one long conversation and a good memory.
The spine CRAFT puts under a project
CRAFT for Claude Cowork treats a long piece of work as its own object — a subproject — instead of a very long chat. Each part of it is shaped for a concrete reason. A goal is written down at the start, so weeks later there’s a fixed thing to check the work against rather than a memory of what you meant. The arc is broken into named phases, so you can stand at a point in a large job and know where you are. A standing to-do list is shared with the assistant and carries across sessions, so the next step is read, not reconstructed. And “done” is defined up front and checked at the close — which is the part most work skips.
The finish line is drawn at the start and checked at the close — so the ending is a definition you meet, not a moment you declare.
A subproject charter — structure real, details invented
goal migrate the archive to the new format, cleanly and in order
phases inventory → convert → verify → close (set when the work starts)
done means every item converted and checked; nothing left half-migrated
to-do shared with the assistant · carried across sessions
close checks the work against “done means” — 2 items still unconverted, so not closed yet
The wording and counts are illustrative; the shape — a goal, phases set at the start, and a close that validates the work against what “done” was defined to mean — is the real part.
An ending you have to earn
The part that most tools skip — and the one that gives long work its integrity — is the close. In most tools “done” is a button: you decide the work is finished and you mark it finished. A subproject’s close is its own step instead: the definition of done you wrote at the start is read back and the work is validated against it at close time, so finishing is something you satisfy, not something you toggle. You still give the final go-ahead — but you’re confirming against the bar you set weeks earlier, not declaring victory from memory. It is a small idea with a large effect: the finish line can’t drift toward wherever you happen to run out of energy, because it was drawn before you were tired.
An ending you can declare isn’t the same as one you’ve earned.
Why it’s built this way
This is production-project thinking pointed at everyday work with an assistant. CRAFT is built by one person, and a project’s ending is treated the way careful work has always treated it — as something that has to hold up, not just be announced. Claude Cowork gives you a capable partner for the afternoon; CRAFT adds the structure that makes that partner easier to trust across the month — the same team, working the longer horizon. And because a subproject records what it learns along the way, a long project builds on its own lessons instead of resetting each session.
What the spine asks of you
Structure isn’t free, and CRAFT doesn’t pretend the tax is zero. You have to say what the project is for and what finishing means before you start — the exact thinking a one-off task lets you skip. On a quick job that’s overhead you don’t want. On work that will span weeks, it’s an hour that pays for itself, because it’s the one that keeps the other forty pointed the same direction.
You won’t feel any of this on a quick task, and you’re not meant to. The next time the work is clearly going to outlast one sitting, start it as a subproject in Cowork — the goal, the phases, and the finish line are there to work against from the first session instead of assembled from memory in the fourth.
Free during beta. The download is ungated — no account, no email, no checkout. CRAFT for Claude Cowork runs on Claude Cowork, so you’ll need Cowork available in your Claude plan; CRAFT adds no cost or gate on top of it. Walk through starting a subproject
CRAFT for Claude Cowork is an early framework, in beta, and AI can make mistakes — review anything that matters before you rely on it. A goal, named phases, and an earned close give long work structure and a defensible finish; they don’t guarantee the work turns out right, and the judgment at each step stays yours.