It knows a lot, but starts off knowing next to nothing about you. It can't look up what you decided last month, today's correction may not make it to tomorrow, and it waits for you to start everything. A harness is what you put around it, so it works like a colleague who's been with you for months.
A harness is half files, half habits. These four are what make the files keep paying off.
Memory and the journal only fill up if each session writes to them.
That's the raw material the learning loop turns into permanent fixes.
Read the memories and rules Claude proposes, so your harness describes you, not Claude's guesses about you.
Everything always loaded costs something every session, so remove what hasn't earned its place.
The kit's guide goes deeper on each, including its tradeoff.
Some tools earn their place later. Setup writes each one into your instructions with the moment it's for, so Claude can suggest it then.
You're facing a decision with real tradeoffs
/decide a decision partner that lays out the real tradeoffs
You're about to commit to something that's expensive to undo
/x-ray a check on what you're both assuming and not asking
You're not sure you're solving the right problem
/reframe names what the work assumes and offers other ways to see it
You're starting a new project
a new-project checklist what a project needs on day one
You're about to build something with real scope
/prd questions the idea before any code gets written
You're building the first thing real people will click through
/ux-thinking checks it will make sense before it's built
You're correcting the look of the same thing twice in a row
/options puts real variants side by side to react to
You're building your first web page
layout rules three rules that load only when Claude works on web files
You're about to show work to other people
/squint a cold read, the way your audience will meet it
You're working on a codebase that really matters
/codehealth reviews for the mistakes AI-written code tends to make
You're about to push code or deploy for the first time
a push gate nothing goes out without your yes, every time
You're coming back to a project after days away
/pickup rebuilds where you left off and what was promised
You're running two sessions at the same time
an append-only guard one session can't overwrite what the other wrote
Two weeks in, at your setup check-in
a session-start check says where you left off, and proves your guards still work
A set of plain text files your Claude reads and works through with you. It builds into your own Claude Code setup in one sitting, finished by a check-in two weeks later, and never overwrites anything you already have without showing you first. It writes to Claude Code's settings folder in your home directory (~/.claude), so what it builds works in every project, and you can start it from any folder.
This grew out of one person's way of working. Some of it will fit you and some won't. Keep what helps, skip what doesn't, and expect your harness to take its own shape.
Download the kit (zip)This harness can also work with Codex. The kit includes an adapter for it.
Read START-HERE.md and SEED.md in <the path to harness-kit-main>, then let's begin.