ai-engineering: Skill for Claude Code

Build the task the plan already approved, and stop where the plan stops

What it produces

Code and tests for one task, and one commit nobody has reviewed yet — HEAD.

Steps

1. Take the task, not the plan: ai-eng spec show <id> --task <n> prints it with the file, the check, the rollback and the digest of the plan it came from. It refuses when that digest no longer matches the file on disk, which is the approval saying so. If the task is not in a plan, or the plan is not approved, stop here: nothing to execute.
2. Write the failing test first, and run it. A test that has never been red is a test that has never been shown to test anything, and the exact focal point matters — assert on the thing that would break, not on the shape around it.
3. Make it green with the smallest change that works. Then refactor with the test still green, or leave it.
4. One logical change per commit. A change that has outgrown the task it belongs to is two changes; split it and say so, rather than widening the commit around it.
5. Run the task's own check and the test files the task names, by path. Not the whole gate: that runs once at block close, over every commit in the block at once, and at close it runs as just check-all — one red never relaunches a full pass; a new failure joins the others, the batch is repaired together, and the one pass lists every red step. --no-verify is denied by hooks/no_verify_guard.py on the way past — if that hook file is missing, stop: a repository whose hooks are not installed cannot attest what this step claims.
6. Label the hand-off UNREVIEWED and say what happened — the parts that did not, too. Say it in the commit message and in the hand-off, and never by editing the plan. The approved bytes are what the approval was given for, so a task that rewrites them to record itself has withdrawn the permission it is running under. That is not a formatting preference; it is the loop this step exists to break.
7. When the plan stops being true — the task is bigger than written, the design does not survive contact, a decision is missing — record the stop before you stop: ai-eng report blocked --what "<the gate>" --why "<what is missing>" --action "<the literal that clears it>", then regenerate the page with ai-eng report intent --html. A halt nobody can see is a halt nobody acts on, and the person is not at the keyboard. Say what is missing; never that it arrived. Then escalate to /ai-spec. Continuing on a plan you have privately rewritten is the failure this skill exists to prevent.