Build

Single-threaded implementation loop that works through tasks in SPEC.md. You invoke it when you want Claude to actually build what's in the spec, and it plans the work, makes the edits, runs tests, and flips task statuses from open to done. The interesting bit is the backprop hook: when a test fails, it stops to ask whether the failure reveals a spec bug or missing invariant, then updates SPEC.md before retrying. No agent swarms, no parallelism, just a straightforward execute-verify-mark-done cycle. Works best when you already have a populated spec and want disciplined execution that learns from failures instead of silently patching around them.

skills.sh installs: 2.7K
GitHub stars: 1.2K
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add juliusbrussee/cavekit --skill build --agent claude-code

Installs into .claude/skills of the current project.

Files

build — implement spec

Single-thread native plan→execute. You are main Claude. No swarm.

LOAD

Read SPEC.md. If missing → tell user to invoke the spec skill first. Stop.
Read FORMAT.md once if not loaded.
Read §R if present — external facts the build must honor, ⊥ re-derive or contradict.
Parse invocation args:
- §T.n → that task only
- --next → lowest-numbered row with status . or ~
- --all or empty → every . row in §T order

High blast radius (shared module, auth, data, money, public §I)? Run /review first. Trivial & reversible? Skip planning ceremony, just do step EXECUTE.