claude-company-of-one

Reproduce → hypothesize → validate one-at-a-time → fix via TDD → regression test → summarize. No shotgun debugging.

Execute a feature end-to-end: predict Human-Owned Core → TDD → verify → update docs → reflect. Gated by spec-conformance-check.sh pre-execute.

Three-layer review: spec-conformance (Layer 1) → reviewer agent red-team (Layer 2) → critique-dialogue (Layer 3). Fresh-context at Layer 2.

Turn a topic into a written spec (REQUIREMENTS + DESIGN + TODO) and, if warranted, an ADR. Never auto-escalates to the reviewer agent.

Fresh-context Layer 2 adversarial reviewer. Reads ONLY REVIEW_INPUT.md + explicitly referenced artifacts. Produces red-team three-bucket output.

Author an Architecture Decision Record under docs/projects/<project>/adr/. Only when a decision has cross-feature, immutable, long-lived consequences.

Reverse-question the user until the feature intent is unambiguous. Drives brief Open Questions to empty before downstream skills run.

/review Layer 3: Main Agent responds to every Layer 2 finding with a disposition. Hard cap 2 needs-user-decision per /review run.

/debug step: enumerate 3–5 plausible root causes with explicit uncertainty labels. No padding.

/debug step: test one hypothesis at a time. Hard rule: no shotgun debugging — no concurrent changes across hypotheses.

3 hooks across 3 events