Trinity Skills — AI skill for Claude Code

trinity-skills

Three Claude Code skills that formalize the **Claude ⇄ Codex review loop** for planning and code review:

Skill	What it does
create-plan	Scaffold a new plan markdown file in docs/ using a canonical template (phases, iterate-review markers, HISTORICAL stubs) — including the project's risk posture: what ships, what the blast radius is, what the ship bar is, which code is a trust boundary.
iterate-plan	Iterate a plan between Claude (the editor — whichever model drives the session) and Codex (structured reviewer) until APPROVE. Runs an architect and a product-manager lens per pass, each fed the plan sections it needs, and merges their findings into one list. Codex never edits — it returns JSON-schema-validated findings + plan corrections + question answers; the editor folds them in.
iterate-review	Same loop applied to code changes — a working-tree diff, a feature branch, or a PR (open or merged). Deterministically selects senior-dev (always) plus security and QA based on what the diff touches, then merges their findings. The editor folds them into the working code.

They're independent — install any one of them — but compose naturally:

                              human
                                │
                ┌───────────────┼───────────────┐
                ▼               ▼               ▼
         create-plan ──▶ iterate-plan ──▶ iterate-review
         (scaffold)      (design review)   (code review)

Version history, including the breaking changes and the known gaps, is in [`CHANGELOG.md`](CHANGELOG.md).

Why this exists

Sending a plan or diff to Codex for review is high-value but tedious to do by hand: copy the markdown out, paste into a Codex prompt, copy the response back, manually decide what to incorporate, repeat. The trinity skills automate the shuttle while preserving the load-bearing invariants:

• Claude is the sole editor. Codex