A your AI Code plugin that takes a change from idea to merged PR on autopilot, scaling rigor to risk: a tiered process, tests that must fail first, cross-model adversarial review (your AI plus Codex) looped to convergence, and knowledge that compounds across runs.
Skills 12 in the Claude listing
brainstorm Entry point for new work. branch Create an isolated git worktree for feature work — clean separation between feature development and main branch. close Structured session closing — commit check, persist learnings to knowledge file, promote pending entries, check loop staleness, branch cleanup. Codex review Direct Codex adversarial review on the current diff, a specific file, or a commit. execute Per-task plan implementation with built-in verification (lint, format, tests) after each task, knowledge gotcha checking, and optional phase-boundary Codex review. plan Write an implementation plan from a spec — exhaustive task decomposition with dependency graph, acceptance criteria, and spec-coverage verification. Prune stale Sweep merged/stale branches, orphan worktrees, and old JSONL logs. Public pr Pre-open gate + cross-fork mechanics for opening a PR to a PUBLIC or not-solely-owned repo (a fork → upstream, an OSS contribution, any world-visible repo). Review doc Multi-round adversarial review of a spec or a plan — research injection, two reviewers with distinct lenses (Claude + Codex, or Claude alone), a verifier pass, and knowledge-file principle checking. ship Classify the risk tier, verify, prove new tests can fail, review the diff to convergence, open the PR, merge on autopilot, then extract learnings and file follow-up loops. Snap compact Extract learnings from the session before context compaction — preserves knowledge that would otherwise be lost when the context window is compressed. triage Two-reviewer loop triage — independently assess each open loop, reconcile verdicts, batch operator confirmation.