# Self Review

> Runs when safeword's review gate asks for a spec review — an inline self-review of the just-authored spec.md that earns its Tier 1 review stamp. Not a user-facing entry point, and never spawns a sub-agent.

- Skill: `arcadeai/self-review-3` (Agent Skill)
- Install (CLI): `npx skillmds@latest add arcadeai/self-review-3`
- Raw SKILL.md: https://api.skillmd.com/api/skills/arcadeai/self-review-3/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: ArcadeAI (https://skillmd.com/u/arcadeai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/arcadeai/self-review-3

---


# Self-Review

Review the artifact you just authored, then earn its review stamp so the next
step is unblocked. This is **Tier 1** — a fast, built-in first pass: you review
your own work, no sub-agent involved. Tier 2 (the phase-exit review) is the
independent check; that happens separately.

**Stakes set depth.** Tier 2 may never run — review as if your stamp is the last
word before code gets built on this spec, because often it is. Cheap floor means
fast, not shallow.

## Earn the stamp

The line below runs the stamp-earning step at render time. It binds a
`review:<scope>` stamp to the active ticket's `spec.md` **at its current
content** and appends it to `skill-invocations.log` under the project namespace root, where the
per-asset gate reads it back. Invoking this skill is what writes the stamp —
hand-editing the log would let you fake this, a known gap this tier accepts to
stay cheap.

!`PROJECT_DIR="${CLAUDE_PROJECT_DIR:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}" && CLAUDE_PROJECT_DIR="$PROJECT_DIR" bun "${CODEX_HOME:-$HOME/.codex}/plugins/cache/safeword/safeword/1.0.0-rc.3/runtime/cli.js" project runtime write-review-stamp -- spec`

If no `[skill-invocation-log] ... ✓` line appears above, run this fallback before stopping:

```bash
PROJECT_DIR="${CLAUDE_PROJECT_DIR:-$(git rev-parse --show-toplevel 2> /dev/null || pwd)}"
CLAUDE_PROJECT_DIR="$PROJECT_DIR" bun "${CODEX_HOME:-$HOME/.codex}/plugins/cache/safeword/safeword/1.0.0-rc.3/runtime/cli.js" project runtime write-review-stamp -- spec
```

The stamp is tied to the spec's exact current content — edit the spec, and the stamp goes stale automatically.

**If the automatic line and fallback both print `[skill-invocation-log] FAILED`, or still do not print `✓`**: STOP.
The stamp was not written and the gate will keep blocking. Most likely the bash
injection was denied, no in_progress ticket was found, or Bun could not run the
installed helper — report it to the user and resolve before retrying.

## Review the spec (do this now, with the stamp written)

The stamp records that a review was invoked; the actual scrutiny is yours. At
review time, run `bun "${CODEX_HOME:-$HOME/.codex}/plugins/cache/safeword/safeword/1.0.0-rc.3/runtime/cli.js" project review-knowledge --json` and use its
current `principles`, `personas`, and `surfaces` source paths and content—not
labels remembered from intake. These resolve from `paths.principles`,
`paths.personas`, and `paths.surfaces` when configured. Read those sources with the active ticket's
`spec.md` and `scope` / `out_of_scope` frontmatter:

- **Every JTBD resolves to a real persona** and reads as a genuine job (`When
I…, I want…, so I can…`), not a restated feature.
- **Each JTBD carries ≥1 numbered Rule** (or legacy Acceptance Criterion)
  stating an observable, product-level invariant — not an implementation detail.
- **The criteria cover the ticket's scope** and stop at its `out_of_scope` line —
  no silent scope creep, no orphan capability.
- **Every affected surface resolves against the configured surfaces inventory**
  or is explicitly marked spec-local; no invented reusable context and no
  configured surface silently renamed in `spec.md`.
- **Nothing leaks implementation** (file names, function names, libraries) into
  spec-level prose.
- **A Product Plan owner's Killer Demo earns its payoff**, using the
  [intake standard](../bdd/DISCOVERY.md#killer-demo): check the value and its
  visible proof, not just filled fields. A child contribution does not need a
  second demo.

If the review surfaces a fix, **edit `spec.md` and re-invoke `$safeword:self-review`** — the
content-bound stamp goes stale on any edit, so the gate correctly re-blocks
until the corrected spec is re-reviewed. That is the point: a review that
changes the artifact must be re-earned.

## Skip valve

If the artifact is genuinely trivial to review (boilerplate, a docs-only
change), log a skip with a reason instead of a review — it clears the same gate
and records why:

```bash
bun "${CODEX_HOME:-$HOME/.codex}/plugins/cache/safeword/safeword/1.0.0-rc.3/runtime/cli.js" project runtime write-review-stamp -- spec --skip "<why this spec needs no review>"
```

To skip, pass `--skip "<reason>"` as one quoted argument. A reason is required — an empty one won't clear the gate.