ArcadeAI 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="$PROJECT_DIR" bun "$PROJECT_DIR/.safeword/hooks/write-review-stamp.ts" spec
If no [skill-invocation-log] ... ✓ line appears above, run this fallback before stopping:
PROJECT_DIR="${CLAUDE_PROJECT_DIR:-$(git rev-parse --show-toplevel 2> /dev/null || pwd)}"
CLAUDE_PROJECT_DIR="$PROJECT_DIR" bun "$PROJECT_DIR/.safeword/hooks/write-review-stamp.ts" 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 .safeword/hooks/resolve-project-knowledge.ts 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_scopeline — 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: 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 /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:
bun .safeword/hooks/write-review-stamp.ts 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.