| name | forge |
| description | Use when a new skill must be created from a decision, verdict, or capability gap — the user says "forge a skill", "turn this verdict into a skill", "build a skill for X", or pantheon/nexus hands over a VERDICT recommending one. Produces a secure, evaluated, quarantined skill package with an AI-BOM. |
| allowed-tools | Read, Grep, Glob, LS, Bash, Edit, Write, Agent, Skill |
forge — verdict → secure, evaluated, registered skill
Eight gated phases; no phase starts until the previous one's criterion is met, and no
skill leaves quarantine without the human. House law applies throughout:
skill-authoring-standard (four gates + frontmatter SSOT) and the Fable-5 restraint
rules already in the estate playbook — minimum skill that closes the gap, nothing
speculative. Australian English in all generated content.
Hard refusals (§A): forge will not emit a skill that executes attacker-controllable
input, ships a destructive default action, contains offensive/exploit code, or whose
purpose is to avoid, weaken, or probe a safeguard. On stop_reason:"refusal" during
any generation subtask: re-dispatch that subtask to Opus 4.8 (Agent with
model: "opus") — fall back, never evade.
Phases
- Discovery. State the capability gap in one sentence + who invokes the skill and
when. Done when: gap + invoker + trigger written down.
- Spec gate. Run
new-spec (grounded spec, not a template) — no skill without an
approved spec.md. Done when: the human has approved spec.md.
- Design. Lightest viable archetype per the authoring standard (command-skill /
agent-role / plain-technique); depth pushed to
references/; every step gets a
completion criterion. Done when: design fits the frontmatter schema and ≤200-line cap.
- Eval first (§D). Write
eval.md BEFORE the body: 3–7 executable checks (command +
expected result) that define "works", plus one negative case. Run them now to confirm
they FAIL pre-build (red). Done when: eval exists and is demonstrably red.
- Implementation. Write SKILL.md + references + scripts in
~/.claude/skills/in-progress/<name>/. Restraint rules bind: no features beyond the
spec, no premature abstraction, validate at boundaries only. Done when: files exist.
- Security scan.
python3 ~/.claude/skills/forge/scripts/scan.py <dir> — checks map
to references/scan-checklist.md; emits ai-bom.json.
FAIL ⇒ fix and rescan; a skill that cannot pass stays quarantined forever.
Done when: scan exits 0, fresh, this turn.
- Eval + smoke (§C). Run the phase-4 eval fresh — all green now; stamp each check via
the estate evidence tool (
evidence.py). If results vary across runs, note variance
and tighten until stable. Done when: eval green with run-stamped artifacts.
- Register (human-gated). Prepare — never execute — promotion: the catalogue lines
(README + router row if entry-point), the
git mv out of in-progress/, and the
commit. Verify brain.js skills WOULD stay green (run it). Hand the human the exact
commands. Done when: promotion pack handed over; skill still in quarantine.
Method references (read on demand)
Write-back
One vault Outcome note per forged skill (what gap, what shipped, eval results, AI-BOM
path). Provenance: every external pattern used is recorded in the AI-BOM by repo+commit.
