Review

SkillDocs & knowledge

Red-team existing skills, constitutions, code, schemas, plans, or documents and return verified, severity-ranked change requests. Use when the target already exists and needs adversarial fidelity, correctness, ownership, operability, or abuse-resistance review. Use research instead when the question is whether a new claim or candidate survives.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Review skill

What this skill tells your AI

The instructions your AI receives, as published by cyberalchemyai/arcanum in arcana/review/SKILL.md and read by ahel’s review.

Review does not create research verdicts on new candidates and does not apply its own fixes. Fixes require a subsequent authorized change task.

  • inline: render the complete document in chat and omit working_folder;
  • persisted: write <working_folder>/review.md after the human confirms the destination.

Do not persist attacker transcripts, attacks.md, or findings.md. The proof of a review finding is the cited or quoted target artifact, not an agent return.

  • fidelity/governance: contradictions with governing authority;
  • mechanics/correctness: failures, broken validation, or doc/code mismatch;
  • ownership/reference integrity: dangling pointers or double ownership;
  • operability: steps a fresh operator cannot execute without invention;
  • abuse/gaming: rules satisfiable in letter while defeated in purpose.

Roles:

  • explorer as attacker: reads the whole target corpus through one lens;
  • writer as synthesizer: deduplicates and severity-ranks findings;
  • skeptic as verifier: tries to refute each finding against the target;
  • optional auditor: checks coverage and ensures refuted findings were removed.

No attacker verifies its own finding. When the parent authored the target, a surviving recommendation to revert that work escalates to the human rather than being self-approved.

  • target file or artifact;
  • exact locator and a short quotation or directly observed behavior;
  • severity: CRITICAL, MAJOR, or MINOR;
  • consequence;
  • one bounded proposed fix.

Use CRITICAL for system breakage, corruption, or direct contradiction of governing law; MAJOR for functional gaps, drift risks, or load-bearing omissions; and MINOR for wording, stale metadata, or fuzzy pointers.

An artifact verdict is FIX when at least one CRITICAL or MAJOR finding survives; otherwise it is KEEP. A verified FIX report is a successful review deliverable, not a failed review.

Coverage

| attacker | lens | targets checked | findings raised | zero-findings defense |

Findings

| # | artifact and locator | evidence | severity | consequence | proposed fix |

Artifact verdicts

| artifact | KEEP or FIX | rationale |

Change requests

Evidence boundary

Signals

GitHub stars
25
Forks
3
Last commit
Sep 2026
Hacker News mentions
20
Advanced
Catalog kind
skill
Gateway key
review-cyberalchemyai
Source
github.com/cyberalchemyai/arcanum