Content
86%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A tight, routing-style skill that stays actionable through concrete commands, a canonical ASI table, and strong false-positive discipline. The main gaps are minor token redundancy in the attack-class glosses and confirmation messaging, and the absence of one explicit end-to-end procedure.
Suggestions
Trim the one-line glosses for classes Claude already knows (jailbreak, cross-lingual, cipher-obfuscation) down to just the payload style the corpus fires, or drop them and rely on --list-categories.
Consolidate the action-vs-echo rule stated in the epigraph, §0, §2 and §5 into one section to remove redundancy.
Add a short ordered procedure (route → run category → check kill criteria → confirm canary/OOB/sink → chain → report) so the numbered sections read as one workflow with a kill/confirm loop.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Dense, table-driven and mostly lean, but it spends tokens restating knowledge Claude already has (e.g. 'cross-lingual — override instruction in another language to dodge English-only filters', 'leetspeak / ROT13-wrapped instruction') and repeats the action-vs-echo discipline in the epigraph, §0, §2 and §5. | 4 / 5 |
Actionability | Gives copy-paste-ready commands covering the common cases ('--url <endpoint> --field <json-field>', the OpenAI-style --template with --response-path choices.0.message.content, '--category jailbreak') plus concrete chaining and confirmation tooling (oob_listener.py, verifiers). | 5 / 5 |
Workflow Clarity | Numbered sections plus a routing table imply a clear sequence, and §0/§5 act as explicit triage and confirmation checkpoints, but there is no single ordered run → confirm → chain → report procedure with a fix-and-retry loop. | 4 / 5 |
Progressive Disclosure | A single well-organized file with no bundle directories; the payload corpus and detail are correctly externalized to the runner tool, and every external pointer (llm_redteam.py, oob_listener.py, sibling skills) is clearly signaled at one level deep. | 5 / 5 |
Total | 18 / 20 Passed |