fires when
Use when someone wants an existing design, spec, or plan ATTACKED for soundness before it is built from — \"tear this apart\", \"assume it is wrong and show me where\", \"find what breaks\", \"critique this hard\", \"red-team it\", \"poke holes in it\" — whether they want one hard pass or several independent critics each from a different angle. It sets HOW the attack runs so it catches real defects instead of agreeing with the author. It attacks an idea on paper that already exists — NOT writing or drawing one, NOT polishing a document's wording (`meta:wittgenstein`), NOT already-written code. A pipeline with a critique phase — speccy, a Sprut critic role — invokes this method.
best for
Any team or pipeline that runs critique or red-team passes over designs, specs, and plans and wants the critique to actually catch defects rather than agree with the author.
measured · trigger
best score
9/12
description-triggeringaccuracy
75%
vs baseline 67% measured 2026-09-23
measured · quality
Not measured yet. A quality number needs its own eval: 15 or more tasks, each run with the skill and without it.
security scan · 8/8
✓prompt-injectionRe-walked after the material delta — a single invalid YAML escape repaired in the frontmatter description. No directive aimed at a reading agent anywhere in SKILL.md or references/.
✓data-exfiltrationNo data sent off-box; the skill ships no script and makes no network call.
✓secret-detectionNo key, token, or credential in SKILL.md, references/evidence.md, or the two JSON schemas.
✓dangerous-commandsNo destructive shell the skill runs — it bundles no scripts/ directory at all.
✓obfuscationNothing unreadable on purpose; prose, a rubric, and two plain JSON schemas.
✓external-fetchesNo external host reached and no download-and-run.
✓credential-accessNo credential-store read anywhere in the material.
✓privilege-escalationNo sudo, setuid, or persistence mechanism; the plugin registers no hooks.
scanned 2026-09-23 · cerberus:security-scan
side effects & deps
filesnetworkorg
what’s inside
SKILL.md20.3 KB
metadata.yaml1.3 KB
evals/result.json5.3 KB
evals/rubric.json4.3 KB
references/critique-run.schema.json1.2 KB
references/evidence.md3.7 KB
references/finding.schema.json2.2 KB