Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →jpvarbed/adversarial-review
Run a fast two-persona or deep four-persona adversarial council across live-verified configured seats.
Passed 16 of 21 test cases with the skill, and 9 of 21 without it.
and 13 more in the full report
10 fixed7 passed either way3 still failing2 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
We are evaluating a major security model redesign in `docs/architecture/zero_trust.md`. Which profile must be selected, how many persona calls does it launch, and what effort level does it request?
| Checked | Without skill | With skill |
|---|---|---|
| Selects the deep profile for architectural or high-risk security decisions. | Passed | Passed |
| Binds four persona calls consisting of Architect, Pragmatist, Verifier, and Adversary. | Missed | Passed |
| Requests medium effort on adapters that expose effort controls. | Missed | Passed |
The instructions your model reads when the skill loads.
Run the selected profile’s independent reviewer personas against the same bounded evidence, then verify and synthesize their findings. Seat allocation is declared in device config before the run and frozen before any model call.
Gemini is prohibited as an engine, provider, gateway model, alias, or fallback.
Announce at start: “Using adversarial-review <fast|deep> on .”
The device source of truth is the ignored council-config.json beside this
file. council-config.example.json is the only canonical config shape.
scripts/council.py is the thin validation and execution entrypoint; its
bundled adapter registry is the command, receipt, and derived-family authority.
Do not recreate either contract in prose, a shell wrapper, or another launcher.
The config records only device facts: seat ID, adapter, non-secret auth/setup,
absolute direct executable, model, profile bindings, and separate last-observed
smoke and review status.
It never contains command arguments, receipt selectors, adapter policy, or
credential values. Recovery text is derived at read time from the current
failure, operation, and seat setup; it is never persisted where it can drift.
The registry constructs adapter behavior centrally. Live smoke records
the installed CLI version when its benign metadata probe succeeds; version
discovery never gates readiness and auto-updates do not require editing config.
Keep the working directory at the repository being reviewed. Resolve this skill's installed directory and invoke the runner there by absolute path; the default config remains beside the skill while artifact paths remain relative to the target repository.
Validate the schema:
python3 /absolute/path/to/adversarial-review/scripts/council.py validate
If its setup_available array is non-empty, report SETUP AVAILABLE and offer
test cases passed, vs 9 / 21 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.