adversarial-agentslisted

# Adversarial Agents Configurable adversarial panel review of any artefact. Generalises the panel-of-personas pattern from Matt Pocock's `grill-me` to arbitrary artefact types — plans, code, design docs, prose, model outputs. Pure conversation. No file output. No "recommended answer" — the critique IS the question; the user's pre-commit defense is the user's answer. ## Triage first Before dispatching, check the artefact is concrete enough to attack: | Signal | Action | |---|---| | Artefact is <~100 words, or has no concrete decisions/values/file paths/named targets | **Refuse.** Reply: "This artefact is too thin to attack productively — flesh it out first, then come back." Do not dispatch. | | User says "review my plan" / "grill this plan" with substantive content | Proceed; panel will be `plan` (YAGNI / Premortem / Hidden Assumptions). | | User pastes code / points at a file | Proceed; panel will be `code` (Saboteur / New Hire / Security Auditor). | | Other concrete artefact (design doc, spec, prose, model output) | Proceed; panel auto-selected per the table below. | Adversarial-agents attacks fleshed-out artefacts. It does not flesh them out — that's `brainstorming`'s job for plans, or the user's drafting process for everything else. ## Pre-commitment gate (after triage, before dispatch) After triage passes, ask the user for a 1-paragraph pre-commitment **before** dispatching the panel: > Before I dispatch the adversaries: in one paragraph, state what you think thi