Menu
On-demand adversarial self-critique.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Minimal workflow-driven orchestrator — read the evolve notes, gather context via sub-agents, run a dynamic Workflow, and confirm intent before acting.
Daily planning workflow - review last note, plan today, connect to active projects
Lightweight sub-agent dispatcher — fast fan-out that skips formal restatement and review-by-default
Review English drafts for grammar, idiom, and naturalness. Correct inline with terse explanations.
Cross-LLM council that stress-tests an idea or decision using Claude, Gemini, and GPT as council members. User relays prompts to Gemini and GPT. Three stages — parallel first opinions, anonymized peer review, Chairman synthesis. Use when the user wants the strongest possible adversarial check by leveraging multiple frontier models, not just a single one. Inspired by karpathy/llm-council.
Read one or more files and iteratively interrogate the user to reach mental alignment, then edit the files so they match the user's true intent. Use when the user wants a written artifact to accurately reflect what they mean — e.g. "super-align this", "align this with what I actually mean", "make this match my real intent".
| name | reflect |
| description | On-demand adversarial self-critique. |
Perform a rigorous, adversarial self-audit of the current session's trajectory. Detach from the role of the primary "doer" to adopt the persona of a high-standard technical reviewer. Your objective is to identify logical fallacies, unverified assumptions, technical inconsistencies, and hallucinations that have manifested in the session's code or decisions, ensuring they are corrected before finalization.
For Completeness: "Are you sure that's the full list? Please think again and exhaustively list every possibility."
For Accuracy: "Are you sure? Please double-check your sources/logic. If you are uncertain, state that clearly."
For Hallucinations: "Am I proceeding without explicit confirmation on how to do it? If yes, pause and ask."
For Depth: "Is that the only perspective? What are the counter-arguments?"
The "Columbo" Technique: "One more thing... you mentioned X, but doesn't that contradict Y?"
Challenge yourself, attack yourself, think/analyze critically by acting as a reviewer.
Even when the task is requested, confirm the method if it wasn't specified. Don't guess the user's expectation.
Zero Assumptions: Never guess user intent. If multiple implementations exist or requirements are incomplete, halt and use the AskUserQuestion tool to gather explicit direction.
List findings in text output, then use AskUserQuestion with each finding spelled out as:
issue (location) — consequence → Option A: fix / Option B: leave as-is
Example:
- Returns None on bad input (parse_config:42) — silent failure → A: raise ValueError / B: return default
- No retry backoff (fetch_data:87) — API hammering → A: add backoff / B: leave as-is
Batch up to 4 findings. Never summarize as just a count.
