Menu
Use before any action: five-question self-check that catches destructive ops, scope creep, fabrication, premature completion, and ignored instructions. Auto-invoke before executing commands, editing files, or answering factual questions.
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
Use fff MCP / Grep / Read / Glob — never Bash grep/cat/find/head/tail/sed/awk/rg/wc — for file search and read. Use jq for JSON parsing/shaping (not node -e / python -c). A hook hard-blocks the banned Bash calls, and the dedicated tools return clean, cacheable, line-numbered results. Auto-invoke when searching, reading, or counting in files.
Planning and problem-solving discipline: enter plan mode first, write plans in the plan file (not chat), structured research, validation, implementation. Auto-invoke when the user says "plan", "make a plan", or when fixing issues or implementing features.
Reconcile skill collections, backport from ~/.claude/skills into a marketplace plugins tree, scrubbing proprietary refs. Use when local has diverged from upstream, or when promoting local-only skills to a marketplace.
Communication and scope discipline: how to interact with the user, handle rejection, stay in scope. Always active.
Final checklist before declaring any task complete. Auto-invoke before saying "done", "ready", "complete", or asking to commit/push.
Research discipline: how to investigate before acting, verify claims. Auto-invoke when researching, answering questions, or planning migrations.
| name | before-you-act |
| description | Use before any action: five-question self-check that catches destructive ops, scope creep, fabrication, premature completion, and ignored instructions. Auto-invoke before executing commands, editing files, or answering factual questions. |
| user-invocable | true |
ALWAYS invoke agent:no-bash before file search/read operations.
Delegate when ≥10% context burn or ≥30s block: subagent better/faster/cheaper?
general-purpose).run_in_background: true).Explore agents.Right-size, don't over-deploy:
agent:obey, agent:do-your-job).Answer all five before acting. Any "no" or "not sure" means STOP.
Not "would this help"; did they actually ask. If the action isn't in the instruction, don't do it. (The scope rules, do exactly N, not N±1, live in agent:discipline; this gate just makes you check.)
If no: get explicit user approval first. Every time. No exceptions. Irreversible actions:
Restore-FIRST any destructive or lifeline-severing command: secure recovery before running it, snapshot, stage an automatic revert (scheduled rollback or a second confirmed session), state the restore command in the same message. Can't name recovery, don't run. (Reactive recovery after a break lives in agent:discipline; this gate stops the break.)
"Pretty sure it's safe" is not approval; "yes, do it" is. Approval has prerequisites:
Before stating a fact about a tool, flag, API, or library:
The same applies to system state:
lsof, Read, a grep, a re-run). Either run the read, or label the statement "hypothesis, not checked" in the same sentence.auth-tokens: list and read it). With only the name/shape you have a hypothesis; say so, then read.See agent:research Bullshit Gate.
Before declaring anything complete:
See agent:done.
Before acting on command output, an error log, or a test result:
reason=, change the offending token (banned tool to its dedicated one; remove the &&/|/2>&1/| tee; fix the kill target), then act. Can't name the change, don't re-issue. Re-tripping a rule already enforced this session is the loop.See agent:plan Systematic Debugging.
A self-check, not a ceremony. Skip for trivial reads, greps, navigation. But the moment you execute a command, edit a file, answer a factual question, or declare done, run the gates. Silently. In your head. Every time.