A kickoff protocol (agent skill) for multi-step tasks. Before writing any code, the agent must: run a spec interview, batch every human-authorization step into one checklist, build a self-verification loop (log sink + verify script + dry-run) *before* the feature itself, and build comparable variants instead of blocking on third-party decisions.
List your recurring tasks; get a keep-as-agent / skill / script verdict and read-only .claude/agents drafts.
Tell it your commands and gotchas, get a working AGENTS.md or CLAUDE.md back in the first reply. It captures fragile build steps, do-not-touch files, past mistakes and testing gotchas, marks gaps as TODOs and asks only what is missing. Ready for Claude Code, Cursor, Copilot and Codex.
Michelin-kitchen audit for your AI coding agent setup: finds where you are still the bottleneck, places you on the trust ladder, and hands you the one upgrade that lets you run more agents with less supervision.
This Skill is built on practical expertise in AI agent safety, prompt-injection defense, source-to-sink privacy review, tool and MCP metadata integrity, memory provenance checks, autonomous action scoping, and evidence-based verification. It helps agents separate instructions from evidence before tool calls, uploads, publishing, deletion, code execution, or network output.
Paste your CLAUDE.md or AGENTS.md and tell it what your agent actually did wrong. It names the rule that failed, quotes the line that should have caught it or says plainly that nothing in your file covers it, and hands you replacement text you can paste back. Built from logged incidents on a production pipeline run with a coding agent, not from general advice. With no incident it still reports what your file itself shows. It makes failures visible; it does not promise they stop.
Audit the memory and instruction files your AI agent reads (AGENTS.md, CLAUDE.md, LEARNINGS.md, system prompts). A bundled offline Python script finds pasted secrets, injected instructions, hidden characters, risky commands, contradictions, duplicates, stale entries and broken file references, and writes cleaned copies. Download package: you bring your own AI agent and Python 3; no model is included.
Before you install an MCP server or agent skill, find out what it reads, runs, and sends. Share a GitHub URL or upload repository files to get an evidence-backed install decision, code citations, documentation mismatches, and targeted fixes. Also reviews coding-agent plugins. Static review; no code execution.
Prevents AI agents from declaring tasks complete until every requirement is verified with a one-line evidence anchor — when a gate fails, the agent stops and reports rather than proceeding, and only a human can unlock the gate to continue.