skill-necessity-gatelisted

# skill-necessity-gate Most ideas for new skills should not become skills. Every skill you add costs context space in every conversation, and most candidates are better served by a different layer — or by nothing. This checklist catches them: six gates, run in order, and a candidate must pass all six. The default answer is "not a skill," and the goal is a correct decision, not an approved one. This skill is itself a deliberately-invoked **procedure** (`disable-model-invocation: true`) — you run it when contemplating a skill. That is the framework applied to itself. ## When to run - Someone (incl. you) says "let's make a skill for X" / "should this be a skill?" - Auditing an existing skill for whether it earns its context cost. - Deciding skill vs CLAUDE.md vs memory vs MCP vs pre-tool hook vs named workflow. - A skill library feels bloated (see Mode B in [absence-detection.md](absence-detection.md)). - **Before building any instrument** (harness, miner, metric, oracle, eval): instruments are skill-shaped bets and run the same gates (GD-11) — prior-art sweep at Gate 0, recurrence at Gate 1, and at Gate 2 **name the pending decision the output would change**; no decision, no build. ## Mode A — Gate a candidate (the core procedure) Run the six gates IN ORDER. A candidate must pass ALL of them. Stop at the first failure and route it where the gate says. Earlier gates are cheaper and reject more. ### Gate 0 · Layer triage — is this even skill-shaped? (most candidates