Human Only Skills

Human-only skills

Four skills carry disable-model-invocation: true. In runtimes that honor that key, their descriptions stay out of the context window and the model cannot invoke them — a person types the command. In runtimes without the switch, the key is stripped at projection time and these remain ordinary advertised skills (see docs/contracts/codex-skill-api.md); the pointer below still tells a person which ones to run themselves. This skill names skills. It never invokes one.

The roster

What they share: each is off the critical path, each is started deliberately rather than inferred from a task, and nothing else in the repository reaches for any of them.

When a model wants one

Say which skill applies and why, and stop. Reproducing the skill's procedure by hand defeats the point of removing it from model reach — the command exists so a person decides the timing.

Before adding a row

Model invocation is load-bearing whenever anything else reaches for the skill, so a skill joins this roster only after all four surfaces come back clean, with the evidence recorded:

1. ao skills consumers <slug> and ao skills graph — declared dependencies and consumes edges.
2. workflows/*.js — a workflow that dispatches the slug.
3. Other skills' SKILL.md bodies — a "routes to <slug>" sentence or a See Also entry is a reach.
4. evals/routing-probes/templates.json — an applicable entry means a probe measures whether the model routes there, which stripping makes unmeasurable rather than merely cheaper.

Retire this skill when the roster is empty.