Write Skills

Write Skills

A skill is not documentation. It is compressed operational memory for an agent that already knows how to code and reason. Its job is predictability — the agent taking the same process every run, not producing the same output. Put only the context that changes what the agent will do.

First Principles

1. Trigger from the description. The frontmatter description is the only part read before the skill loads. Say what the skill does and the concrete situations that should trigger it. Do not hide trigger rules in the body.

2. Spend tokens like they are scarce. Assume the agent is already good at general reasoning. State the direction and let smart models figure out implementation details. A wrapper skill can be two sentences: name the capability to use and the skill to follow. Add mechanics only when a demonstrated failure requires them.

3. Write procedures, not essays. Prefer imperative rules, decision points, and small examples. A good skill changes behavior in the next turn; it does not merely explain the topic. And match the procedure to the deliverable's shape: a catalog of techniques with "pick what fits" produces tool-picking, not a flow — if the job is a shaped interaction (an interview, a staged walk), the workflow must be that shape, with the techniques demoted to steps inside it.

4. Use progressive disclosure. Keep the main skill file short. Push long schemas, examples, provider docs, or variant-specific guidance out into linked references/ files, reached by a pointer that fires only when needed. Put repeatable fragile operations in scripts/. Put reusable output material in assets/.

5. Validate by use. A skill is good when a fresh agent applies it correctly on a realistic task — run it blind, and on the weakest model that will run it; a skill that only drives correct behavior on the frontier model that authored it is too fragile to ship. After editing, read it as if you had no conversation history and remove anything that would not affect action.

6. Examples document the PROBLEM, not the solution. An example earns its tokens by teaching the agent to recognise a recurring problem — the smell, the symptom, how you knew it was wrong. That is durable. The fix you happened to apply is not: code changes, and a baked-in solution goes stale, or worse prescribes a move that won't fit next time. Write the failure mode and its tell; let the agent derive the fix fresh against the current code. "A directional question gated on a centroid distance read wrong from every bearing" teaches; "so we keyed it on FRONT_ARC" rots. When in doubt, state what was broken and how you spotted it, and stop there.

Leading Words

A leading word is a compact concept already in the model's pretraining that the agent thinks with while running the skill (e.g. fog of war, tracer bullets, red, tight). One word recruits priors the model already holds and anchors a whole region of behavior in the fewest tokens.

• It anchors execution in the body (same word → same behavior every time) and invocation in the description (when that word also lives in the user's prompts, docs, and code, the skill fires more reliably).
• Hunt for restatements that a leading word retires. "fast, deterministic, low-overhead" → a tight loop. "a loop you believe in" → the loop goes red. You win twice: fewer tokens and a sharper hook.
• A weak leading word is a no-op (be thorough when the agent already is). The fix is a stronger word (relentless), not more sentences.

Invocation

Choose how the skill is reached; each choice spends a different cost.

• Model-invoked (default): keep a description so the agent can fire it on its own and other skills can reach it. Costs context load — the description sits in the window every turn. Write rich trigger phrasing.
• User-invoked: set disable-model-invocation: true. Only the user typing its name can invoke it; zero context load, but the user must remember it exists. The description becomes a human-facing one-liner.

Pick model-invocation only when the agent or another skill must reach it unprompted. When user-invoked skills pile up past memory, add a router skill that names the others and when to reach for each.

Description

The description does two jobs: state what the skill is, and list the distinct situations (branches) that trigger it.

• Front-load the leading word — invocation work happens here.
• One trigger per branch. Synonyms that rename one branch are duplication; collapse them and keep only genuinely distinct branches.
• Cut identity already in the body. Keep triggers plus any "when another skill needs…" reach clause.

Information Hierarchy

A skill is steps and reference, mixed freely. Rank each piece by how immediately the agent needs it:

1. In-skill step — an ordered action in SKILL.md. Each ends on a completion criterion: make it checkable (can the agent tell done from not-done?) and, where it matters, exhaustive ("every modified model accounted for", not "produce a change list"). A vague criterion invites premature completion.
2. In-skill reference — a definition or rule consulted on demand. A flat peer-set of rules is fine, not a smell.
3. External reference — pushed out of SKILL.md into a linked file, loaded only when its pointer fires.

Keep a concept's definition, rules, and caveats under one heading (co-location) so reading one part brings its neighbors. Push too little down and the top bloats; push too much down and you hide what the agent needs.