Writing Great Skills

A skill exists to wrangle determinism out of a stochastic system. Predictability — the agent taking the same process every run, not producing the same output — is the root virtue; every lever below serves it.

Bold terms are defined in GLOSSARY.md; look them up there for the full meaning.

Local Codex use

This copy is adapted for the local shared skill roots:

• /Users/stevmq/.agents/skills for shared assistant skills.
• /Users/stevmq/.codex/skills for Codex-specific skills.

Treat those folders as source of truth. Installer output and registry listings are supporting evidence; the files on disk decide what is actually installed.

When adapting a skill for this environment:

• Improve an existing skill before creating a near-duplicate. If an existing skill covers roughly 70 percent of the workflow, tighten that skill.
• Keep global operating preferences in AGENTS.md or durable instructions, not buried inside a narrow skill.
• Encode operational contracts, not advice: scope, trigger, default action, evidence, permission boundary, stop rule, validation, and closeout.
• Prefer action-first language. Ask the user only when a wrong assumption would be costly and local context cannot answer the question.
• Preserve evidence boundaries. Say what is validated locally, verified live, blocked, inferred, or memory-derived.
• Treat memory as routing context, not proof. Recheck drift-prone facts from current files, tools, logs, rows, browser state, or official docs.
• Keep powerful or mutating behavior opt-in: installs, issue comments, label changes, repo-wide rewrites, credentials, deployment, and destructive cleanup need explicit scope.

Invocation

Two choices, trading different costs:

• A model-invoked skill keeps a description, so the agent can fire it autonomously and other skills can reach it (you can still type its name too). It contributes to context load — the description sits in the window every turn. Mechanics: omit disable-model-invocation, and write a model-facing description with rich trigger phrasing ("Use when the user wants…, mentions…").
• A user-invoked skill strips the description from the agent's reach: only you, typing its name, can invoke it — and no other skill can. Zero context load, but it spends cognitive load: you are the index that must remember it exists. Mechanics: set disable-model-invocation: true; the description becomes human-facing — a one-line summary, trigger lists stripped.

Pick model-invocation only when the agent must reach the skill on its own, or another skill must. If it only ever fires by hand, make it user-invoked and pay no context load.

When user-invoked skills multiply past what you can remember, that piled-up cognitive load is cured by a router skill: one user-invoked skill that names the others and when to reach for each.

Writing the description

A model-invoked description does two jobs — state what the skill is, and list the branches that should trigger it. Every word increases context load, so a description earns even harder pruning than the body:

• Front-load the skill's leading word — the description is where it does its invocation work.
• One trigger per branch. Synonyms that rename a single branch are duplication — "build features using TDD … asks for test-first development" is one branch written twice. Collapse them; keep only genuinely distinct branches.
• Cut identity that's already in the body. Keep the description to triggers, plus any "when another skill needs…" reach clause.

Information hierarchy

A skill is built from two content types — steps and reference — that mix freely: a skill can be all steps, all reference, or both. The core decision is which to use and where each sits on the information hierarchy, a ladder ranked by how immediately the agent needs the material:

1. In-skill step — an ordered action in SKILL.md, the primary tier: what the agent does, in order. Each step ends on a completion criterion, the condition that tells the agent the work is done. 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, rule, or fact in SKILL.md, consulted on demand. Often a legitimately flat peer-set (every rule of a review on one rung) — a fine arrangement, not a smell. This skill is all reference.
3. External reference — reference pushed out of SKILL.md into a separate file, reached by a context pointer, loaded only when the pointer fires. (Spans disclosed reference — a sibling file like GLOSSARY.md, still part of the skill — through fully external reference that lives outside the skill system and any skill can point at.)

A demanding completion criterion drives thorough legwork — the digging the agent does within the work — whether the skill has steps or not, since "every rule applied" binds flat reference just as "every step done" binds a sequence.

Push too little down and the top bloats; push too much and you hide material the agent actually needs. That tension is the whole decision.

Progressive disclosure is the move down the ladder — out of SKILL.md into a linked file — so the top stays legible. Mechanics: a linked .md file in the skill folder, named for what it holds (this skill discloses its full definitions to GLOSSARY.md). Some skills are used in more than one way, and each distinct way is a branch — different runs taking different paths through the skill. Branching is the cleanest disclosure test: inline what every branch needs, and push behind a pointer what only some branches reach. A context pointer's wording, not its target, decides when and how reliably the agent reaches the material.

Where the ladder decides how far down a piece sits, co-location decides what sits beside it once there: keep a concept's definition, rules, and caveats under one heading rather than scattered, so reading one part brings its neighbours with it.

When to split

Granularity is how finely you divide skills, and each cut spends one of the two loads, so split only when the cut earns it. Two cuts:

• By invocation — split off a model-invoked skill when you have a distinct leading word that should trigger it on its own, or another skill must reach it. You pay context load for the new always-loaded description, so that independent reach has to be worth it.
• By sequence — split a run of steps when the steps still ahead (a step's post-completion steps) tempt the agent to rush the one in front of it (premature completion). Keeping them out of view encourages the agent to do more legwork on the current task.

Pruning

Keep each meaning in a single source of truth: one authoritative place, so changing the behaviour is a one-place edit.

Check every line for relevance: does it still bear on what the skill does?

Then hunt no-ops sentence by sentence, not just line by line: run the no-op test on each sentence in isolation, and when one fails, delete the whole sentence rather than trim words from it. Be aggressive — most prose that fails should go, not be rewritten.

Leading words

A leading word is a compact concept already living in the model's pretraining that the agent thinks with while running the skill (e.g. lesson, fog of war, tracer bullets). Repeated throughout the text (though not necessarily - a strong leading word might only be needed once), it accumulates a distributed definition and anchors a whole region of behaviour in the fewest tokens, by recruiting priors the model already holds.

It serves predictability twice. In the body it anchors execution: the agent reaches for the same behaviour every time the word appears. In the description it anchors invocation: when the same word lives in your prompts, docs, and code, the agent links that shared language to the skill and fires it more reliably.

Hunt for opportunities to refactor skills to use leading words. A triad spelled out at three sites (duplication), a description spending a sentence to gesture at one idea — each is a passage begging to collapse into a single token. Examples include:

• "fast, deterministic, low-overhead" -> tight — one quality restated across a phase — into a single pretrained word (a tight loop).
• "a loop you believe in" -> red — converts a fuzzy gate into a binary observable state (the loop goes red on the bug, or it doesn't).

You win twice over: fewer tokens, and a sharper hook for the agent to hang its thinking on. Assume every skill is carrying restatements that leading words retire — go find them.

Failure modes

Use these to diagnose issues the user may be having with the skill.

• Premature completion — ending a step before it's genuinely done, attention slipping to being done. Defence, in order: sharpen the completion criterion first (cheap, local); only if it is irreducibly fuzzy and you observe the rush, hide the post-completion steps by splitting (the sequence cut).
• Duplication — the same meaning in more than one place. Costs maintenance and tokens, and inflates a meaning's prominence on the ladder past its real rank.
• Sediment — stale layers that settle because adding feels safe and removing feels risky. The default fate of any skill without a pruning discipline.
• Sprawl — a skill simply too long, even when every line is live and unique. Hurts readability and maintainability and wastes tokens. The cure is the ladder: disclose reference behind pointers, and split by branch or sequence so each path carries only what it needs.
• No-op — a line the model already obeys by default, so you pay load to say nothing. The test: does it change behaviour versus the default? A weak leading word (be thorough when the agent is already thorough-ish) is a no-op; the fix is a stronger word (relentless), not a different technique.