Setting the file. One moment.
Skill 04 · Writing Great Skills
Subchapter 4.1
GLOSSARY.mdMarkdown17 KBView on GitHub
The domain model for what makes a skill great. A skill exists to wrangle determinism out of a stochastic system; the root virtue is Predictability, and every term below is a lever on it. This is the disclosed reference for writing-great-skills.
The terms are grouped by axis: Invocation (how a skill is reached), Information Hierarchy (how its content is arranged), Steering (how the agent’s runtime behaviour is shaped), and Pruning (how it is kept lean). Each failure mode lives beside the lever that cures it, tagged failure mode.
Bold terms in any definition are themselves defined in this glossary; find them by their heading.
The degree to which a skill makes the agent behave the same way on every run — the same process, not the same output (a brainstorming skill should diverge; its tokens vary, its behaviour doesn’t). The root virtue every other term serves — cost and maintainability are symptoms of it, not rivals.
Avoid: consistency, reliability, robustness, output-determinism
How a skill is reached — and the two loads you pay for the choice.
A skill that keeps its description field, so the agent can see it and fire it autonomously — and the human can still type its name, so model-invocation always includes user reach. There is no model-only state: a description only ever adds agent discovery, never removes the human’s. Pays a permanent context load on every turn in exchange for that discoverability. Reachable by other skills, because the description that makes it agent-discoverable makes it invocable. A model-invoked skill whose content is all reference is also one home for shared reference: another skill can invoke it, so reference needed by several skills lives in one place. Pick model-invocation only when the agent must reach the skill on its own; if it never fires except by hand, drop the description and pay no context load.
Avoid: ability, tool, capability
A skill with its description stripped — invisible to the agent and reachable only by the human typing its name (user-only, where model-invoked is user-and-agent). Trades agent-discoverability for zero context load. Because it has no description, nothing but the human can reach it: no other skill can fire it.
Avoid: procedure, workflow, command
The skill’s machine-readable trigger, and the one context pointer a model-invoked skill is forced to keep loaded at all times. Its mere presence is the invocation axis: keep it and the skill is model-invoked (and reachable by other skills); delete it and the skill is user-invoked, reachable only by the human. The source of a model-invoked skill’s context load.
Avoid: frontmatter, summary
A reference held in the agent’s context that names some out-of-context material and encodes the condition for reaching it. The description is the top-level context pointer (context window → skill); pointers to disclosed files are the same object one level down. Its wording, not the target, decides when the agent reaches — and how reliably. A must-have target behind a weakly worded pointer is a variance bug: fix the wording first, and inline the material only if sharpening fails.
Avoid: link, reference, import
The cost a model-invoked skill imposes on the agent’s context window — its description, always loaded, spending both tokens and attention. What user-invoked skills escape by having no description, and the brake on splitting into more model-invoked skills.
Avoid: token cost, context bloat
The cost a user-invoked skill imposes on the human — what they must hold in their head: which skills exist and when to reach for each (the human is the index). What model-invocation removes by being agent-discoverable, and the brake on splitting into more user-invoked skills. Not a cost to minimise: it is the price of human agency, the reason some skills stay user-invoked. Spend it where human judgement matters; remove it where it does not.
Avoid: human index, burden, overhead
A user-invoked skill whose job is to point at your other user-invoked skills — naming each and when to reach for it — so the human has one skill to remember instead of many. It can only hint, never fire them: user-invoked skills have no description, so nothing but the human can reach them. The cure for cognitive load when user-invoked skills multiply.
Avoid: dispatcher, menu, registry, index, router procedure
How finely you divide skills. Finer division spends one of the two loads: more model-invoked skills spend context load (more descriptions crowding the window and competing for attention); more user-invoked skills spend cognitive load (more for the human to remember and reach for). Two cuts guide the division. By invocation, split off a model-invoked skill where you have a distinct leading word to trigger it — a trigger word you actually use in your prompts. By sequence, split a run of steps where a step’s post-completion steps need hiding, since isolating it in its own context clears what follows. Beware the reverse: merging sequences exposes each step’s post-completion steps to what follows, inviting premature completion.
Avoid: chunking, modularity
How a skill’s content is arranged, and how far down the ladder each piece sits.
A skill’s content ranked by how immediately the agent needs it — a single ladder, produced by two cuts: in-file or behind a pointer, and step or reference. The rungs:
A skill with no steps uses just the bottom two rungs — often a legitimately flat peer-set (e.g. every rule of a review on one rung), which is a fine arrangement, not a smell. The hierarchy is independent of invocation: a skill can be model- or user-invoked whether it is all steps, all reference, or both. When a skill has steps, in-file reference that should be disclosed buries them and turns attending to them into a coin-flip — a variance lever, not just a legibility one. Keep the top of the ladder legible; push down it whatever you can.
Avoid: structure, organization, layout
The ordered actions the agent performs — when a skill has them, the primary tier of its content, and the part that earns its place in SKILL.md. Not every skill has steps: a skill can be all steps (tdd), all reference (a review), or both, independent of invocation. Every step ends on a completion criterion, clear or vague.
Avoid: workflow, instructions, choreography
Material the agent refers to on demand — definitions, facts, parameters, examples, conditional instructions. When a skill has steps it is secondary to them; when a skill has none it is the entire content; or it lives outside any skill entirely — see External Reference. Reached via context pointers, and the prime candidate for progressive disclosure.
Avoid: supporting material, docs, background
Reference that lives outside the skill system — a plain file, no description, no steps, not invocable — that any skill can point at. The home for shared reference that needn’t fire on its own, and the only shared home two user-invoked skills can use, since neither has a description and so neither can fire the other.
Avoid: doc, resource, knowledge base
Moving reference down the ladder — out of SKILL.md and behind a context pointer — so the top stays legible. Not primarily a token optimisation; it is how the information hierarchy is protected. Licensed by branching: disclose what only some branches need, inline what every path needs, and if a pointer fires unreliably on must-have material, sharpen its wording, and pull it back inline only if that fails.
Avoid: lazy loading, chunking
Keeping the material an agent needs at once in one place — a concept’s definition, rules, and caveats under a single heading, not scattered across the file — so reading one part brings its neighbours with it. The within-file companion to the Information Hierarchy: the hierarchy ranks how far down a piece sits; co-location decides what sits beside it once there. There is no formula for the right format of a body of reference; the test is that a skill should read like documentation written for the agent, and grouped material reads that way where scattered material does not. Distinct from Duplication: that repeats one meaning in two places, where scattering fragments a single meaning across many.
Avoid: grouping, clustering, cohesion
Failure mode. A skill that is simply too long — too many lines in SKILL.md — independent of whether they are stale or repeated. Even an all-live, all-unique skill can sprawl. It costs readability (the agent wades through more before it can act, and attention thins across the excess), maintainability (every extra line is one more to keep relevant), and tokens. The cure is the information hierarchy: push reference down behind context pointers, and split by branch or sequence so each path carries only what it needs. Distinct from sediment (length from stale accumulation) and duplication (length from repeated meaning) — sprawl is length itself, whatever its cause.
Avoid: bloat, length, size, verbosity
The levers that shape the agent’s runtime behaviour toward Predictability.
A distinct way a skill can be invoked — a case the skill handles — so different runs take different paths through it. A skill with many steps may carry many branches; a linear one has none.
Avoid: path, case, fork
A compact concept — also called a Leitwort — already living in the model’s pretraining, that the agent thinks with while running the skill. It encodes a behavioural principle in the fewest possible tokens by invoking priors the model already holds (e.g. lesson, proximal zone of development, fog of war, tracer bullets). Repeated as a token, never as a sentence, it accumulates a distributed definition across the skill and anchors a whole region of behaviour. Coining your own works if you define it clearly, but a made-up word recruits no priors — you pay in definition tokens what a pretrained word gives free. Reach for an existing word first.
A leading word serves predictability twice. In the body it anchors execution — the agent reaches for the same behaviour every time the concept appears, and inside flat reference it focuses attention on a class of thing to look for, recruiting the right checks each run. In the description it anchors invocation — and not only within the skill: when the same word lives in your prompts, your docs, and your codebase, the agent links that shared language to the skill and fires it more reliably. Word a description with the leading words you actually use when you want the skill.
Avoid: keyword, term, motif
The condition that tells the agent a unit of work is done — the target it judges against. Two properties make it a lever, not just a quality. Its clarity (can the agent tell done from not-done?) resists premature completion — a vague bound (“understanding reached”) lets the agent declare done and slip to the next step; this axis needs steps to bite, since premature completion is a between-steps failure. Its demand (how much it requires) sets legwork — “every modified model accounted for” forces thorough work where “produce a change list” does not — and this axis is not step-bound: it can bind a body of flat reference too, which is how a skill with no steps still carries an exhaustiveness bar (“every rule applied”). The strongest criteria are both checkable and exhaustive.
Avoid: done condition, exit condition, stopping rule
The work an agent does behind the scenes within a single step — reading files, exploring the codebase, making changes, digging up what it needs rather than offloading to the user. It lives below the step structure: never written as its own step, latent in the wording, controlled by the agent rather than the skill. The within-step counterpart to post-completion steps‘ across-step pull. Raised by a leading word (comprehensive, thorough) or a completion criterion that demands the work be exhaustive — including the demand axis applied to flat reference, which is what drives a skill of flat reference to cover all its rungs. Goes thin either when that demand is missing or when premature completion cuts the step short.
Avoid: scope, effort, diligence, coverage
The steps that follow the current step. Visible, they pull the agent forward into premature completion — the more it sees, the stronger the tug; the defence is to hide them by splitting the sequence of steps into two.
Avoid: horizon, fog of war, lookahead
Failure mode. Ending the current step before it is genuinely done, because the agent’s attention slips to being done rather than to the work. A between-steps failure: it needs steps to occur — a skill with no steps that quits early isn’t premature completion but thin legwork under an unmet demand. A tug-of-war between two forces: visible post-completion steps (the pull forward) and the completion criterion‘s clarity (the resistance — a sharp, checkable bar holds; a vague one gives way). Fuzziness is the necessary condition: a sharp bound resists the pull no matter how many later steps are visible, so a step that never rushes needs no defending. Two levers hold a step that does, but reach for them in order: sharpen the bound first — it is local and cheap. Only when the criterion is irreducibly fuzzy and you actually observe the rush do you hide the later steps — and hiding only works across a real context boundary (a user-invoked hand-off or a subagent dispatch; an inline model-invoked call leaves the later steps in context and clears nothing). One cause of thin legwork, but distinct from it: legwork can be thin even when a step runs to full completion.
Avoid: premature closure, the rush, rushing, shortcutting
Keeping a skill lean — each remedy paired with the failure it cures.
The desired state where each meaning lives in exactly one authoritative place, so a change to the skill’s behaviour is a change in one place. Duplication is its violation.
Avoid: home, canonical location
Failure mode. The same meaning given more than one single source of truth. It costs maintenance (change one place, you must change the others), costs tokens, and inflates prominence — repeating a meaning weights it on the ladder past its real rank. The accidental inverse of a leading word, which raises attention on purpose by repeating a token, never the meaning.
Avoid: repetition, redundancy
Whether a line still bears on what the skill does — the lens for what to keep. A line loses relevance either by never bearing on the task (mere exposition, or a branch that should be disclosed) or by going stale: drifting out of date as the behaviour or world it describes changes. Shorter skills are easier to keep relevant, because each line is cheaper to check. Distinct from no-op: relevance asks whether a line bears on the task, not whether it changes behaviour.
Avoid: load-bearing, staleness, freshness
Failure mode. Layers of old content that settle in a skill and are never cleared, because adding feels safe and removing feels risky — so stale and irrelevant lines accumulate and you must core down through them to find what is still live. The default fate of any skill without a pruning discipline; the slow erosion of relevance, as opposed to duplication‘s repeated meaning.
Avoid: accretion, bloat, cruft, rot
Failure mode. An instruction that changes nothing because the model already does it by default — you pay load to tell the agent what it would do anyway. The test: does a line change behaviour versus the default? A line can be perfectly relevant and still be a no-op. The same priors that make a leading word free make a no-op worthless.
A leading word is a technique; No-Op is a verdict on a line — and they cross. A leading word too weak to beat the default is a no-op (be thorough when the agent is already thorough-ish), and the fix is a stronger word that passes the verdict (relentless), not a different technique. So the No-Op test — does it change behaviour versus the default? — is also how you grade whether a leading word is earning its repetitions. This is model-relative, not reader-relative: two people disagreeing over whether a line is a no-op disagree about the default, and settle it by running the skill, not by debate.
Avoid: redundant instruction, restating the obvious, belaboring