Writing For Agents

Writing for Agents

A document an agent reads is a budget, not a manuscript. Every line either changes what the agent does or costs tokens for nothing.

The two loads

Context load — material present every turn: CLAUDE.md lines, and every skill's description. Charged whether or not it is used. Minimise it.

Cognitive load — what a human must remember exists. Paid by you, not the model. It is the price of your own agency; do not optimise it to zero by inlining everything.

A pointer trades context load for cognitive load. Material with no pointer relies on you remembering it.

The description is the pointer

A skill's description is loaded on every turn of every session. It is the highest-leverage sentence in the file and the only part most sessions ever read.

Write it as what it does + when to reach for it. Front-load the trigger word.

• "Ship a spec or a single ticket from start to merged PR." — the work is named first
• "A helpful skill for various shipping-related tasks." — no trigger, no branch, pure cost

One trigger per real branch. Collapse synonyms rather than listing them.

disable-model-invocation: true when the skill should only run when you type it. Use it for anything that publishes, mutates a tracker, or restructures files — prd, spec, grill, triage. Leave it off for skills the model should reach for on its own — ship, implement, review.

Information hierarchy

Three tiers, by immediacy:

1. Step — an ordered action the agent takes. The primary tier.
2. In-file reference — a flat peer-set consulted on demand: a rubric, a label table, a template.
3. Disclosed reference — a sibling file reached by a pointer. Loads only when the pointer fires.

Promote to a sibling file when a section is consulted rather than followed, and is long enough to thin attention around it. ship/TRACKERS.md, triage/LABELS.md, spec/SPEC-RENDER.md, and prd/TEMPLATES.md are all this. Point at them by filename in backticks — bare from the skill's own directory, path-qualified when pointing across skills.

Co-locate: a concept's definition, its rules, and its caveats live under one heading. Splitting them across sections costs a lookup every time.

Completion criteria

Every skill ends on a ## Completion section, and it does real work:

• Checkable — name the command that proves it. "Done when it works" invites the agent to declare victory early.
• Demanding — "every ticket is a sub-issue with correct labels and blocking edges" forces legwork that "tickets are published" does not.

Split a sequence into separate steps only when a later step's visibility tempts the agent to stop short of the earlier one.

Leading words

A leading word recruits a prior the model already holds, so it changes behaviour in one token instead of a paragraph. tracer bullet, seam, prefactor, expand-contract, frontier each replace a paragraph of explanation.

Prefer them to restatement. When you catch yourself explaining the same behaviour twice in different words, there is a leading word for it.

Negation is the failure mode. Prohibiting a behaviour drags it into context. Never guess the slug from the issue title works only because it is immediately followed by what to do instead. A bare prohibition is worse than nothing — state the positive target.

Pruning

Cut, in this order:

1. No-ops — instructions the model already follows by default. "Write clean code", "be careful", "use good judgment". Pure cost.
2. Duplication — each meaning gets one source of truth. Two copies drift, and the duplicate inflates the idea's apparent importance.
3. What the environment already says — file layout, script names, config values. The agent can look. Document only what looking will not reveal.
4. Stale specifics — file paths and code snippets go out of date within a week. Name the module, not the line.

Shorter documents stay correct longer. A document that grows every time something surprises you becomes a document nobody reads.

This repo's conventions

skills/<name>/
├── SKILL.md          ← frontmatter + steps
└── REFERENCE.md      ← SCREAMING-KEBAB, one concern per file

• Frontmatter name: must equal the directory name. The build fails otherwise.
• Every skill in skills/ must appear in skills.enable in modules/home/packages/skills.nix, and nothing may appear there that is not on disk. The nix allowlist is validated at eval.
• Cross-skill references use the backticked flat form: /ship, /implement. An external skill is /<idPrefix>_<skill> — /cursor-team-kit_deslop, never a slash path.
• Reference the PRD by its real files: prd/glossary.md, prd/behaviours.md, prd/architecture.md, prd/decisions.md, prd/product.md, prd/features/<slug>.md. Treat every one as optional — read what exists, say what is missing, continue.
• AI_NOTES.md documents the workflow as a whole. A new skill, a rename, or a cut belongs there too, or it is invisible.

Installing a skill — flake input, source, enable list — is /add-skill. This skill is only about what goes inside the file.

Verify

Run the checks in CHECKS.md. They catch the failures that are invisible on reading: a pointer to a file that does not exist, a name that does not match its directory, an enable list out of step with disk.

Completion

Done when: the description names its trigger, every pointer resolves, no section duplicates another, ## Completion is checkable, and every check in CHECKS.md prints nothing. Checkable: run all seven checks and confirm zero findings.