Moo Authoring

---
name: moo-authoring
description: Use when adding or changing a skill, hook, or runtime file in this repo — including deciding which unit a new capability should be.
metadata:
  internal: true
---

Doctrine for authoring moo's own surfaces.

## Frontmatter

A skill's frontmatter carries `name` and `description`; read any shipped `SKILL.md` for the current key set. The description is one line, and Claude Code caps it at 1024 characters.

No `when_to_use`. Claude Code reads that key and renders it dash-joined onto the description, but every other agent copies `SKILL.md` verbatim and never reads it, so the text deciding when a skill fires would be invisible wherever it is installed outside Claude Code.

A condition that follows what the skill does is marked `Use when <condition>.` — a finite clause, never a bare noun phrase or a bare dash. The key name was what marked the condition as a condition; in one string nothing else does, and an unmarked trailing clause reads as more of what the skill produces. A description may also open with the condition and stop there, where naming the capability would discriminate nothing.

A skill that stays in this repo declares `metadata.internal: true`. The skills.sh installer walks the tree and offers every `SKILL.md` it finds; that flag is what it filters on.

Version lives in `plugin.json` only (DRY). The official Claude Code spec does not allow `version` in SKILL.md frontmatter.

## Sizing

- Size a skill by a deletion pass, never a numeric cap.
- Put branch-specific content in files the skill opens when the branch is taken, and name them where they are needed.
- Decision tables > prose explanations.

## Token Efficiency

- Challenge every sentence: "Does Claude need this?"
- Bullets for enumerable items; a paragraph when the point is one connected argument.
- No vague terminology; pick one term per concept.

## Skill Design

Phrase design decisions as "X over Y: reason".

**Unit choice** — behavior inlines at build, data references at runtime; a skill's firing is probabilistic, so behavior that must run every time is a hook:

| The new thing is... | Unit |
|---|---|
| Data selected per use (catalog, profile, corpus) | Runtime file |
| A trigger + procedure that stands alone | Skill |
| A trigger only the human perceives | Skill with `disable-model-invocation: true` |
| Behavior that must run every time, deterministically | Function hook |
| An unproven idea | hunch skill + `HYPOTHESIS.md`; graduates or dies |
| Doctrine for this repo's own surfaces | Repo-local skill in `.claude/skills/`; never ships |

**Composition — name-calls, never imports:**

- Skills compose by name: a composer names a move and invokes it with the Skill tool; a body never inlines another body.
- New skill when the cognitive mode changes (clarify WHAT ≠ decide HOW ≠ judge unsupervised work; find ≠ fix). One skill = one mode + one gate.
- Two skills over one when triggers differ; a shared explanation the pair needs lives in this file, which owns skill-design reasoning, not in either skill.
- A body names a sibling in exactly two places: a refusal, and a route — wherever in the body those sit. There the name *is* the refusal's or the route's content, and a reader without it is stranded.
- Ending over hand-off: a finished run states what is now settled and what is still open, in its own words. Naming who handles it next hardcodes the graph n², so one trigger change edits every sibling.

**Output form — bound by what the skill hands back, never by one shared sentence.** Plain words is the only part that holds across all three:

| The Output hands back | Bound it by |
|---|---|
| Something the agent restated in its own words | As short as it goes with nothing *the skill can't afford to drop* lost — name that thing, it differs per skill |
| The user's own words — a record, an amended proposal | Preserving them. A compression clause here destroys the result |
| A structure the reader navigates — verdict then evidence, a case set then a rule | Bounding each part where a reader would pad it, never the whole |

A cap counted in lines is a compression budget on a handed-back artifact — replace it. A cap on one turn of a loop is pacing, and the rest arrives on the user's next pull — keep it.

## Hook Design

Every hook is a function hook: it reaches the human with no model turn. The API is early access; `plugin-authoring` and `/plugin-types` are the reference for every event and call.

- A hook never starts a turn the human didn't ask for: no wake, no `$.prompt.submit` of its own.
- What a hook finds waits where the human looks next, the band above the prompt, and appears there only while it holds something. Detail unfolds on a press and folds back.
- A hook hands to Claude only on the human's Enter: `$.prompt.fill` a sentence start they finish, with any bulk added as `prompt.submit` context.
- Work longer than one dispatch runs unawaited, so the turn's end is never held.
- State lives in `$.state` (this session) or `$.store` (across sessions), never in temp files or module variables. A hot reload drops module variables.
- A failure shows once, as a toast, and the event passes on.
- A pane or band is verified on events copied from real session transcripts, one per shape it renders; a synthetic event alone never verifies it.
- One hooks module per plugin; a second part composes into it. A second registration on the same event and matcher in one plugin is dropped silently.

A hook that spawns headless `claude -p` copies its flag set from `hope/hooks/judge.sh`, where each flag is commented at the point of use.