Skill Maintenance

Skill Maintenance

Editing conventions for files under .claude/skills/** (including SKILL.md and any references/*.md). Does not apply to markdown elsewhere in the repo. Apply on every edit. Do not retroactively rewrite skills that drift from these rules. Refactor only when you are already editing the skill for another reason.

For the upstream specification, see Anthropic's Skill authoring best practices. The conventions below are repo-local extensions and reinforcements of that guide.

Markdown style

Wrap markdown at 100 characters. In rendered Markdown, single newlines fold into spaces, so wrapping does not affect the output, but it makes side-by-side diffs and review readable.

Preserve the document language. Follow the language established by the skill and its nearby references. Do not switch a French document to English to satisfy a character-set preference. French prose must use UTF-8 with its normal accents and punctuation. Never strip diacritics to make French text ASCII. English prose should stay ASCII when practical.

Skill files are internal instructions, not human-facing project prose. Do not apply humanizer or boileau when creating, translating, rewriting or reviewing them. Keep French instructions clear, idiomatic and correctly accented without loading either writing skill.

No emoji, symbolic icons (warning signs, check marks, decorative arrows, prohibition signs, etc.) or emoticons. For emphasis use bold, headings, or a leading word like NEVER.

No em dashes or en dashes. Do not use -- or - as a substitute; restructure instead: a period (new sentence), a comma (a tight aside), a colon (an explanation), or parentheses (a true aside). This applies in every language.

Frontmatter

Only name and description fields. No version, allowed-tools, or other keys unless a tool actively reads them.

Use the YAML folded block scalar (>-) for any description longer than one line. Wrap lines at 100 chars inside the scalar; YAML folds them into one logical string.

description: >-
  ThreadWaiter pattern for background work in Plastic SCM.
  Use when implementing long-running/network/file operations, replacing `async/await`,
  wiring completion callbacks, or handling `waiter.Exception`.

Avoid:

• description: | (literal: preserves newlines; the trigger string becomes multi-line)
• A single 400-char unwrapped line (works, but unreadable in diffs)

Description anatomy: WHAT + WHEN

A description has two parts:

1. WHAT: one topic sentence covering what the skill knows. No triggers, no rule content.
2. WHEN: concrete user phrases or task contexts that should trigger the skill ("Use when...").

Annotated example:

WHAT:  Plastic SCM UI localization rules.
WHEN:  Use when adding/changing/removing localized UI text, creating
       `PlasticLocalization.Name` keys, editing localization files under
       `bin/client/localization/`, or auditing localization-key usages.

Token economy

The description is the trigger surface, always loaded into the system prompt for every conversation. Every word costs.

Hard ceiling: the folded description MUST be under 1024 characters. This is a hard limit. After writing or editing any description, measure it before saving: count the length of the single string YAML produces (wrapped lines folded together with spaces, blank lines becoming newlines), and confirm it is under 1024. A description that reaches 1024 chars is a broken skill. Do not save it; summarize or cut triggers until it fits. When in doubt, measure rather than eyeball.

Optimize:

• Drop synonyms. "Edit / modify / update" is one trigger, not three. Pick the verb users type.
• Drop duplicates. If the topic sentence already says "for Plastic SCM", the trigger list does not need "in Plastic SCM".
• Drop quality words. "Comprehensive", "robust", "powerful": none affect when the skill loads.
• Drop hedges. "May be useful for", "can help with". Replace with concrete triggers.
• Drop body content. Rules, syntax details, code snippets belong in the SKILL.md body, not the description.

Trigger precision

Goal: load when needed, not when not.

• Too narrow: skill never triggers; capability is dead weight.
• Too broad: skill loads on unrelated tasks; wastes context budget.

Heuristic: a trigger phrase is well-calibrated when (a) you can imagine a user typing it verbatim or in a near-paraphrase, and (b) you cannot imagine the phrase appearing in a task where this skill is irrelevant.

When to refactor a description

Rewrite a description on the next edit pass to that skill when:

• It reaches the 1024-char hard ceiling -- e.g. a needed new trigger would push the folded length over the limit. This is the only length-based trigger: do not trim a description merely for being near the ceiling. A long description that still fits under 1024 and reads clearly is fine and needs no refactor, even at 950+ chars.
• It restates the same trigger in three different ways.
• It contains marketing or hedging language.
• The body has gained or lost a capability the description does not reflect.

Do not refactor pre-emptively. Proximity to the ceiling, on its own, is never a reason to act -- only an edit that would actually cross 1024 forces the length-driven rewrite.

After Editing a SKILL.md Body

Whenever you change a SKILL.md body, re-read its description and decide:

• Add triggers if the body now covers a capability whose user-facing phrasing is not in the description.
• Remove triggers if the body no longer covers something the description still claims.
• Tighten triggers if a phrase has become a duplicate or synonym of another.
• State explicitly that no change is needed if the description still covers the body. Do not skip the check silently.

The same check applies when editing a references/*.md file if the reference introduces a new capability or removes one.

SKILL.md vs references/

SKILL.md is always-loaded once the skill triggers. Keep it lean: rules, decisions, pointers.

Push to references/<topic>.md:

• Setup, install, troubleshooting (loaded on demand only when broken)
• Verbose examples, full schemas, edge-case tables
• Project-specific build/test commands

If a rule is needed only during a sub-workflow that already loads a reference, put the rule in the reference. SKILL.md should not pre-empt it.

Canonical guidelines (Anthropic)

The repo-local rules above defer to the upstream Skill authoring best practices. Highlights worth re-stating:

• Hard limits. name <= 64 chars (lowercase letters, digits, hyphens only). description < 1024 chars, a hard ceiling. Measure the folded length before saving (see "Token economy").
• Body size. Keep SKILL.md under ~500 lines / ~5000 tokens. Split into references/ past that.
• Description structure. WHAT (what the skill does) plus WHEN (concrete triggers). Use the imperative ("Use when ...") rather than vague hedges.
• Avoid undertriggering. Claude tends to skip skills that should fire. Make triggers explicit enough, and slightly "pushy" when the topic is one the user will not name directly.
• Test trigger coverage. When adding or refactoring a description, run a few realistic user phrasings through it (literal, paraphrased, abbreviated, casual). Confirm the skill would load for the ones that should match and would not load for the ones that should not.
• Progressive disclosure, three tiers:
  1. Frontmatter: always loaded (cheap, every word counts).
  2. SKILL.md body: loaded when triggered.
  3. references/, scripts/, assets/: loaded only when SKILL.md instructs.

Third-party / upstream skills

Some skills are vendored from an external source and tracked verbatim so they can be re-synced from upstream. humanizer and boileau are examples: leave their files untouched, including frontmatter keys that these conventions would otherwise strip. Editing them locally would create merge friction the next time they are pulled from upstream. If either genuinely needs a change, push it upstream or fork it under a new name rather than diverging the vendored copy in place.

Validating a skill change

A skill's deliverable is agent behavior, not the Markdown files. Validate a change by exercising its procedure: prompt a Coding Agent (ideally a fresh session/subagent on the most powerful model available, opus) with a trigger phrase, and confirm the skill loads and the agent follows it to the intended outcome. Re-reading the edited files is review, not validation.