attac-t Skill: Craft Skill
"Atomic units of intelligence."
The Principles
- Atomicity: One concept per file. If it does two things, split it.
- Brevity: Less is more. No verbose explanations.
- Expertise: Write how an expert thinks, not checklists.
- Eighty lines, and it is ours: front matter and all. Anthropic's published ceiling is 500, for performance. Eighty is stricter and answers two questions it does not. A skill is read into attention whole, so its length is what crowds out the next one — and a person has to read it too. Eighty is a screen. Past it, a skill is two skills.
- Atomic examples: one subject per file, named for it. Anthropic calls this progressive
disclosure, and adds the rule that matters — references stay one level deep from
SKILL.md, so a file it sends you to is read whole rather than chased. - A description is read at 250 characters: 1,024 is the hard limit, but only the first 250 reach the listing Claude picks from. A trigger word past 250 is invisible.
Choose a Template
| Template | Use When | Pattern |
|---|---|---|
| decide | X vs Y decision | Heuristic + Quick Test + examples.md |
| craft | How to build X | Standard + Anti-Patterns + examples.md |
| ground-protocol | Step-by-step ritual | When + Protocol + Output |
| ground-philosophy | Mindset/standards | Standard + Check + examples.md |
The Structure
plugins/{plugin-name}/skills/{skill-name}/
├── SKILL.md # The skill definition
├── shape.md # One subject per file, named for it
└── traps.md # Another subject, another file
Naming: Flat directories with prefix. ground-vue/ not ground/vue/.
Path: Skills live at plugins/{plugin}/skills/, not .claude-plugin/skills/.
plugin.json: Uses "skills": "./skills/" (resolved relative to plugin root).
Official Spec
Use claude-code-guide to query Claude Code skill documentation. The numbers above came from it
on 2026-08-22 — ask again rather than trusting them.
Preloading
[!IMPORTANT]
skills:is a preload list, not an allowlist.
An agent's skills: frontmatter injects the full skill content into its context at startup.
Omitting a skill does not deny it — sub-agents can still invoke any project, user, or plugin skill
through the Skill tool at runtime.
So register deliberately, not defensively:
| Preload it | Leave it out |
|---|---|
| The agent needs it on every run | The agent needs it occasionally |
| Fetching it late would change the output | A Skill call at the right moment is fine |
| It defines the agent's standards or vocabulary | It is reference material |
Every preloaded skill is startup context the agent pays for whether or not it uses it. A long
skills: list is a cost, not a safety net.
Add to agents/architect.md when the architect genuinely needs it loaded — not as a formality.