| name | skill-authoring |
| description | Write, refactor, or improve Claude Code / Claude skill files (SKILL.md). Use when the user asks to "create a skill", "write a SKILL.md", "add a skill for X", "make Claude better at Y", "turn this into a skill", "optimize skill triggers", "my skill isn't firing", or when building any reusable capability that Claude should auto-invoke on matching prompts. Covers trigger-phrase engineering, progressive disclosure, reference-library pattern, scripts, and eval design. |
Skill Authoring
You write SKILL.md files that follow Anthropic's skill-creator best practices. A skill is a folder (skills/<name>/) containing SKILL.md plus optional reference.md, scripts/, and examples.
Anatomy of a good SKILL.md
- Frontmatter —
name (kebab-case, matches folder) and description (the single most important field; this is what the model matches against at runtime).
- Body — instructions the model loads WHEN the skill fires. Keep it tight. Push long content into
reference.md or sibling files loaded on demand.
Description engineering (the whole ball game)
The description is used by Claude's skill-selection classifier. It must:
- Lead with the capability ("Write cold outreach emails", not "An AI-powered assistant for...").
- Enumerate trigger phrases verbatim. Include 10-30 phrase variants: "Write...", "Draft...", "Make this more...", "My email sounds X", "turn this into Y". Pattern:
Use when the user asks to "X", "Y", "Z", or mentions "A", "B", "C".
- Disambiguate vs. sibling skills. Always include
Not for <neighbor skill>; use <that> for....
- Include negative triggers for common false-positives.
- Stay under ~400 tokens — descriptions load every turn.
Body structure (progressive disclosure)
# Skill Name
One-line purpose.
## When to use / not use
Explicit boundaries.
## Core rules (5-10 max, most load-bearing first)
...
## Workflow
Numbered steps. Each step either asks a question, reads a reference file, or produces output.
## References (load on demand)
- @reference.md — deep guide
- @examples/*.md — few-shot
Anti-patterns
- Stuffing the description with "AI-powered, intelligent, comprehensive". Classifier ignores marketing words and over-triggers.
- Inlining 200-line guides in SKILL.md. Move to
reference.md and use @reference.md imports.
- Writing skills that overlap with existing ones without negative triggers.
- Skills without concrete examples. Always include at least one before/after or worked example.
Eval harness
Every skill should ship an evals/ folder with:
positive.md — prompts that MUST fire the skill (10+).negative.md — prompts that MUST NOT fire (10+, includes near-misses).golden/ — expected outputs for positive prompts.
Run evals with the meta-builder's scripts/eval.py (or manually review invocations).
Scripts sidecar
If the skill performs deterministic work (parsing, validating, rendering), ship a scripts/ folder and have SKILL.md instruct the model to invoke them via Bash rather than reimplementing in prose.
References
- @../../agents-architect/references/skill-triggers.md — trigger-phrase patterns and classifier heuristics
- @../../agents-architect/templates/skill.template.md — fill-in-the-blanks scaffold
- @../../agents-architect/references/anthropic-best-practices.md — condensed from Anthropic docs
Output
When asked to author a skill, produce (or edit):
skills/<name>/SKILL.md with tight description + structured body.skills/<name>/reference.md for anything over 150 lines of body.skills/<name>/evals/positive.md and negative.md.- If deterministic:
skills/<name>/scripts/.
Finish with a 3-line summary: name, trigger surface, load-on-demand references.
