Create Skill

Create Skill

6-phase workflow: Discover → Design → Plan → Build → Validate → Deliver.

Core Principles

Skill Anatomy

skill-name/
├── SKILL.md         # Behavioural guidance (how to think, what to check)
├── scripts/         # Deterministic, fragile operations (tested code)
├── references/      # Look-up material, loaded on demand
└── assets/          # Output templates, structure only

Three-Layer Separation

Every piece of content belongs in exactly one layer:

• Behavioural (SKILL.md) — how to think, what to check, what principles
• Format (assets/) — what the output structure looks like
• Tool (references/) — CLI flags, API params, external commands

No mixing layers is the most common new-skill mistake.

Degrees of Freedom

Match instruction tightness to task fragility:

• High (text guidance) — exploratory, context-dependent decisions
• Medium (pseudocode/script with params) — preferred pattern, some variation OK
• Low (specific script, few params) — fragile, consistency critical

See references/degrees-of-freedom.md for the full decision tree.

CSO Rule for Descriptions

The description field is the only metadata the agent sees to decide whether to load a skill. Never summarise the workflow — describe what the skill does and when to use it, not how it works.

See references/anti-patterns.md for bad/good examples.

Workflow

Phase 1: DISCOVER — Understand the Skill's Purpose

Collect answers before writing anything.

1a — Five Core Questions

Ask the user:

1. One-liner: What does this skill do in one sentence?
2. Inputs: What does it read or receive?
3. Outputs: What does it produce?
4. Consumer: Who or what consumes the output?
5. Pitfalls: Known gotchas or environment-specific facts?

1b — Degrees of Freedom

For each operation, determine instruction tightness by asking:

• Flexibility or exact sequence?
• Fragile or exploratory?
• One right way or many valid approaches?

Use references/degrees-of-freedom.md for the decision tree.

Phase 2: DESIGN — Preview the Expected Interaction

Before building anything, show the user what the skill will look like.

2a — Mock Dialogue

Construct a 3–4 turn dialogue: user triggers the skill → agent responds, including one edge case (missing input, failure recovery). Format:

User: <realistic trigger request>
Agent: <expected response>

User: <follow-up or edge case>
Agent: <handling or final output>

Present to the user: "Is this the interaction you expect?"

• YES → proceed.
• NO → revise and reconfirm.

Phase 3: PLAN — Map Content to Files

3a — Classify Every Element

One item → one bucket. No mixing.

3b — Design Directory

proposed-name/
├── SKILL.md
├── scripts/      # only if Step 3a produced script items
├── references/   # only if Step 3a produced reference items
└── assets/       # only if Step 3a produced asset items

Minimal viable skill: just SKILL.md.

Phase 4: BUILD — Create the Files

4a — Scaffold

mkdir -p skills/<skill-name>/{scripts,references,assets}
rmdir <empty-dirs>

Or use scripts/init-skill.sh <skill-name> with optional --with-* flags.

4b — Write SKILL.md

Use assets/skill-template.md for structure. Rules:

• name matches directory name (lowercase + hyphens, < 64 chars)
• description follows CSO Rule — concrete triggers + negative triggers
• Third-person imperative: "Verify..." not "You should verify..."
• Under 500 lines. Split dense content into references/
• Include ## Gotchas section with environment-specific facts
• Reference files explicitly: "See references/schema.md for structure"

4c — Write Supporting Files

• scripts/: Tested executables. Accept CLI args/stdin, output to stdout, descriptive error messages so the agent can self-correct. Run before finalising.
• references/: One topic per file. ToC if >100 lines. No behavioural guidance.
• assets/: Structure only — {{placeholder}} tokens. No instructions or rules.

Phase 5: VALIDATE — Ensure Quality

Run these checks in order. Fix failures before proceeding.

5a — Discovery Validation

Present the frontmatter to yourself: "Based on this name + description, generate 3 prompts that SHOULD trigger this skill and 3 that should NOT. Critique the description and propose revisions if needed."

Fix the description if it produces false positives/negatives.

5b — Simulated Execution

Walk through the SKILL.md step-by-step as an agent running it. Flag any line where you'd have to guess or hallucinate. Clarify those instructions.

5c — Edge Case Attack

Ask: "What if a referenced file is missing? Input violates assumptions? Script fails mid-execution?" Fix gaps found.

5d — Structural Checks

Run python scripts/validate-skill.py skills/<skill-name> if available.

Manually verify:

• Same inputs → same outputs (core deliverable preserved)
• Description includes negative triggers
• No behavioural guidance in templates — structure only
• No references marked as required reading
• SKILL.md under 500 lines
• No duplicate or overlapping steps
• All cross-references valid
• Gotchas section present with concrete facts
• No instructions the model would follow correctly without being told
• No README, CHANGELOG, or human docs in skill directory

Phase 6: DELIVER — Present the Result

Show the user:

1. Directory structure of the new skill
2. Summary of each file's purpose
3. Key decisions (script vs inline, reference extracted, etc.)
4. Point to optimise-skill for a token-efficiency pass

✅ Skill created: skills/<name>/

<name>/
├── SKILL.md          — <N> lines, core workflow + gotchas
├── scripts/          — <what scripts do>
├── references/       — <what references cover>
└── assets/           — <what templates provide>

Example: "The script was extracted because JSON parsing is fragile when written from scratch. Reference material was split out to keep SKILL.md under 200 lines."

Gotchas

• Empty subdirectories confuse agents. Only create scripts/, references/, or assets/ when they have actual files.
• All routing metadata goes in the description frontmatter field. Don't add a "When to Use" section in the body.
• name must match the directory name (lowercase, hyphens, < 64 chars). If they diverge, some agents won't load the skill.
• No human-facing docs (README, CHANGELOG, LICENSE) inside a skill directory. Skills are for agents, not humans.
• Default assumption: the model is already smart. Don't explain basic concepts (HTTP, JSON, git). Add only what's project-specific, non-obvious, or where the default behaviour would be wrong.
• Mock dialogues are design tools, not part of the output. Delete before delivering.

References

• references/spec.md — agentskills.io spec summary
• references/anti-patterns.md — common skill-writing mistakes with fixes
• references/degrees-of-freedom.md — decision tree for instruction tightness
• assets/skill-template.md — structure template for the generated SKILL.md

Scripts

• scripts/validate-skill.py — structural validation for generated skills
• scripts/init-skill.sh — scaffold an empty skill directory