Skill Creator

Skill Creator

Guidance for creating and maintaining high-quality Agent Skills across any AI agent ecosystem.

What Is a Skill

A skill is a self-contained folder that gives an AI agent specialized knowledge, workflows, and tools for a specific domain. Skills use the open AgentSkills specification.

Directory Structure

skill-name/
├── SKILL.md          # Required — frontmatter + instructions
├── scripts/          # Optional — executable code agents can run
├── references/       # Optional — docs loaded on demand into context
└── assets/           # Optional — templates, images, data files used in output

Core Design Principles

1. Brief, Clear, Specific, Useful

Assume the agent already knows common domain facts and standard tool use. Include only context, constraints, decision criteria, or reusable mechanics that materially change its work. Match specificity to risk: describe outcomes and choices when several approaches work; prescribe exact steps or scripts only for fragile or deterministic operations.

2. Progressive Disclosure

Design for staged loading to keep the context clean:

• Discovery: name + description only
• Activation: full SKILL.md body (baseline workflow, routing, and task guidance)
• On demand: explicit triggers load files in scripts/, references/, assets/

If a workflow gets deeply specific, move it to references/ so the agent only loads it when that specific subtask triggers.

3. Agent-Neutral Language

Use agent-neutral wording for portable behavior. Name a product only when its runtime, metadata, tools, or distribution are part of the capability; keep those details scoped and declare relevant compatibility.

4. No Meta-Justification

Keep SKILL.md and references/ files stripped of benchmarks, "why we built this" defenses, and generic README material. Only include actionable rules and necessary constraints. Tell the agent what to do and the operational why (e.g., "because command X hangs the service"), not the philosophical why.

Skill Creation Process

Follow these steps to build or refactor a skill:

Step 1: Understand or Audit

Use the conversation and target workspace before asking questions.

• New skill: Capture representative requests, inputs, expected outputs or behavior, dependencies, and near misses. Ask only for missing information that changes the design.
• Update: Read SKILL.md, relevant resources, and repository conventions. Name the concrete gap, success and failure criteria, and what must remain stable.

Separate durable requirements from one-off examples, failures, and preferences. Preserve the skill's name, scope, supported metadata, and authorization boundaries unless the user requests a change.

Step 2: Plan Resources

Start instruction-only. Add a resource only when it repeatedly helps the agent execute the skill:

• scripts/: Use when the same code is rewritten each time or deterministic output is required.
• references/: Use for specific subtasks, schemas, or guides needed dynamically. They must not fill context with non-actionable material.
• assets/: Use for boilerplate or templates the agent copies.

Step 3: Scaffold

For a new skill, run the init script:

python scripts/init_skill.py <skill-name> --path <output-dir>
# Add only the resource directories the workflow needs:
python scripts/init_skill.py <skill-name> --path <output-dir> --resources references

Request only justified resource directories. Use --examples only when placeholders clarify a real need, then replace or remove them. For an existing skill, edit in place; do not re-scaffold.

Step 4: Author

SKILL.md — Frontmatter

Start with the required fields:

---
name: my-skill                  # lowercase, hyphens, max 64 chars, matches folder name
description: "Single coherent paragraph covering what it does + when to use it; max 1024 chars."
---

Add optional fields only when they change use or distribution. Use compatibility for non-obvious OS, package, network, or tool requirements; most skills do not need it.

Description rules: The description is the primary routing signal. Front-load the capability and natural task context so matching survives hosts that shorten discovery metadata. Add a near-miss boundary only when it prevents likely misrouting. Avoid implementation details, catchalls, exhaustive synonym lists, and exact wording copied from failed test prompts.

SKILL.md — Body

State the desired outcome, non-obvious constraints, decision criteria, and verification. Explain operational intent so agents can generalize. Use fixed sequences only where deviation causes a concrete failure. Link each resource where it becomes relevant, or in a compact Resources section, and state exactly when to read, run, or use it. Omit the section when the skill has no resources.

Scripts (scripts/)

• Bundle only repeated or deterministic logic. Make inputs, outputs, dependencies, and failures explicit.
• Execute every new or changed script against representative input.

References (references/)

• Reference files must extend the skill for a specific subtask.
• Do not use them as a README, training manual, or catalog.
• Add a table of contents at the top of any file over 100 lines.
• Never duplicate content between SKILL.md and a reference file.

What NOT to Include

Do not create: README.md, CHANGELOG.md, INSTALLATION_GUIDE.md. Evict any file that doesn't direct agent behavior.

Step 5: Pressure-Test Behavior

Choose evaluation depth in proportion to the change. A small or subjective edit may need one clean-context scenario and qualitative review. For substantial, risky, or objectively verifiable work, use 2–3 realistic prompts with expected and forbidden behavior, then compare the candidate with the pre-edit or no-skill baseline under the same conditions. See references/pressure-testing-skills.md.

Use references/skill-triggering-tests.md when activation may be too broad or too narrow. Do not test only whether the agent can repeat the skill text.

Step 6: Validate and Package

From the target repository, resolve these scripts relative to this skill:

python <skill-creator-dir>/scripts/quick_validate.py <skill-dir>
python <skill-creator-dir>/scripts/sweep_skills.py <skill-dir>
python <skill-creator-dir>/scripts/check_changed_files.py

Fix validation errors, resolve placeholder findings, and triage workstation-path hits. The sweep exit status enforces broken links; its other findings are report-only. These checks prove structure and hygiene, not behavior. Package with package_skill.py only when a distributable archive is requested.

Step 7: Iterate

After real usage:

• If it under- or over-triggers, revise the underlying intent or boundary and rerun the trigger matrix.
• If a section doesn't improve output, remove it.
• If it grows too large, push depth into references/.
• If several runs recreate the same helper logic, consider bundling it in scripts/.
• Review execution traces as well as final outputs; remove instructions that cause repeated unproductive work.
• Generalize patches: fix the underlying instruction gap, not just the single failing prompt.

Skill Naming Conventions

• Lowercase letters, digits, and hyphens only (e.g., pdf-extractor).
• Max 64 characters; no leading/trailing/consecutive hyphens.
• Folder name must match the name field exactly.

Reference Files

• See references/patterns.md for progressive disclosure patterns and anti-patterns.
• See references/spec.md for the full AgentSkills frontmatter field reference.
• See references/pressure-testing-skills.md for proportional clean-context behavior tests and baselines.
• See references/skill-triggering-tests.md for natural-prompt activation and description regression checks.

Scripts