Custom Skill Builder

Custom Skill Builder

Role

Build skills that agents can actually use: small, specific, triggerable, validated, and easy to route. This skill adapts the repository's skill format for user-defined skills while preserving the core skill-creator principles: concise body, clear trigger metadata, progressive disclosure, deterministic resources when needed, and validation before handoff.

Start By

1. Read references/skill-authoring-contract.md.
2. Clarify the skill intent in one sentence: what recurring job should this skill make easier?
3. Decide whether this is a new skill, an update to an existing skill, or a project-local overlay.
4. Check for overlap with existing skills. Extend an existing skill when the new behavior is a narrow variant; create a new skill when it has a distinct trigger, workflow, or ownership boundary.
5. If the skill depends on current libraries, APIs, CLIs, MCP servers, or platform behavior, verify those docs with Context7 MCP before writing procedural guidance.

Procedure

1. Define trigger metadata. Write name and description so an agent can decide when to load the skill. The description must include when to use it.
2. Choose scope. Keep the skill focused on one repeatable job. Move detailed tables, examples, schemas, and provider-specific variants into references/.
3. Create structure. Use SKILL.md, agents/openai.yaml, and optional references/, scripts/, or assets/ only when they directly support the skill.
4. Write the workflow. Include Role, Start By, Procedure, Principal-Level Defaults, Output Artifacts, Quality Bar, Handoff, and References.
5. Add metadata. Create agents/openai.yaml with display name, short description, default prompt, and Context7 MCP dependency when the skill can touch external technology.
6. Add routing. Register the skill in agentic/routing/skills.json and add it to entrypoints only where it is genuinely useful.
7. Validate. Run python3 agentic/scripts/validate.py, check links, check placeholder markers, and run install dry-run when installer behavior changes.
8. Forward-test. Try at least one realistic prompt the skill should handle and one prompt it should not capture. Adjust trigger wording if it is too broad.

Principal-Level Defaults

• Follow ../../routing/principal-operating-model.md before adding a skill that changes routing, permissions, tool access, or memory behavior.
• Use Context7 MCP for current external technical guidance included in the skill or its references.
• Prefer small skills with strong trigger descriptions over large skills that try to cover a whole discipline.
• Do not copy long external docs into a skill. Link primary sources and summarize only the operational rules the agent needs.
• Do not create auxiliary files like skill-level README, install guides, or changelogs unless the repository validator explicitly requires them.

Output Artifacts

• New or updated skill folder
• SKILL.md with complete frontmatter and required sections
• agents/openai.yaml metadata
• Reference files, scripts, or assets only when justified
• Routing update in agentic/routing/skills.json
• Documentation update when user-facing behavior changes
• Validation and forward-test evidence

Quality Bar

• Skill name is lowercase hyphen-case and matches its folder.
• Description is specific enough to avoid accidental activation.
• SKILL.md stays lean; detailed material goes into directly linked references.
• Every reference named in SKILL.md exists.
• No unresolved placeholder markers remain.
• The skill passes repository validation and does not introduce machine-specific paths.
• The final handoff tells the user how to invoke the skill and where it is installed.

Handoff

Hand off with: skill name, trigger summary, files created or changed, routing impact, validation result, and one example prompt. If the skill came from a failed agent behavior, hand off also to self-improvement-loop for regression tracking.

References

• references/skill-authoring-contract.md: compact rules for skill scope, anatomy, progressive disclosure, metadata, validation, and forward testing.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ashermahonin
• Source: ashermahonin/agentic-skills
• License: MIT

Install and usage instructions live in the source repository linked above.