| name | create-skill |
| description | Guide creation of a new GAAI skill following the agentskills.io spec and GAAI best practices. Activate when adding a new skill to the .gaai/core/skills/ catalog. |
| license | ELv2 |
| compatibility | Works with any filesystem-based AI coding agent |
| metadata | {"author":"gaai-framework","version":"1.0","category":"cross","track":"cross-cutting","id":"SKILL-CRS-016","updated_at":"2026-02-26T00:00:00.000Z","status":"stable"} |
| inputs | [{"skill_intent":"description of what the skill should do"},{"skill_category":"discovery|delivery|cross"},{"skill_track":"discovery|delivery|cross-cutting"},{"existing_skills":"list of current skills to check for overlap"}] |
| outputs | [".gaai/core/skills/{category}/{skill-name}/SKILL.md","updated .gaai/core/skills/README.skills.md entry"] |
Create Skill
Purpose / When to Activate
Activate when:
- A new capability is needed that no existing skill covers
- A skill is being extracted from agent logic that has become reusable
- Extending a forked GAAI installation with project-specific skills
This skill is self-referential: it uses the agentskills.io spec to author a spec-compliant skill.
Prerequisites
Before activating:
- Confirm no existing skill already covers the intended capability (check
README.skills.md)
- Confirm the new skill is a pure execution unit — it executes, it does not reason
- Confirm the skill belongs to exactly one category (discovery / delivery / cross)
- Have a clear, one-sentence description of: WHAT it does and WHEN to activate it
If a skill "appears to think" in the design, it is wrongly scoped. Redesign or split.
Process
Step 1 — Name and classify
Derive the skill name from the action it performs:
- Use lowercase kebab-case
- Name should be a verb phrase:
generate-stories, qa-review, memory-compact
- Category: discovery (produces artefacts) | delivery (implements/validates) | cross (shared utility)
Verify name does not conflict with an existing skill in README.skills.md.
Step 2 — Define inputs and outputs explicitly
List every input the skill requires. Be specific:
- What file paths does it read?
- What structured data does it receive from the invoking agent?
- What context is assumed to be present?
List every output the skill produces:
- File paths created or modified
- Structured data returned to the agent
- Side effects (backlog state changes, memory updates)
If inputs or outputs are unclear, the skill is not ready to be written. Clarify first.
Step 3 — Write the SKILL.md frontmatter
---
name: {skill-name}
description: {one sentence}
license: ELv2
compatibility: Works with any filesystem-based AI coding agent
metadata:
author: {author}
version: "1.0"
category: {discovery|delivery|cross}
track: {discovery|delivery|cross-cutting}
id: SKILL-{CAT}-{NNN}
updated_at: {YYYY-MM-DD}
status: stable|experimental
inputs:
- {input_1}: {description}
- {input_2}: {description}
outputs:
- {output_path_or_description}
---
Description rule: Must be one sentence, ≤ 1024 characters. Format:
"{Verb phrase that describes the output} when {activation condition}."
Example: "Generate Stories from an Epic with testable acceptance criteria when Discovery is refining a new feature scope."
Step 4 — Write the skill body
Required sections:
# {Skill Title}
## Purpose / When to Activate
When and why an agent should invoke this skill. Be explicit.
## Process
Numbered steps. Each step produces a specific output or advances the state.
Steps must be deterministic: same inputs → same outputs.
## Quality Checks
What must be true for the output to be acceptable.
List as assertions, not descriptions.
## Outputs
Exact file paths or structured outputs produced.
## Non-Goals
What this skill must NOT do. Prevents scope creep.
Add a final line if applicable:
No silent assumptions. Every {X} becomes explicit and governed.
Step 5 — Create directory and file
.gaai/core/skills/{category}/{skill-name}/SKILL.md
SKILL.md filename is always uppercase (spec requirement).
Optional subdirectories (create only if needed):
references/ — supporting documents the skill references
assets/ — templates or static files the skill produces
Step 6 — Update ALL relevant indices (MANDATORY)
This step is non-negotiable. A skill that is not indexed is invisible to agents.
6a. Skills index YAML:
- Invoke
build-skills-index to regenerate both .gaai/core/skills/skills-index.yaml (core) and .gaai/project/skills/skills-index.yaml (project)
- Verify the new skill appears in the generated index with correct
id, name, path, and tags
6b. Skills README:
- Core skills: update
.gaai/core/skills/README.skills.md — add the skill to the appropriate category section
- Project skills: update
.gaai/project/skills/README.md — add the skill to the appropriate section
6c. Domain index (if domain-scoped skill):
- If the skill belongs to a domain (e.g.,
domains/content-production/), update the domain's index.md in memory:
.gaai/project/contexts/memory/domains/{domain}/index.md — add a row to the skills table
- This ensures domain sub-agents discover the skill when loading their domain context
Failure to update any of these indices means the skill is effectively invisible — agents cannot discover it, and it will not be loaded during delivery or discovery sessions.
Step 7 — Reference in agent file
If the skill is intended for a specific agent (Discovery) or a delivery sub-agent, add it to the agent's skill list in:
.gaai/core/agents/discovery.agent.md — Skills Used section
- Or the relevant sub-agent file in
.gaai/core/agents/sub-agents/ (planning, impl, qa)
- Or neither, if it's a general cross skill invoked opportunistically
Quality Checks
- Skill name matches directory name exactly
SKILL.md frontmatter has name and description at top level (spec requirement)
- Description is ≤ 1024 characters, one sentence
- All inputs are explicitly declared
- All outputs are explicitly declared with file paths where applicable
- Skill body has all 5 required sections
- Non-Goals section explicitly prevents scope creep
README.skills.md updated
health-check.sh passes after creation
Non-Goals
This skill must NOT:
- Design agent behavior or orchestration logic (that is an agent concern)
- Make decisions about when to activate the skill (the invoking agent decides)
- Merge or replace an existing skill (deprecation requires explicit human decision)
A skill that appears to think is wrongly designed. Redesign before authoring.