Create A Skill

Create-a-Skill — Author & Improve Repo Skills

End-to-end skill authoring for skills-i-use: collect → challenge → review → write → validate → register in repo.

Works with any coding agent that can run subagents. Follows CONTRIBUTING.md.

Supporting files:

• questions.md — intake and hard questions
• reviewers.md — Skeptic / Maintainer / Practitioner subagents + synthesis
• design-brief-template.md — pre-build design doc
• templates/ — SKILL.md and README starters
• scripts/validate-skill.sh — convention checker

Scripts: read scripts/validate-skill.sh before running. Double-check paths and flags every time.

Output: {SKILL_OUTPUT_DIR}/create-a-skill/ — see ../OUTPUT.md

Step 0 — Mode

Locate repo root: skills-i-use/ (this repo) or ask user for path.

Step 1 — Intake (hard questions)

Read questions.md. Ask required + hard questions in one message.

Do not proceed until required items are answered. Push back on weak answers — use follow-ups from the red-flag table.

Scan existing skills before intake completes:

ls skills/

For each existing skill, read description in frontmatter (first 20 lines of SKILL.md). Note overlap in the brief.

If user provided verbatim wording for the skill, preserve it exactly in output files.

Step 2 — Build Skill Brief

Compile answers into:

WORKFLOW: [repeatable steps]
FREQUENCY: [how often used]
TRIGGERS: [phrases]
OUTPUTS: [artifacts + skill-outputs path]
STOP_CONDITION: [done when]
MODE: [new | improve]
TARGET: [skill name or existing folder]
OVERLAP: [existing skills]
OUT_OF_SCOPE_V1: [explicit]
TRUST_BOUNDARY: [commit/push/delete/spend — yes/no]
VERBATIM: [any exact copy from user]

Step 3 — Reviewer Subagents (parallel)

Read reviewers.md. Run Skeptic, Maintainer, and Practitioner in parallel with the skill brief.

Then run Synthesis (fourth subagent or orchestrator) → DECISION: BUILD NEW | EXTEND | REJECT.

If REJECT — explain why and suggest alternative (script, rule, extend X). Stop.

If EXTEND — read existing skill folder fully before writing. Prefer minimal diff.

Step 4 — Design Brief

Fill design-brief-template.md.

Path: {output_dir}/design-YYYY-MM-DD-[skill-name].md

Present brief to user. Ask: "Proceed to write files?" — stop on no.

Step 5 — Write Skill Files

New skill

Create skills/[skill-name]/ using templates/ as starters. Replace {{placeholders}}.

Typical layout per CONTRIBUTING.md:

skills/[skill-name]/
├── SKILL.md
├── README.md
├── [principles.md | scripts/ | templates/ as needed]

Improve existing

1. Read entire existing folder
2. Apply changes from brief — do not duplicate a new folder unless synthesis says split
3. Preserve working triggers unless intentionally breaking (note in brief)
4. Extract bloat from SKILL.md into new supporting files if over ~250 lines

Rules while writing:

• Agent-agnostic — subagents/shell, not vendor APIs
• Output → {SKILL_OUTPUT_DIR}/[skill-name]/
• Safe defaults — no auto-commit/push
• Third-person description with what + when

Step 6 — Validate

bash skills/create-a-skill/scripts/validate-skill.sh skills/[skill-name]

Fix all FAIL| items. Address WARN| items or document why waived in design brief.

Step 7 — Register in Repo

Update registry files (skip if improving and row already exists):

1. README.md — add row to skills table (alphabetical or logical order)
2. skills/OUTPUT.md — add subfolder to layout tree if skill produces output
3. Cross-skill section in OUTPUT.md — if skill reads/writes another skill's index

Update {output_dir}/index.md:

# Create-a-Skill Designs

| Date | Skill | Decision | Brief |
|------|-------|----------|-------|

Step 8 — Ship (user confirms)

1. List all created/changed files
2. Show validation result
3. Suggest @ship-check if scripts were added
4. Ask: "Commit to repo?" — only on explicit yes

git add skills/[skill-name]/ README.md skills/OUTPUT.md
git commit -m "feat: add [skill-name] skill"

Ask separately before push. Never push without explicit request.

Edge Cases

• Skill outside skills-i-use — write files where user points; skip Step 7 registry
• Meta: improving create-a-skill itself — same workflow; extra care on regressions
• User says "just write it" — still run validate; abbreviated intake OK if context is rich
• Duplicate skill name — fail validation; pick new name with user
• Subagents unavailable — orchestrator plays all three reviewers sequentially

Invocation Examples

@create-a-skill I want a skill for retrieval benchmarks
improve ship-check — add Python project support
create a skill from this conversation
new skill: legal contract review with persona debate
author a skill, ask me hard questions first