create-skill

type	Skill
name	Create Skill
category	evolution
description	Generate a complete new skill from a one-line prompt and ship it as a PR
var
tags	["dev","meta"]

${var} — A natural-language description of the skill to create. Required. Example: "monitor Hacker News for AI papers and send a summary" or "track gas prices on Ethereum and alert when below 10 gwei".

If ${var} is empty, exit CREATE_SKILL_NO_VAR:

./notify "create-skill aborted: var empty — pass a description e.g. \"monitor X for Y\""

Then stop.

Today is ${today}. Your task is to generate a complete, production-ready skill from ${var}, score it against a quality bar, and ship it as a PR — never commit directly to main.

Steps

1. Parse the request. Extract from ${var}:

   • Core action verb (monitor, fetch, generate, analyze, alert, track, scan, etc.)
   • Data source(s) — APIs, websites, RSS, on-chain, GitHub, etc.
   • Output format — notification, article, file, PR, dashboard, etc.
   • Configurable parameter(s) the new skill will accept via its own ${var}
   • Suggested cadence (daily, hourly, weekly, on-demand)

   Save a one-paragraph structured request summary; you'll use it in the PR body.

2. Duplicate detection (deep — not just ls). Find functional overlap, not just name collision.

   keywords=$(echo "${var}" | tr '[:upper:]' '[:lower:]' | grep -oE '[a-z]{4,}' \
     | grep -vE '^(send|with|from|that|this|when|each|into|over|some|like|just|than|then|also|will|have|been|using|monitor|track|fetch|alert)$' \
     | sort -u)
   for kw in $keywords; do
     grep -liE "$kw" skills/*/SKILL.md | head -5
   done
   

   Read the top 3 candidates fully. For each, judge: does it already do this? Could the request be solved by running an existing skill with a different var=?

   • Near-duplicate exists → exit CREATE_SKILL_DUPLICATE. Notify with the existing skill name and a one-line suggestion ("use existing {skill} with var={...} instead"). Stop.
   • Functionally adjacent → design the new skill to complement (different angle/cadence/output). Document the boundary in the PR body.

3. Research the data sources. For every API or data source the new skill needs:

   • WebSearch for the current API documentation. Cross-check against a secondary source when feasible (a recent GitHub repo using it, an official changelog, or a Stack Overflow answer dated ≥2026) to confirm the endpoint isn't deprecated.

9b. Dry-run gate (blocks a broken generated skill from auto-merge). Before opening the PR, execute the new skill once with synthetic secrets, so a generated skill never reaches production having only ever run with real credentials: bash DRYRUN_VERDICT="output/.dry-run/$name.json" bash scripts/dry-run.sh run "$name" || true - The script self-checks the SKILL_DRYRUN repo variable (default on) and returns a skipped verdict when it is 0. - Read output/.dry-run/$name.json. passed: true (or skipped: true) means continue. passed: false means delete skills/$name/, revert the aeon.yml edit, and exit CREATE_SKILL_DRYRUN_FAILED with a notify listing the verdict reasons[]. Do not open the PR. - Put the verdict JSON in the PR body under a ## Dry-run section either way, so a reviewer sees the gate ran. The gate is structural (exit 0, non-empty output, no write outside the declared mode, no secret outside requires:), and no real credential is ever placed in the run's environment. It does not re-score content; the Haiku scorer already does that.

10. Open as a PR (never commit to main).

    name="{skill-name}"
    git checkout -b create-skill/$name
    git add skills/$name/SKILL.md aeon.yml
    git commit -m "create skill: $name
    
    {one-sentence description}
    
    Generated by create-skill from var: \"{request summary, ≤80 chars}\""
    git push -u origin create-skill/$name
    gh pr create --title "create skill: $name" --body "$(cat <<'EOF'
    ## Skill
    **Name**: `{skill-name}`
    **Description**: {description}
    **Tags**: {tags}
    **Schedule**: `{cron}` (disabled by default)
    **Model**: {model}
    **Var**: {var-doc}
    
    ## Request
    

    ${var}

    
    ## Sources researched
    - {URL 1}
    - {URL 2}
    - {URL 3}
    
    ## Required secrets
    {list of NEW_SECRET_REQUIRED env vars OR "None — uses existing secrets"}
    
    ## Quality scores
    | Criterion | Score |
    |-----------|-------|
    | Frontmatter | X/5 |
    | Var doc | X/5 |
    | API calls | X/5 |
    | Fallback behavior | X/5 |
    | Output spec | X/5 |
    | Network note | X/5 |
    
    ## Trigger manually
    Workflow dispatch with `skill={skill-name}` and `var={example-var}`.
    EOF
    )"
    

    Capture the PR URL.

11. Log. Append to memory/logs/${today}.md:

    ### create-skill
    - Request: {var, ≤80 chars}
    - Created: skills/{skill-name}/SKILL.md
    - Registered in aeon.yml: schedule={cron}, model={model}
    - Required secrets: {list or "none"}
    - Quality scores: F/V/A/Fb/O/N = X/X/X/X/X/X
    - PR: {url}
    - Exit: CREATE_SKILL_OK (or CREATE_SKILL_NEW_SECRET_REQUIRED)
    

12. Notify. Send via ./notify:

    *create-skill — {skill-name}*
    {one-line description}
    Schedule: `{cron}` (disabled by default)
    {Required secrets line if any}
    PR: {url}
    Trigger: dispatch skill=`{skill-name}` var=`{example}`
    

Exit taxonomy

Code	When	Action
CREATE_SKILL_OK	New skill created, validated, PR opened	Notify with PR link
CREATE_SKILL_NEW_SECRET_REQUIRED	Same as OK plus operator must add a new secret before enabling	Notify with PR link + secret call-out
CREATE_SKILL_NO_VAR	${var} empty	Notify abort reason; stop
CREATE_SKILL_DUPLICATE	Existing skill covers the request	Notify with existing-skill suggestion; stop
CREATE_SKILL_INSUFFICIENT_RESEARCH	Couldn't confirm ≥1 working data source after WebSearch + WebFetch	Notify with what was tried; stop
CREATE_SKILL_VALIDATION_FAILED	Quality enforcement or post-write checks failed	Delete partial files; revert aeon.yml; notify with failed criteria; stop
CREATE_SKILL_DRYRUN_FAILED	The dry-run gate (step 9b) returned passed: false	Delete partial files; revert aeon.yml; notify with verdict reasons; do NOT open the PR; stop

Network note

There is no network sandbox — curl works, with WebFetch as the fallback for a flaky public GET during research. For an auth'd API the new skill will call, route it through ./secretcurl with a {ENV_NAME} placeholder (the key injected via the skill's requires:), and gh api for GitHub. Irreversible side-effects (email, spend, on-chain writes, deploys) run in-run via ./secretcurl as the skill's final, fail-closed action — there is no deferred/postprocess step, and never defer a read (see CLAUDE.md).

Constraints

• Never commit a generated skill directly to main. Always open a PR.
• Never enable a generated skill in aeon.yml (enabled: false always — operator decides).
• Never add an API key/secret to the workflow that isn't already there. Surface as NEW_SECRET_REQUIRED and document in the PR body.
• Never ship a skill that fails validation. Aborting cleanly is always better than shipping broken.
• Never overwrite an existing skills/{name}/SKILL.md — name collisions are blocking errors.