Make Skill

Create one new workspace Skill from the raw current conversation. Use the scripts in this directory for the lifecycle below; do not rely on a make-skill-specific core tool, mode, or state store.

Run each documented python scripts/... command through execute_shell_command, setting cwd to this Skill's <dir> from the available-skills entry; every script reads one JSON object from stdin (or --input <file>) and writes one JSON object to stdout.

Plan

The focus in /make-skill <focus> is required. For a natural-language request, infer it from the request and current conversation. Later user corrections replace conflicting earlier rules. Preserve stable guidance, contracts, templates, and workflows that should change future behavior; exclude one-off data, temporary paths, secrets, and retry chatter.

Read primary type and package, then choose one primary type and only the files it needs. When the proposed test mode is not off, read behavior testing before defining its target.

Batch workflows

A stored batch is a parameterized run_tool_batch program bundled with a workflow Skill. Use batch: true when a reusable region's actions, branches, and success condition can be stated before execution and one stored entrypoint saves meaningful agent-tool round trips. The region may be the whole workflow, one substantial helper, or one semantic tool-native action; action count is not the criterion. Runtime data, observations, and a final agent review do not prevent batching when the rule for handling them is already known.

Use batch: false only when execution must invent the next action or success condition at runtime, or a shared entrypoint has no practical reuse value. If the user explicitly requests Batch, treat it as an approved plan refinement and revise the plan without reopening eligibility.

Only after selecting batch: true, read run batch before finalizing the workflow and file tree. When batch: false, do not read it.

Planning is read-only except for running python scripts/create_plan.py: use conversation evidence and existing artifacts, but do not execute or probe the proposed workflow, create files, or initialize a draft. Pass the candidate through stdin:

{
  "revision": 1,
  "focus": "One-sentence extraction scope",
  "name": "lowercase-hyphen-name",
  "goal": "Outcome for a future agent",
  "type": "workflow",
  "batch": true,
  "steps": ["A user-reviewable workflow step"],
  "package": ["SKILL.md", "scripts/run.batch.json"],
  "execution": "foreground",
  "test": {"mode": "off", "target": ""},
  "warnings": []
}

Render the normalized plan in English and show the selected value together with every available choice so the user can revise it without knowing the schema. The user-visible plan must contain this compact options table; do not replace it with prose or an approval hint. Omit the Batch row for a non-workflow:

Also show the name, goal, workflow, complete file tree, test target when applicable, and warnings. Pass the internal values instruction/template/workflow, true/false, foreground/background, and off/smoke/eval to the script. Do not invent a full enum or any choice outside the script schema. Do not show a Batch closing reason, schema, revision, or internal enum. Ask the user to approve, modify, or cancel, then end the response without further tool calls. Continue in a later turn only after a new user message clearly approves the displayed revision.

• After a modification, merge the latest feedback, increment revision, rerun create_plan.py, and show the complete returned plan.
• Stop on cancellation; ask one short confirmation for an ambiguous reply.
• Do not ask separately about execution or testing.

This version creates new Skills only. Resolve a name conflict through a newly approved revision; never overwrite an existing Skill.

Build

After approval, run python scripts/init_draft.py with JSON through stdin:

{"workspace": "/workspace/path", "plan": {"...": "complete normalized plan"}}

Create only approved files under the returned skill_dir. Start the generated SKILL.md with valid frontmatter:

---
name: lowercase-hyphen-name
description: Briefly state the capability and when to use it.
---

Keep the body to essential procedure and constraints without repeating the description. Type metadata is unnecessary.

Before validation, read the package from the perspective of a future agent that cannot see the source conversation. Remove references to source task directories, prior outputs, temporary IDs, current-case examples, or make-skill draft/publish language unless that resource is deliberately packaged and reusable. When adapting an existing helper, generalize its paths, docstrings, and reporting, and check that its implementation still matches the final reusable rules. Keep this as one authoring pass; do not add case-specific lifecycle checks.

For background execution, give the generic subagent the complete approved plan, latest corrections, draft path, and the remaining gates; it must not request approval again.

Validate, test, and publish

Before executing any draft script or batch, run python scripts/validate_skill.py with JSON through stdin:

{"workspace": "/workspace/path", "draft_id": "returned-draft-id"}

Fix reported static or security errors in the draft and validate again. Testing is independent of Batch: run exactly the approved behavior test according to behavior testing, and let off perform no draft execution. When a test or Batch run fails, retain the draft, report the concrete error, revise the Skill if the correction is clear, then validate again; do not hide the failure behind a fallback.

Publish the unchanged validated draft by running python scripts/publish_skill.py with JSON through stdin:

{"workspace": "/workspace/path", "draft_id": "returned-draft-id", "expected_digest": "digest-from-validation"}

On success, report the package tree, validation summary, test result when one ran, and invocation /<name>. On conflict or failure, retain the draft and report the error. Publishing a Skill is already persistent; do not also write it to MEMORY.md or daily memory unless the user separately asks.