make-agent

name	make-agent
description	Creates new repo-native agent profiles from plan to scaffolded `agents/<name>.md` and catalog alignment. Use when adding an agent profile to this repository.
user-invocable	true
disable-model-invocation	true

Make Agent

Core Contract

Create one new agent profile in this repository as a single markdown file under agents/. Start with a brief plan and explicit assumptions. Follow CLAUDE.md / AGENTS.md on conflict.

Agent Profile Contract

An agent profile is one .md file whose body is the subagent system prompt.

Location contract:

• Preferred path: agents/<name>.md.
• Optional grouping directories are cosmetic (agents/<group>/<name>.md).
• Frontmatter name: is the identity and must stay repo-unique.

Required frontmatter:

name: agent-name
description: Third-person summary ending with "Delegate when ..."
tools: Read, Grep, Glob, Bash
model: inherit

• name: required; lowercase-hyphenated; repo-unique; should match filename.
• description: required; third person; ends with Delegate when ....
• tools: optional; keep least-privilege; omit to inherit all.
• model: optional; inherit | sonnet | haiku | opus.

Body contract:

• Imperative, single-role, and self-contained instructions.
• Include stop-and-ask guidance when ambiguity affects behavior or safety.
• Include the override rule: target repo CLAUDE.md/AGENTS.md wins on conflict.

Required Inputs

1. Agent purpose and boundaries.
2. Proposed name (if any).
3. Tooling constraints / least-privilege requirements.
4. Model preference (if any).
5. Expected outputs or reporting style.

Workflow

1. Restate request and produce a compact creation plan.
2. List assumptions (safe default vs needs confirmation).
3. Ask only high-impact clarifying questions.
4. Normalize/derive agent name and validate repo uniqueness.
5. Create agents/<name>.md with compliant frontmatter and prompt body.
6. Ensure prompt includes stop-and-ask gates and override rule.
7. Add a concise README.md catalog entry under the Agents section.
8. Verify naming alignment (name, filename, and references) and catalog sync.

Stop and ask if request spans multiple agents, name conflicts with existing tools/agents, or ambiguity changes behavior/safety.

Safety Rules

• Never scaffold an agent before name uniqueness is confirmed in this repo.
• Never grant more tools than required for the agent's role.
• Never hide assumptions; label and confirm high-impact ones.
• Never skip clarifications when ambiguity affects behavior, safety, or permissions.
• Never silently broaden scope beyond one requested agent without user approval.
• Never leave README.md out of sync after adding, renaming, or removing an agent.

Output Style

Report final agent name/path, assumptions confirmed, defaults applied, and any optional follow-up improvements.