Agents Claude Creator

Claude Agent Creator

Build precise, focused, production-grade Claude agents without re-researching the platform. The default product is a Claude Code subagent: a Markdown file with YAML frontmatter plus a system-prompt body, stored under .claude/agents/ (project) or ~/.claude/agents/ (personal), that Claude automatically delegates to when a task matches its description. The subagent runs in its own fresh context window, with its own tools and model, and returns only a summary.

Scope: this skill creates the agents that do work. For repo instruction files (AGENTS.md/CLAUDE.md) use agent-md-creator; for portable Agent Skills use skill-creator; for OpenCode agents use opencode-agent-creator. A subagent often preloads a Skill to get its methodology — the two compose via the skills: field. For a cloud/programmatic agent on the Claude API, see references/managed-agents.md.

The five things that define an agent

Internalize this — every decision below maps to one of these:

1. Identity — name (lowercase-hyphens; how Claude and @-mentions address it).
2. Routing — description (third person; the trigger that makes Claude delegate).
3. Capability boundary — tools / disallowedTools / mcpServers / skills (what it can touch).
4. Brain — model (cost vs. capability: haiku / sonnet / opus / fable / inherit).
5. Behavior — the Markdown body (the system prompt: role, workflow, output contract).

The one constraint that shapes every prompt

A subagent starts cold. It does not see the conversation history, the files Claude already read, or the skills already invoked. Its entire world at startup is: its own system-prompt body, the one-shot delegation message Claude writes, the CLAUDE.md hierarchy, a git-status snapshot, and any skills: you preload. Any context the agent needs must be in the body, in the description, preloaded as a skill, or restated in the task you delegate — otherwise it does not cross. (The built-in Explore and Plan agents skip even CLAUDE.md.) Design as if briefing a brand-new contractor who can read the repo but was in none of your meetings.

Workflow

1. Is a subagent even the right tool?

Do not recreate the built-ins: Explore (read-only codebase search, Haiku), Plan (read-only research), general-purpose (full tools). Create a custom subagent only when a focused role, tool boundary, or model choice beats those.

2. Scope to one responsibility

One agent, one bounded job, one deliverable, one clear hand-off. If you can't state its success signal in a sentence, split it. Single-responsibility agents route more reliably (the description stays sharp) and stay debuggable.

Infer; don't interrogate. Derive the role, tools, and model from the request plus sensible defaults, state the assumptions you made, and build. Ask the user only when a choice genuinely changes the build and can't be inferred (e.g. project vs. personal scope when it matters, or a required secret). A one-line assumption the user can correct beats a questionnaire.

3. Choose scope (location) and name

Higher-priority scope wins on name collisions (managed → CLI --agents → project → user → plugin). Identity comes from the name field, not the filename; keep name unique across the tree. Subfolders are allowed for organization and don't change the name.

4. Write the frontmatter (least privilege, right brain)

Minimum viable agent — only name and description are required:

---
name: code-reviewer
description: Expert code-review specialist. Proactively reviews changes for quality, security, and maintainability. Use immediately after writing or modifying code.
tools: Read, Grep, Glob, Bash
model: inherit
---

• description is the router. Write it in third person, name the concrete trigger, and state what it does. Add "use proactively" / "use immediately after X" when you want Claude to delegate without being asked. Vague (Helps with code) → never triggers. (Full guidance + good/bad examples: references/system-prompt-patterns.md.)
• Tools = least privilege. Omitting tools inherits ALL tools — rarely what you want. Use an allowlist (tools: Read, Grep, Glob) or a denylist (disallowedTools: Write, Edit). A read-only reviewer/researcher must exclude Write and Edit; a fixer needs Edit.
• Model = cost routing. haiku for fast/cheap/bulk search & summarizing; sonnet for balanced analysis; opus for hard reasoning; inherit (default) to match the main session. Route cheap work to cheap models.
• skills: preloads the full content of named skills into the agent at startup — the cleanest way to hand a cold agent your methodology/conventions.
• Optional dials: memory (cross-session learning), permissionMode, mcpServers, hooks, color, effort, isolation: worktree. Every field, precedence, and model-resolution order: references/frontmatter-reference.md.

5. Write the system-prompt body

The body is the system prompt. Proven shape:

You are a <role> specializing in <domain>.

When invoked:
1. <first concrete step — often "run git diff" / "read the failing test">
2. ...
3. <produce the deliverable>

<Checklist or key practices the agent must apply>

<Output contract: exactly how to format the result that returns to the caller>

<Focus rule: the one thing to optimize / not do>

Match degrees of freedom to the task: tight numbered steps for fragile/destructive work (migrations, batch edits); high-level direction for open-ended judgement (code review). Because the agent is cold, put any non-obvious repo fact, convention, or constraint in the body. Worked examples and the full pattern: references/system-prompt-patterns.md. Copy-ready skeleton: assets/agent-template.md.

6. Validate and test

• Frontmatter is valid YAML between ---; name is lowercase-hyphens; tools names are real; model is a valid alias/ID/inherit.
• Manual file edits load at session start — restart the session. Use --agents JSON flag for session-only testing without a file (as of v2.1.198, /agents no longer opens a creation wizard).
• Triggering test: give a natural prompt that should delegate and confirm Claude picks the agent; if not, sharpen the description. Behavior test: run it on a real task and confirm it stays in its tool boundary and returns the contracted output. Method: references/triggering-and-testing.md.

For teams of agents (a coordinator that delegates to specialists), keep each specialist single-responsibility and let the coordinator route — see the orchestration notes in references/system-prompt-patterns.md.

Anti-patterns

• Omitting tools "to be safe" → inherits everything, including Write/Edit/Bash on a "reviewer". Always set the boundary.
• Assuming the agent sees the chat → it starts cold. Restate context, or preload a skill.
• A vague description → Claude never delegates. Name the trigger; add "use proactively".
• One mega-agent doing review + fix + test + deploy → split by responsibility.
• Recreating Explore/Plan/general-purpose → use the built-ins.
• bypassPermissions by default → only when truly justified; prefer tools/disallowedTools + hooks.
• Premium model for bulk work → route summarizing/search to haiku.

Resources

• references/frontmatter-reference.md — every frontmatter field (tools, disallowedTools, model, skills, memory, permissionMode, mcpServers, hooks, effort, isolation, color, initialPrompt, maxTurns), scopes & precedence, model-resolution order, tool-inheritance rules, what loads at startup. Load when writing or debugging frontmatter.
• references/system-prompt-patterns.md — how to write the body and the description: role + "When invoked" workflow + checklist + output contract, degrees-of-freedom, cold-context briefing, orchestration/teams, plus annotated example agents. Load when writing prompts.
• references/managed-agents.md — Claude Managed Agents (cloud harness): agent = model + system prompt + tools + MCP + skills; when to use vs. a Code subagent; create-agent/session essentials and the beta header. Load when the user wants a programmatic/cloud agent.
• references/triggering-and-testing.md — triggering tests (does it delegate?) and behavior/pressure tests (does it stay in bounds and deliver?). Load before finalizing.
• assets/agent-template.md — fill-in subagent skeleton (frontmatter + body).
• assets/examples/ — ready-to-adapt agents: code-reviewer.md (read-only), debugger.md (fix-capable), safe-researcher.md (restricted tools), test-runner.md (isolates verbose output). Read for concrete patterns.