Claude subagent creator

Create well-scoped custom subagents Claude Code can auto-delegate to. A subagent is a Markdown file with YAML frontmatter plus a system-prompt body; it runs in its own context window with its own tool surface.

When to use

• The user wants a reusable persona (reviewer, explorer, planner, scoped worker).
• A workflow recurs often enough to deserve a named role fan-out can target.
• You need to pin a model or restrict tools for a class of delegated work.

For a one-off delegation, don't author a file — just use fan-out's inline spawn contract.

Agent file anatomy

./.claude/agents/<name>.md (project) or ~/.claude/agents/<name>.md (user):

---
name: <hyphen-case, matches filename>
description: <what it does + WHEN to delegate to it — explicit triggers; this drives auto-selection>
model: opus            # or sonnet / haiku / omit to inherit the session model
tools: Read, Grep, Glob, Bash   # smallest surface that works; omit = inherit all
---
<system prompt: role, exact scope, output contract, hard boundaries>

Authoring workflow

1. Name + intent — hyphen-case name; one clear job. Avoid catch-alls.
2. Description for delegation — lead with the action, then Use when ... with concrete triggers

("Use when reviewing a diff for security issues"). The main agent matches this to decide delegation; vague descriptions don't get auto-selected.

1. Model policy — opus for reasoning/review/debugging/planning; sonnet/haiku or inherit for

read-heavy scans and inventories. Omit model: to inherit the session model.

1. Tool scoping — grant the minimum. Default read-only (Read, Grep, Glob, Bash); add

Edit, Write only for workers that own a disjoint write surface.

1. System prompt — state the role, the exact scope, the return contract (status, evidence with

file:line, commands run, findings + confidence, risks), and hard boundaries ("never edit outside your assigned files"; "you inherit no parent transcript — work only from this prompt").

1. Install + smoke test — write the file, then confirm via /agents (or the Task tool with

subagent_type: <name>) that it loads and a trivial scoped task returns the contracted shape.

Description-for-delegation rules

• Action verb first; then explicit "Use when" triggers and the keywords a request would contain.
• Name the boundary ("read-only", "does not commit") so the parent scopes it correctly.
• One responsibility per agent — overlapping agents compete and dilute auto-selection.

Scope

• Project (./.claude/agents/) — committed, shared with the repo, available in that project.
• User (~/.claude/agents/) — personal, available across all your projects.
• Project agents shadow user agents of the same name.

Failure modes

Use fan-out to delegate to the roles you create here.