create-agent

Scaffolds new sub-agent definitions as

.claude/agents/<category>/<name>.md

and automates the process up to quality checks.

Usage

/create-agent quality/lint-runner     # Create with specified category
/create-agent research/doc-fetcher    # Add to research category
/create-agent author/template-writer  # Add to author category
/create-agent                         # Interactive confirmation when arguments are omitted

Refer to

sample/agent.sample.md

included with this skill for the sub-agent template.

Category and Model Selection Guidelines

Category	Path	Role	Recommended Model
research/	`.claude/agents/research/`	Research, information gathering, WebFetch	`sonnet`
author/	`.claude/agents/author/`	Creation, editing (uses Edit/Write/Bash)	`sonnet`
quality/	`.claude/agents/quality/`	Review, linting, validation (read-only)	`sonnet` (inference/review) / `haiku` (mechanical linting only)

Detailed Model Criteria

Use Case	model
Mechanical, aggregation, linting, frontmatter validation	`haiku`
Reading comprehension, generation, review, research	`sonnet`
Complex design decisions, architecture	`opus`

Tools Least Privilege Principle

Agent Category	Allowed Tools
Read-only (research/・quality/)	Only `Read` , `Glob` , `Grep`
Creation/editing (author/)	Above + `Edit` , `Write` , `Bash`

Do not include

Edit

Write

, or

Bash

in read-only Agents.

Subagent Fallback (for skills add adopters)

The subagents delegated by this skill (

agent-author

・

frontmatter-linter

) assume the

.claude/agents/

definitions from the Fandhe-AI/agent-cli-skills repository. If the corresponding subagent does not exist in the target repository, do not delegate; instead, the main skill will directly execute the input, application rules, and required items described in the delegation prompt for each step to create the same artifact. Manual confirmation of the validation points in Step 4 is sufficient as an alternative to

frontmatter-linter

. Apply only the rules that exist in the target repository (such as

dotclaude-via-temp.md

); if they do not exist, you may create them directly in

.claude/agents/

Flow

Step 1: Confirm Category and Responsibilities

Parse

<category>/<agent-name>

from the arguments.

Verify that the category is one of
```
research/
```
,
```
author/
```
, or
```
quality/
```
If a category outside the above is specified, confirm with the user to select an appropriate category
Confirm the Agent's responsibilities (scope, scenarios for delegation) with the user

Step 2: Delegate to agent-author to Create File

Delegate to agent-author (subagent_type: agent-author) to create the file (if it does not exist, the main skill will create it directly according to "Subagent Fallback").

Important:

agent-author

follows

dotclaude-via-temp.md

and finalizes the placement via

_/dotclaude/agents/<category>/<name>.md

(does not write directly to

.claude/agents/

Example Delegation Prompt:

subagent_type: agent-author
prompt: |
  Purpose: Create a sub-agent definition for <category>/<agent-name>
  Input:
    - Agent Name: <agent-name>
    - Category: <category> (research/author/quality/)
    - Role & Responsibilities: <content received from user>
    - Template: sample/agent.sample.md included with this skill
    - Reference Existing Agents: Existing files under .claude/agents/
  Output Path: .claude/agents/<category>/<agent-name>.md
  Application Rules:
    - .claude/rules/agent-authoring.md (frontmatter, tools least privilege, category placement, body outline)
    - .claude/rules/dotclaude-via-temp.md (create via _/dotclaude/ and finalize with mv)
  Required Items:
    - frontmatter: name (kebab-case), description (specifically describe delegation scenarios),
                   model (follow selection criteria), tools (least privilege principle)
    - Body: # Agent Name → ## Role → ## Target Scope → ## Compliance Rules → ## Procedures/Perspectives →
            ## Completion Criteria → ## Report Format
    - Do not include Edit, Write, or Bash in read-only Agents

Step 3: Confirm Model Selection and Tools Configuration

Review the creation result from

agent-author

and verify the following:

Whether
```
model
```
is appropriately selected for the category and use case
Whether
```
tools
```
adhere to the least privilege principle
Whether
```
Edit
```
,
```
Write
```
, or
```
Bash
```
are not included for
```
quality/
```
category Agents

Step 4: Validate with frontmatter-linter

Delegate to frontmatter-linter (subagent_type: frontmatter-linter) for validation (if it does not exist, substitute with manual confirmation of the following points).

Validation Points:

Whether
```
name
```
is correctly set in kebab-case
Whether
```
model
```
is one of the specified values (haiku/sonnet/opus)
Whether
```
tools
```
do not include items exceeding least privilege
Whether
```
description
```
specifically explains delegation scenarios
Whether it is correctly placed in
```
.claude/agents/<category>/
```

If there are issues, send it back to

agent-author

for correction.

Step 5: Guide to Update Sub-agents Table in CLAUDE.md

After Agent addition is complete, guide the user with the following:

✅ Created .claude/agents/<category>/<agent-name>.md

Agent frontmatter summary:
  name: <agent-name>
  model: <model>
  tools: [<tool-list>]

To update the Sub-agents table and structure tree in CLAUDE.md:
  /update-docs
Please execute the above command.

To call this sub-agent within a skill, use:
  subagent_type: <agent-name>

Validation

Confirm that
```
.claude/agents/<category>/<agent-name>.md
```
exists in the correct path (since
```
agent-author
```
follows dotclaude-via-temp to execute everything from temporary creation in
```
_/dotclaude/agents/
```
to
```
mv
```
to
```
.claude/agents/
```
, there is no need for
```
create-agent
```
to execute
```
mv
```
separately)
Confirm that the
```
name:
```
field in frontmatches
```
<agent-name>
```
Confirm that the
```
tools
```
list complies with the category's least privilege policy
Confirm that it can be called with
```
subagent_type: <agent-name>
```

Notes

dotclaude-via-temp is mandatory:
```
agent-author
```
does not write directly to
```
.claude/agents/
```
, but executes the entire process from temporary creation in
```
_/dotclaude/agents/
```
to
```
mv
```
to
```
.claude/agents/
```
. There is no need for
```
create-agent
```
to execute
```
mv
```
separately (details:
```
.claude/rules/dotclaude-via-temp.md
```
)
rm -rf _/dotclaude
is prohibited: Only delete empty directories with
```
rmdir
```
(shared directory with other concurrent operations)
Tools for read-only Agents: Do not include
```
Edit
```
,
```
Write
```
, or
```
Bash
```
in
```
research/
```
and
```
quality/
```
categories
Name resolution: Calls via
```
subagent_type:
```
are resolved using the
```
name:
```
field in frontmatter. No changes to calling code are needed even if the category is moved
Execute update-docs: Always run
```
/update-docs
```
to update
```
CLAUDE.md
```
after adding an Agent

create-agent

NPX Install

Tags

SKILL.md Content (Chinese)

create-agent

Usage

Category and Model Selection Guidelines

Detailed Model Criteria

Tools Least Privilege Principle

Subagent Fallback (for skills add adopters)

Flow

Step 1: Confirm Category and Responsibilities

Step 2: Delegate to agent-author to Create File

Step 3: Confirm Model Selection and Tools Configuration

Step 4: Validate with frontmatter-linter

Step 5: Guide to Update Sub-agents Table in CLAUDE.md

Validation

Notes