subagent-creator

Name: subagent-creator
Author: luongnv89

97stars15forks

How to install subagent-creator?

luongnv89/skills/subagent-creator

$npx -y skills add luongnv89/skills --skill subagent-creator

Installs into the current project.

Summary

Create, evaluate, or improve Claude Code subagent files (.claude/agents/*.md) — the frontmatter + system prompt defining a delegatable specialist. Don't use for skills (skill-creator), CLAUDE.md/AGENTS.md (agent-config), or running an agent.

Files · 1

View on GitHub

SKILL.md

1	# Subagent Creator
2
3	Create, evaluate, and improve Claude Code subagents — the `.md` files (in `.claude/agents/` for a project or `~/.claude/agents/` for personal use) whose YAML frontmatter plus system prompt define a specialist the main agent can delegate to with an isolated context window and restricted tools.
4
5	This is not about skills (`skill-creator` owns those) or `CLAUDE.md`/`AGENTS.md` (`agent-config` owns those). A subagent is one file: frontmatter (`name`, `description`, `tools`, `model`, …) + a system-prompt body.
6
7	## When to Use
8
9	Use when the user asks to create, review, or fix a Claude Code subagent definition file (`.claude/agents/.md` or personal `~/.claude/agents/.md`). Do not use for skills or for editing CLAUDE.md/AGENTS.md.
10
11	## Pick the branch first
12
13	Three branches. Identify which one the user is on before doing anything else — they don't share a starting step.
14
15	\| Branch \| Trigger \| Go to \|
16	\|--------\|---------\|-------\|
17	\| Create \| "make a subagent that…", "I need an agent for X", no file exists yet \| [Creating a subagent](#creating-a-subagent) \|
18	\| Evaluate \| "review this agent", "is this subagent any good", "audit `.claude/agents/foo.md`" \| [Evaluating a subagent](#evaluating-a-subagent) \|
19	\| Improve \| "fix this agent", "this subagent isn't triggering / over-reaches", points at an existing file \| [Improving a subagent](#improving-a-subagent) \|
20
21	If the request is ambiguous ("look at my agent"), assume Evaluate and confirm before editing anything.
22
23	## Mandatory rules (all branches)
24
25	These apply on every write, regardless of branch.
26
27	### Repo Sync Before Edits (when target is a project repo)
28
29	A subagent written to `.claude/agents/` lives in a version-controlled repo. Before creating or editing a file there, sync to avoid clobbering remote work:
30
31	```bash
32	branch="$(git rev-parse --abbrev-ref HEAD)"
33	git fetch origin
34	git pull --rebase origin "$branch"
35	```
36
37	If the working tree is dirty: stash → sync → pop. If `origin` is missing or a conflict occurs: stop and ask the user. Skip this section only when the target is `~/.claude/agents/` (personal, not a repo) — confirm the scope with the user before assuming.
38
39	### Frontmatter safety + version
40
41	- Subagent file `name` must be lowercase letters/digits/hyphens, no consecutive hyphens. Identity comes from `name`, not the filename — they need not match, but keeping `name` == filename stem (`code-reviewer` → `code-reviewer.md`) is a recommended convention.
42	- Quote any frontmatter string value containing `:` `#` `-` `<` `>` `\|` `,` `&` `?` `!` `[` `]` `{` `}` `*` to keep YAML valid. `description` almost always needs quotes.
43	- This SKILL.md itself follows the catalog's own rule: bump `metadata.version` on every edit.
44
45	### Scope check (one question, not an interview)
46
47	Confirm where the file goes — `.claude/agents/<name>.md` (project, team-shared, version-controlled) or `~/.claude/agents/<name>.md` (personal, cross-project). This decides Repo Sync above and nothing else. The source guideline encodes the rest of the intent, so don't over-interview.
48
49	---
50
51	## Creating a subagent
52
53	Goal: one well-formed `<name>.md` file that triggers reliably, stays in its lane, and holds only the tools it needs.
54
55	### 1. Capture the specialist in one sentence
56
57	A good subagent has a single responsibility. Force the purpose into one sentence: "A read-only security auditor that flags auth and data-exposure bugs in changed files." If you need "and" to join two unrelated jobs, that's two subagents. Read `references/system-prompt-guide.md` → Single responsibility if the scope feels fuzzy.
58
59	### 2. Choose frontmatter
60
61	Start from the minimal set and add only what the job demands. Read `references/frontmatter-schema.md` for every field, which are safe-and-common vs. advanced-and-rot-prone, and the tool-restriction syntax.
62
63	Minimal viable frontmatter:
64
65	```yaml
66	---
67	name: <kebab-name>
68	description: "<when to invoke + what it does. Add 'Use PROACTIVELY' to encourage auto-delegation.>"
69	tools: Read, Grep, Bash # omit entirely to inherit ALL tools
70	model: inherit # inherit \| sonnet \| opus \| haiku
71	---
72	```
73
74	Decisions, in order:
75	- `description` — the primary trigger. Say when to invoke and what it does. Add `Use PROACTIVELY` (or `MUST BE USED`) when the agent should auto-delegate. This is the single highest-leverage field; see the description guidance in `references/system-prompt-guide.md`.
76	- `tools` — least privilege. Prefer read-only (`Read, Grep, Glob`) for analysis/review agents; add `Edit`/`Write` only for agents that change code; scope `

People who liked this also viewed

6 picks

Type

subagent-creator

97stars15forks

View on GitHub

How to install subagent-creator?

luongnv89/skills/subagent-creator

$npx -y skills add luongnv89/skills --skill subagent-creator

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Subagent Creator
2
3	Create, evaluate, and improve Claude Code subagents — the `.md` files (in `.claude/agents/` for a project or `~/.claude/agents/` for personal use) whose YAML frontmatter plus system prompt define a specialist the main agent can delegate to with an isolated context window and restricted tools.
4
5	This is not about skills (`skill-creator` owns those) or `CLAUDE.md`/`AGENTS.md` (`agent-config` owns those). A subagent is one file: frontmatter (`name`, `description`, `tools`, `model`, …) + a system-prompt body.
6
7	## When to Use
8
9	Use when the user asks to create, review, or fix a Claude Code subagent definition file (`.claude/agents/.md` or personal `~/.claude/agents/.md`). Do not use for skills or for editing CLAUDE.md/AGENTS.md.
10
11	## Pick the branch first
12
13	Three branches. Identify which one the user is on before doing anything else — they don't share a starting step.
14
15	\| Branch \| Trigger \| Go to \|
16	\|--------\|---------\|-------\|
17	\| Create \| "make a subagent that…", "I need an agent for X", no file exists yet \| [Creating a subagent](#creating-a-subagent) \|
18	\| Evaluate \| "review this agent", "is this subagent any good", "audit `.claude/agents/foo.md`" \| [Evaluating a subagent](#evaluating-a-subagent) \|
19	\| Improve \| "fix this agent", "this subagent isn't triggering / over-reaches", points at an existing file \| [Improving a subagent](#improving-a-subagent) \|
20
21	If the request is ambiguous ("look at my agent"), assume Evaluate and confirm before editing anything.
22
23	## Mandatory rules (all branches)
24
25	These apply on every write, regardless of branch.
26
27	### Repo Sync Before Edits (when target is a project repo)
28
29	A subagent written to `.claude/agents/` lives in a version-controlled repo. Before creating or editing a file there, sync to avoid clobbering remote work:
30
31	```bash
32	branch="$(git rev-parse --abbrev-ref HEAD)"
33	git fetch origin
34	git pull --rebase origin "$branch"
35	```
36
37	If the working tree is dirty: stash → sync → pop. If `origin` is missing or a conflict occurs: stop and ask the user. Skip this section only when the target is `~/.claude/agents/` (personal, not a repo) — confirm the scope with the user before assuming.
38
39	### Frontmatter safety + version
40
41	- Subagent file `name` must be lowercase letters/digits/hyphens, no consecutive hyphens. Identity comes from `name`, not the filename — they need not match, but keeping `name` == filename stem (`code-reviewer` → `code-reviewer.md`) is a recommended convention.
42	- Quote any frontmatter string value containing `:` `#` `-` `<` `>` `\|` `,` `&` `?` `!` `[` `]` `{` `}` `*` to keep YAML valid. `description` almost always needs quotes.
43	- This SKILL.md itself follows the catalog's own rule: bump `metadata.version` on every edit.
44
45	### Scope check (one question, not an interview)
46
47	Confirm where the file goes — `.claude/agents/<name>.md` (project, team-shared, version-controlled) or `~/.claude/agents/<name>.md` (personal, cross-project). This decides Repo Sync above and nothing else. The source guideline encodes the rest of the intent, so don't over-interview.
48
49	---
50
51	## Creating a subagent
52
53	Goal: one well-formed `<name>.md` file that triggers reliably, stays in its lane, and holds only the tools it needs.
54
55	### 1. Capture the specialist in one sentence
56
57	A good subagent has a single responsibility. Force the purpose into one sentence: "A read-only security auditor that flags auth and data-exposure bugs in changed files." If you need "and" to join two unrelated jobs, that's two subagents. Read `references/system-prompt-guide.md` → Single responsibility if the scope feels fuzzy.
58
59	### 2. Choose frontmatter
60
61	Start from the minimal set and add only what the job demands. Read `references/frontmatter-schema.md` for every field, which are safe-and-common vs. advanced-and-rot-prone, and the tool-restriction syntax.
62
63	Minimal viable frontmatter:
64
65	```yaml
66	---
67	name: <kebab-name>
68	description: "<when to invoke + what it does. Add 'Use PROACTIVELY' to encourage auto-delegation.>"
69	tools: Read, Grep, Bash # omit entirely to inherit ALL tools
70	model: inherit # inherit \| sonnet \| opus \| haiku
71	---
72	```
73
74	Decisions, in order:
75	- `description` — the primary trigger. Say when to invoke and what it does. Add `Use PROACTIVELY` (or `MUST BE USED`) when the agent should auto-delegate. This is the single highest-leverage field; see the description guidance in `references/system-prompt-guide.md`.
76	- `tools` — least privilege. Prefer read-only (`Read, Grep, Glob`) for analysis/review agents; add `Edit`/`Write` only for agents that change code; scope `

People who liked this also viewed

6 picks

Type