agent-development

Name: agent-development
Author: anthropics

byanthropics· 237 skills

Installs

16k

Stars

139k

Forks

22k

How to install agent-development?

anthropics/claude-code/agent-development

$npx -y skills add anthropics/claude-code --skill agent-development

Installs into the current project.

›Prefer a prompt? Paste this to your agent

Use this skill

Run `npx skills use "https://github.com/anthropics/claude-code" --skill "anthropics/claude-code/agent-development"` and follow the generated skill instructions now. Read its complete output, redirecting it to a temporary file first if necessary. Resolve relative paths from the supporting-files directory it provides.

Use the whole pack

Use the skills in "https://github.com/anthropics/claude-code" that are relevant to the current task. Run `npx skills add "https://github.com/anthropics/claude-code"` and select the relevant skills, then follow their instructions.

Files · 1

View on GitHub

SKILL.md

1	# Agent Development for Claude Code Plugins
2
3	## Overview
4
5	Agents are autonomous subprocesses that handle complex, multi-step tasks independently. Understanding agent structure, triggering conditions, and system prompt design enables creating powerful autonomous capabilities.
6
7	Key concepts:
8	- Agents are FOR autonomous work, commands are FOR user-initiated actions
9	- Markdown file format with YAML frontmatter
10	- Triggering via description field with examples
11	- System prompt defines agent behavior
12	- Model and color customization
13
14	## Agent File Structure
15
16	### Complete Format
17
18	```markdown
19	---
20	name: agent-identifier
21	description: Use this agent when [triggering conditions]. Examples:
22
23	<example>
24	Context: [Situation description]
25	user: "[User request]"
26	assistant: "[How assistant should respond and use this agent]"
27	<commentary>
28	[Why this agent should be triggered]
29	</commentary>
30	</example>
31
32	<example>
33	[Additional example...]
34	</example>
35
36	model: inherit
37	color: blue
38	tools: ["Read", "Write", "Grep"]
39	---
40
41	You are [agent role description]...
42
43	Your Core Responsibilities:
44	1. [Responsibility 1]
45	2. [Responsibility 2]
46
47	Analysis Process:
48	[Step-by-step workflow]
49
50	Output Format:
51	[What to return]
52	```
53
54	## Frontmatter Fields
55
56	### name (required)
57
58	Agent identifier used for namespacing and invocation.
59
60	Format: lowercase, numbers, hyphens only
61	Length: 3-50 characters
62	Pattern: Must start and end with alphanumeric
63
64	Good examples:
65	- `code-reviewer`
66	- `test-generator`
67	- `api-docs-writer`
68	- `security-analyzer`
69
70	Bad examples:
71	- `helper` (too generic)
72	- `-agent-` (starts/ends with hyphen)
73	- `my_agent` (underscores not allowed)
74	- `ag` (too short, < 3 chars)
75
76	### description (required)
77
78	Defines when Claude should trigger this agent. This is the most critical field.
79
80	Must include:
81	1. Triggering conditions ("Use this agent when...")
82	2. Multiple `<example>` blocks showing usage
83	3. Context, user request, and assistant response in each example
84	4. `<commentary>` explaining why agent triggers
85
86	Format:
87	```
88	Use this agent when [conditions]. Examples:
89
90	<example>
91	Context: [Scenario description]
92	user: "[What user says]"
93	assistant: "[How Claude should respond]"
94	<commentary>
95	[Why this agent is appropriate]
96	</commentary>
97	</example>
98
99	[More examples...]
100	```
101
102	Best practices:
103	- Include 2-4 concrete examples
104	- Show proactive and reactive triggering
105	- Cover different phrasings of same intent
106	- Explain reasoning in commentary
107	- Be specific about when NOT to use the agent
108
109	### model (required)
110
111	Which model the agent should use.
112
113	Options:
114	- `inherit` - Use same model as parent (recommended)
115	- `sonnet` - Claude Sonnet (balanced)
116	- `opus` - Claude Opus (most capable, expensive)
117	- `haiku` - Claude Haiku (fast, cheap)
118
119	Recommendation: Use `inherit` unless agent needs specific model capabilities.
120
121	### color (required)
122
123	Visual identifier for agent in UI.
124
125	Options: `blue`, `cyan`, `green`, `yellow`, `magenta`, `red`
126
127	Guidelines:
128	- Choose distinct colors for different agents in same plugin
129	- Use consistent colors for similar agent types
130	- Blue/cyan: Analysis, review
131	- Green: Success-oriented tasks
132	- Yellow: Caution, validation
133	- Red: Critical, security
134	- Magenta: Creative, generation
135
136	### tools (optional)
137
138	Restrict agent to specific tools.
139
140	Format: Array of tool names
141
142	```yaml
143	tools: ["Read", "Write", "Grep", "Bash"]
144	```
145
146	Default: If omitted, agent has access to all tools
147
148	Best practice: Limit tools to minimum needed (principle of least privilege)
149
150	Common tool sets:
151	- Read-only analysis: `["Read", "Grep", "Glob"]`
152	- Code generation: `["Read", "Write", "Grep"]`
153	- Testing: `["Read", "Bash", "Grep"]`
154	- Full access: Omit field or use `["*"]`
155
156	## System Prompt Design
157
158	The markdown body becomes the agent's system prompt. Write in second person, addressing the agent directly.
159
160	### Structure
161
162	Standard template:
163	```markdown
164	You are [role] specializing in [domain].
165
166	Your Core Responsibilities:
167	1. [Primary responsibility]
168	2. [Secondary responsibility]
169	3. [Additional responsibilities...]
170
171	Analysis Process:
172	1. [Step one]
173	2. [Step two]
174	3. [Step three]
175	[...]
176
177	Quality Standards:
178	- [Standard 1]
179	- [Standard 2]
180
181	Output Format:
182	Provide results in this format:
183	- [What to include]
184	- [How to structure]
185
186	Edge Cases:
187	Handle these situations:
188	- [Edge case 1]: [How to handle]
189	- [Edge case 2]: [How to handle]
190	```
191
192	### Best Practices
193
194	✅ DO:
195	- Write in second person ("You are...", "You will...")
196	- Be specific about responsibilities
197	- Provide step-by-step process
198	- Define output format
199	- Include quality standards
200	- Address edge cases
201	- Keep under 10,000 characters
202
203	❌ DON'T:
204	- Write in first person ("I am...", "I will...")
205	- Be vague or generic
206	- Omit process steps
207	- Leave output format undefined
208	- Skip quality guidance
209	- Ignore error cases
210
211	## Creating Agents
212
213	### Method 1: AI-Assisted Generation
214
215	Use this prompt pattern (extracted from Claude Code):
216
217	```
218	Create an agent configuration based on this request: "[YOUR DESCRIPTION]"
219
220	Requirements:
221	1. Extract core intent and responsibilities
222	2. Design expert persona for the domain
223	3. Create comprehensive system prompt with:
224	- Clear behavioral boundaries
225	- Specific methodologies
226	- Edge case handling
227	- Output format
228	4. Create identifier (lowercase, hyphens, 3-50 chars)
229	5. Write description with triggering conditions
230	6. Include 2-3 <example> blocks showing when to use
231
232	Return JSON with:
233	{
234	"identifier": "agent-name",
235	"whenToUse": "Use this agent when... Examples: <example>...</example>",
236	"systemPrompt": "You are..."
237	}
238	```
239
240	Then convert to agent file format with frontmatter.
241
242	See `examples/agent-creation-prompt.md` for complete template.
243
244	### Method 2: Manual Creation
245
246	1. Choose agent identifier (3-50 chars, lowercase, hyphens)
247	2. Write description with examples
248	3. Select model (usually `inherit`)
249	4. Choose color for visual identification
250	5. Define tools (if restricting access)
251	6. Write system prompt with structure above
252	7. Save as `agents/agent-name.md`
253
254	## Validation Rules
255
256	### Identifier Validation
257
258	```
259	✅ Valid: code-reviewer, test-gen, api-analyzer-v2
260	❌ Invalid: ag (too short), -start (starts with hyphen), my_agent (underscore)
261	```
262
263	Rules:
264	- 3-50 characters
265	- Lowercase letters, numbers, hyphens only
266	- Must start and end with alphanumeric
267	- No underscores, spaces, or special characters
268
269	### Description Validation
270
271	Length: 10-5,000 characters
272	Must include: Triggering conditions and examples
273	Best: 200-1,000 characters with 2-4 examples
274
275	### System Prompt Validation
276
277	Length: 20-10,000 characters
278	Best: 500-3,000 characters
279	Structure: Clear responsibilities, process, output format
280
281	## Agent Organization
282
283	### Plugin Agents Directory
284
285	```
286	plugin-name/
287	└── agents/
288	├── analyzer.md
289	├── reviewer.md
290	└── generator.md
291	```
292
293	All `.md` files in `agents/` are auto-discovered.
294
295	### Namespacing
296
297	Agents are namespaced automatically:
298	- Single plugin: `agent-name`
299	- With subdirectories: `plugin:subdir:agent-name`
300
301	## Testing Agents
302
303	### Test Triggering
304
305	Create test scenarios to verify agent triggers correctly:
306
307	1. Write agent with specific triggering examples
308	2. Use similar phrasing to examples in test
309	3. Check Claude loads the agent
310	4. Verify agent provides expected functionality
311
312	### Test System Prompt
313
314	Ensure system prompt is complete:
315
316	1. Give agent typical task
317	2. Check it follows process steps
318	3. Verify output format is correct
319	4. Test edge cases mentioned in prompt
320	5. Confirm quality standards are met
321
322	## Quick Reference
323
324	### Minimal Agent
325
326	```markdown
327	---
328	name: simple-agent
329	description: Use this agent when... Examples: <example>...</example>
330	model: inherit
331	color: blue
332	---
333
334	You are an agent that [does X].
335
336	Process:
337	1. [Step 1]
338	2. [Step 2]
339
340	Output: [What to provide]
341	```
342
343	### Frontmatter Fields Summary
344
345	\| Field \| Required \| Format \| Example \|
346	\|-------\|----------\|--------\|---------\|
347	\| name \| Yes \| lowercase-hyphens \| code-reviewer \|
348	\| description \| Yes \| Text + examples \| Use when... <example>... \|
349	\| model \| Yes \| inherit/sonnet/opus/haiku \| inherit \|
350	\| color \| Yes \| Color name \| blue \|
351	\| tools \| No \| Array of tool names \| ["Read", "Grep"] \|
352
353	### Best Practices
354
355	DO:
356	- ✅ Include 2-4 concrete examples in description
357	- ✅ Write specific triggering conditions
358	- ✅ Use `inherit` for model unless specific need
359	- ✅ Choose appropriate tools (least privilege)
360	- ✅ Write clear, structured system prompts
361	- ✅ Test agent triggering thoroughly
362
363	DON'T:
364	- ❌ Use generic descriptions without examples
365	- ❌ Omit triggering conditions
366	- ❌ Give all agents same color
367	- ❌ Grant unnecessary tool access
368	- ❌ Write vague system prompts
369	- ❌ Skip testing
370
371	## Additional Resources
372
373	### Reference Files
374
375	For detailed guidance, consult:
376
377	- `references/system-prompt-design.md` - Complete system prompt patterns
378	- `references/triggering-examples.md` - Example formats and best practices
379	- `references/agent-creation-system-prompt.md` - The exact prompt from Claude Code
380
381	### Example Files
382
383	Working examples in `examples/`:
384
385	- `agent-creation-prompt.md` - AI-assisted agent generation template
386	- `complete-agent-examples.md` - Full agent examples for different use cases
387
388	### Utility Scripts
389
390	Development tools in `scripts/`:
391
392	- `validate-agent.sh` - Validate agent file structure
393	- `test-agent-trigger.sh` - Test if agent triggers correctly
394
395	## Implementation Workflow
396
397	To create an agent for a plugin:
398
399	1. Define agent purpose and triggering conditions
400	2. Choose creation method (AI-assisted or manual)
401	3. Create `agents/agent-name.md` file
402	4. Write frontmatter with all required fields
403	5. Write system prompt following best practices
404	6. Include 2-4 triggering examples in description
405	7. Validate with `scripts/validate-agent.sh`
406	8. Test triggering with real scenarios
407	9. Document agent in plugin README
408
409	Focus on clear triggering conditions and comprehensive system prompts for autonomous operation.

6 picks

Type

Run `npx skills use "https://github.com/anthropics/claude-code" --skill "anthropics/claude-code/agent-development"` and follow the generated skill instructions now. Read its complete output, redirecting it to a temporary file first if necessary. Resolve relative paths from the supporting-files directory it provides.

Use the skills in "https://github.com/anthropics/claude-code" that are relevant to the current task. Run `npx skills add "https://github.com/anthropics/claude-code"` and select the relevant skills, then follow their instructions.

agent-development

How to install agent-development?

Related

agent-development

How to install agent-development?

Related