Skip to content

Commit 27a1631

Browse files
p-iknow이영창(piknow)
p-iknow
and
이영창(piknow)
authored
agent: add meta-skill-creator and meta-agent-creator skills (#28)
agent: update meta-skill description format to match Anthropic docs Requirement: Current meta-skill description format uses custom `Triggers:` syntax and arbitrary constraints that don't align with official Anthropic documentation. Implementation: - Remove `Triggers:` format, use third-person natural language instead - Update name constraint from max 40 chars to 64 chars (per official docs) - Replace ~100 words limit with "keep concise" guideline - Update validation checklist and anti-patterns sections - Convert skill-template.md to English with new format Co-authored-by: 이영창(piknow) <apricotsoul@gmail.com>
1 parent e436e31 commit 27a1631

11 files changed

Lines changed: 1893 additions & 0 deletions

File tree

‎.claude/skills/meta-agent-creator/SKILL.md‎

Lines changed: 203 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,203 @@
1+
---
2+
name: meta-agent-creator
3+
description: Guide for creating specialized AI agents (subagents). This skill should be used when users want to create a new agent, define agent configurations, or need guidance on agent architecture.
4+
---
5+
6+
# Meta-Agent Creator
7+
8+
A meta-skill that helps create new specialized agents for multi-agent orchestration systems.
9+
Supports Claude Code and OpenCode platforms.
10+
11+
> **Output Language**: All generated agent content (prompts, descriptions) will be in **English**.
12+
13+
## Quick Start
14+
15+
### Create a New Agent
16+
17+
```bash
18+
# Initialize a new agent
19+
bun scripts/init-agent.ts <agent-name> --path <output-directory> --platform <platform>
20+
21+
# Examples
22+
bun scripts/init-agent.ts security-auditor --path src/agents --platform opencode
23+
bun scripts/init-agent.ts data-analyst --path .claude/agents --platform claude-code
24+
```
25+
26+
### Validate Agent
27+
28+
```bash
29+
bun scripts/validate-agent.ts <agent-file>
30+
31+
# Examples
32+
bun scripts/validate-agent.ts src/agents/my-agent.ts
33+
bun scripts/validate-agent.ts .claude/agents/my-agent.ts
34+
```
35+
36+
## Agent Categories
37+
38+
Choose the appropriate category based on agent purpose:
39+
40+
| Category | Purpose | Model Tier | Examples |
41+
| --------------- | ------------------------------- | ---------------- | ----------------- |
42+
| `exploration` | Fast search, codebase discovery | LOW (haiku/fast) | explore, searcher |
43+
| `specialist` | Domain-specific implementation | MEDIUM (sonnet) | frontend, backend |
44+
| `advisor` | Read-only consultation | HIGH (opus) | oracle, architect |
45+
| `utility` | General helpers | LOW (haiku/fast) | writer, formatter |
46+
| `orchestration` | Multi-agent coordination | MEDIUM (sonnet) | orchestrator |
47+
48+
## Agent Creation Workflow (5 Phases)
49+
50+
### Phase 1: DEFINE PURPOSE
51+
52+
**Goal**: Clarify agent's role and responsibilities.
53+
54+
**Questions to Ask**:
55+
56+
1. "What specific tasks should this agent handle?"
57+
2. "When should the orchestrator delegate to this agent?"
58+
3. "What domain expertise is required?"
59+
4. "Should it be read-only or have write access?"
60+
61+
**Deliverables**:
62+
63+
- Clear role definition
64+
- Trigger conditions for delegation
65+
- Required capabilities list
66+
- Access level decision (readonly vs full)
67+
68+
### Phase 2: CLASSIFY
69+
70+
**Goal**: Determine category and model tier.
71+
72+
**Decision Matrix**:
73+
74+
| If the agent needs... | Category | Model | Cost |
75+
| ------------------------------- | ------------- | -------------- | --------- |
76+
| Fast codebase search | exploration | haiku/fast | FREE |
77+
| Specific domain implementation | specialist | sonnet/inherit | CHEAP |
78+
| Complex reasoning, architecture | advisor | opus/inherit | EXPENSIVE |
79+
| Simple transformations | utility | haiku/fast | FREE |
80+
| Delegate to other agents | orchestration | sonnet/inherit | CHEAP |
81+
82+
**Deliverables**:
83+
84+
- Category assignment
85+
- Model tier selection
86+
- Cost classification
87+
88+
### Phase 3: DESIGN PROMPT
89+
90+
**Goal**: Write effective system prompt.
91+
92+
**Prompt Structure**:
93+
94+
```markdown
95+
## Role
96+
97+
[1-2 sentence identity and expertise]
98+
99+
## Core Capabilities
100+
101+
- [Capability 1]
102+
- [Capability 2]
103+
104+
## Workflow
105+
106+
1. [Step 1]
107+
2. [Step 2]
108+
109+
## Output Format
110+
111+
[Structured output definition]
112+
113+
## Constraints
114+
115+
- [Constraint 1]
116+
- [Constraint 2]
117+
```
118+
119+
**Best Practices**:
120+
121+
- Keep prompts concise (<500 words)
122+
- Use imperative form
123+
- Include specific output format
124+
- Define clear constraints
125+
126+
### Phase 4: CONFIGURE
127+
128+
**Goal**: Create platform-specific configuration.
129+
130+
#### OpenCode / Claude Code (TypeScript)
131+
132+
Location: `src/agents/<agent-name>.ts`
133+
134+
```typescript
135+
export function createMyAgent(model: string): AgentConfig {
136+
return {
137+
name: "my-agent",
138+
description: "Agent description...",
139+
mode: "subagent",
140+
model,
141+
temperature: 0.1,
142+
tools: { include: ["read", "glob", "grep"] },
143+
prompt: `System prompt...`,
144+
};
145+
}
146+
```
147+
148+
### Phase 5: REGISTER & TEST
149+
150+
**Goal**: Add agent to registry and verify functionality.
151+
152+
**Registration Checklist**:
153+
154+
- [ ] Add to agent definitions/index
155+
- [ ] Update orchestrator's available agents list
156+
- [ ] Add to delegation table if applicable
157+
- [ ] Test with sample delegation
158+
159+
**Testing**:
160+
161+
1. Invoke agent with representative task
162+
2. Verify output format matches spec
163+
3. Check tool restrictions are enforced
164+
4. Test edge cases
165+
166+
## Platform Configurations
167+
168+
### OpenCode / Claude Code Agent Fields
169+
170+
| Field | Required | Description |
171+
| ------------- | -------- | ------------------------ |
172+
| `name` | Yes | Agent identifier |
173+
| `description` | Yes | Selection guide |
174+
| `mode` | Yes | Always `"subagent"` |
175+
| `model` | Yes | Model string |
176+
| `temperature` | No | Defaults to 0.1 |
177+
| `tools` | No | Tool whitelist/blacklist |
178+
| `prompt` | Yes | System prompt |
179+
180+
## Common Agent Templates
181+
182+
See `references/agent-templates.md` for ready-to-use templates:
183+
184+
- Exploration Agent
185+
- Verification Agent
186+
- Debugger Agent
187+
- Security Auditor Agent
188+
189+
## Anti-Patterns to Avoid
190+
191+
| Anti-Pattern | Problem | Solution |
192+
| ----------------- | -------------------------- | ------------------------- |
193+
| Vague description | Orchestrator can't decide | Include specific triggers |
194+
| Too many tools | Security risk | Minimal tool set |
195+
| Long prompts | Slower, harder to maintain | Keep under 500 words |
196+
| Generic "helper" | No clear purpose | Single responsibility |
197+
| Missing readonly | Security for advisors | Set readonly: true |
198+
| Wrong model tier | Cost or quality issues | Match tier to complexity |
199+
200+
## References
201+
202+
- **Agent Templates**: See [references/agent-templates.md](references/agent-templates.md) for ready-to-use templates
203+
- **Agent Patterns**: See [../meta-skill-creator/references/agent-patterns.md](../meta-skill-creator/references/agent-patterns.md) for architecture guidance

0 commit comments

Comments
 (0)