Files

T

Antigravity Agent 504fd5fb42 [G1-Sync] Manual knowledge update

2026-05-10 22:08:15 +09:00

6.1 KiB

Raw Blame History

id, title, category, status, source_trust_level, verification_status, created_at, updated_at, tags, tech_stack, applied_in, aliases

title

Anthropic Skills

Claude / agent 의 modular capability. Markdown skill files + structured tools. Re-usable + composable.

📖 핵심 개념

매 skill = self-contained doc + tool.
Loaded on-demand.
Modular (mix and match).
LLM 가 자체 invoke.

💻 코드 패턴

Skill structure

skills/
├── pdf-processor/
│   ├── SKILL.md           # 의 description + 사용법
│   ├── helpers.ts         # tool implementation
│   └── examples/          # sample input
├── data-analyst/
│   └── SKILL.md
└── code-reviewer/
    └── SKILL.md

SKILL.md

---
name: PDF Processor
description: Extract text and structure from PDF files
---

# PDF Processor

When user asks about PDF processing:
1. Use `extract_text(path)` to get raw text.
2. Use `extract_tables(path)` for tables.
3. Use `summarize(text)` for long docs.

## Tool examples

To extract: extract_text('/path/to/doc.pdf')
To find tables: extract_tables(path, page=2)

→ LLM 가 SKILL.md 읽음 + 자체 use.

Discovery

Agent loop:
1. User query.
2. List available skills.
3. LLM 가 relevant skill 선택.
4. Skill description + tools 가 context 에.
5. Use.

→ Lazy load (token 절약).

Skill 의 ingredient

- Description: 언제 use.
- Examples: input → output.
- Tools: 호출 가능.
- Constraints: 안 do.
- Workflow: step-by-step.

Multi-skill

User: "Analyze this PDF and write a summary email".

LLM 가:
1. PDF Processor skill (extract).
2. Data Analyst skill (analyze).
3. Email Writer skill (write).

→ Compose multiple.

vs MCP server

MCP: standardized protocol (Anthropic).
Skills: skill 의 unit (markdown + tool).

→ Skills 가 MCP 위에 사용 가능.

→ AI_MCP_Server_Building.

Code skill 예

---
name: Code Reviewer
description: Review code for bugs, style, performance
---

# Code Reviewer

For each file:
1. Check for common bugs (null check, off-by-one).
2. Verify style (lint, format).
3. Suggest performance improvement.
4. Output JSON: { issues: [...], suggestions: [...] }

## Don't
- Don't rewrite entire file.
- Don't suggest opinion-based changes (subjective style).

Versioning

skills/
├── pdf-processor/
│   └── v2/
│       ├── SKILL.md
│       └── ...

→ Multiple version, A/B test.

Internal company skills:
- Marketing Copywriter.
- Customer Support Triage.
- Code Reviewer.
- Data Analyst.

→ Library 가 매 team 가 share.

Public skill marketplace

2026 의 가능성:
- "Skills marketplace" (Hugging Face 식).
- 매 skill 가 review / rating.
- 매 use case 의 best skill.

Custom Claude

Claude Pro / Team:
- Custom instruction (system prompt).
- Memory (persistent).
- Skills (loaded on demand).

→ Specialized assistant.

Anthropic Computer Use + Skills

Skill: 'Search the web for X and summarize'.
- Computer Use tool.
- Browser navigation.
- Reading + summary.

→ Modular agent.

Skill 작성 가이드

1. Description 가 specific (애매 X).
2. Examples (input → expected output).
3. Constraints (안 do).
4. Tool 가 명확.
5. Output format (JSON 등).
6. Eval (test cases).

Iterative improve

1. Initial skill.
2. Test cases.
3. Fail case 분석.
4. Skill update.
5. Re-test.

→ "Skill engineering" 의 process.

vs prompt template

Prompt template: 1 prompt + variable.
Skill: doc + tool + workflow.

→ Skill 가 더 큰 unit.

Production

- Code review bot (PR 매번).
- Data analyst (Slack bot).
- Customer support (1-line).
- Personal assistant.

→ 매 task = own skill.

Skills + RAG

Skill 가 own knowledge base:
- Doc 가 vector DB.
- Skill 가 retrieve.
- LLM 가 answer.

→ Skill 가 expert in domain.

Skills + sub-agent

Main agent: orchestrator.
Skill A → sub-agent A.
Skill B → sub-agent B.

→ Multi-agent coordination.

→ AI_Multi_Agent_Coordination.

Anthropic 의 official format

2026 시점 evolving.
- Markdown + frontmatter.
- Tool definitions.
- Examples.
- Constraints.

Implementation 예 (TS)

class SkillManager {
  skills = new Map<string, Skill>();
  
  register(skill: Skill) {
    this.skills.set(skill.name, skill);
  }
  
  list(): SkillSummary[] {
    return [...this.skills.values()].map(s => ({
      name: s.name,
      description: s.description,
    }));
  }
  
  async load(name: string): Promise<string> {
    const skill = this.skills.get(name);
    return skill ? skill.fullContent : '';
  }
}

// LLM 의 system prompt 에 skill list.
// LLM 가 'load_skill(name)' tool 호출.
// Loaded skill content 가 context 에 추가.

함정

- Skill 너무 많음 = LLM 가 헷갈림.
- Description 가 vague.
- Tool overlap.
- No eval.
- Version 관리 X.
- Cross-skill state share.

Best practice

1. Specific name + description.
2. Self-contained (depend X).
3. Test case (10+).
4. Versioning (skill v1, v2).
5. Composable (다른 skill 와 stack).
6. Idempotent (re-run OK).

🤔 의사결정 기준

작업	추천
Specific task	Skill
일반 prompt	System prompt
External integration	MCP
Multi-step	Skill + sub-agent
Re-usable across team	Skill library

❌ 안티패턴

Skill 너무 많음: LLM 헷갈림.
Vague description: 안 invoke.
No example: bad output.
Tool overlap: confusion.
No eval: silent regression.

🤖 LLM 활용 힌트

Skills = modular capability.
Markdown + tool 가 unit.
MCP 위 가능.
Specific name + description 가 핵심.

6.1 KiB Raw Blame History