사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/everyinc/compound-writing --skill cw-emergent시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
Emergent Capability
Purpose
Enable the agent to handle requests that don't map directly to a specific skill by composing primitives and existing skills into novel solutions.
This is what makes the system agent-native: the ability to figure out how to accomplish an outcome, not just execute predefined workflows.
Read ../../references/context-contract.md before composing a multi-artifact workflow. Respect the active project's source-of-truth files, provenance, destinations, and approval rules.
When This Applies
Use emergent composition when:
- Request spans multiple drafts or files
- Request requires pattern analysis across the body of work
- Request combines multiple types of feedback
- No single skill handles the full request
- User describes an outcome, not a command
Do NOT use when:
- A skill directly handles the request (use the skill)
- The request is a simple file operation
- The user invokes a specific command
The Composition Loop
When a request doesn't map to a skill:
- Understand the outcome
- What does success look like?
- What would the user see/know when this is done?
- Identify capabilities needed
- What artifacts must be found, read, compared, or revised?
- What maintained context or connected source must be loaded?
- What draft, source, or destination operations are authorized?
- What analysis operations? (compare, analyze voice, check AI)
- Identify skills that help
- Which existing skills apply to parts of the request?
- Can skills be chained for the full solution?
- Compose a solution
- Chain primitives and skills together
- Plan the sequence of operations
- Execute in a loop
- Work through the plan
- Adapt if something unexpected comes up
- Continue until outcome achieved or blocked
- Report results
- Summarize what was found/done
- Offer next steps if applicable
Example Requests and Approaches
| Request | Primitives | Skills | Approach |
|---|---|---|---|
| "Compare voice in my last 3 drafts" | Find and read the relevant pieces | cw-voice-check | Analyze each against the same live voice sources, then compare |
| "Find patterns in my hooks" | Find and read the relevant pieces | — | Extract openings, classify moves, and report with examples |
| "What objections keep coming up?" | Read the relevant drafts and reviews | cw-objections | Compare repeated resistance and distinguish writer pattern from topic pattern |
| "Am I overusing certain transitions?" | Search the requested corpus | cw-transition | Count, inspect context, and report only meaningful repetition |
| "Make this article better" | Load the draft, brief, voice, and sources | cw-dev-edit, cw-line-edit, cw-ai-check | Diagnose, revise, then validate |
| "Help me finish this piece" | Inspect the active artifact and open decisions | varies | Route to the smallest unfinished outcome |
| "Cross-reference notes with outline" | Read both authoritative artifacts | — | Compare claims, examples, sources, and omissions |
Judgment: When to Compose
Good candidates for composition:
- "Compare X across Y" — Analysis pattern
- "Find patterns in Z" — Discovery pattern
- "Check if A matches B" — Validation pattern
- "Make this better" — Improvement pattern
- "Help me with X" (vague) — Needs investigation first
Not good candidates:
- "Give me a hook" → Use /cw-hook directly
- "Does this bury the lede?" → Use /cw-bluf directly
- "Check my voice" → Use /cw-voice-check directly
- "Run the nemesis read" → Use /cw-nemesis directly
The Ultimate Test
Can the agent accomplish tasks you didn't explicitly design for?
Example: User asks "Cross-reference my meeting notes with my task list and tell me what I've committed to but haven't scheduled."
This wasn't designed as a feature. But if the agent can:
- Read the meeting notes
- Read the task list
- Extract commitments from notes
- Compare against tasks
- Report the gaps
...then it's agent-native.
Failure Modes
If composition fails:
- Missing capability or access: Report what is unavailable
- Skill not available: Suggest manual alternative
- Too complex: Break into smaller requests
- Ambiguous: Ask for clarification
Never say "I don't have a feature for that" when composition could work.
Reporting Results
When composition completes, report:
- What was requested
- What was done (briefly)
- What was found/achieved
- Suggested next steps (if applicable)
Keep reports focused on insights, not process details.
Lessons
[Skill-specific lessons will be added here as they're captured]
