メニュー
Optimize context loading for efficient token usage. Use when working with large codebases, context limits, or when the user mentions "context", "token", "optimize", "summarize", or asks to reduce context size.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | context-optimizer |
| description | Optimize context loading for efficient token usage. Use when working with large codebases, context limits, or when the user mentions "context", "token", "optimize", "summarize", or asks to reduce context size. |
Optimize AI context loading for efficient token usage and focused development sessions.
| Factor | Weight | Description |
|---|---|---|
| Direct relevance | 40% | Directly related to current task |
| Dependency chain | 25% | Required by relevant files |
| Recent access | 20% | Recently read or modified |
| Reference frequency | 15% | Often referenced in codebase |
| Strategy | Load Scope | Token Savings |
|---|---|---|
| A: Essential Only | Modified files + type defs + immediate deps | 60-80% |
| B: Focused Context | Working files + 1-level deps + docs + config | 40-60% |
| C: Summarized | Full working files + summaries + index | 30-50% |
| Priority | Files | When |
|---|---|---|
| 1 (Always) | docs/CONTEXT.md, docs/PROGRESS.md | Every session |
| 2 (Phase) | docs/phases/phase-N/SPEC.md, TASKS.md | Phase work |
| 3 (On-Demand) | docs/PRD.md, docs/TECH-SPEC.md, src/**/* | As needed |
| Session Type | Budget | Loading Strategy |
|---|---|---|
| Quick check | ~2K | CONTEXT.md + PROGRESS.md |
| Standard dev | ~10K | CONTEXT + PROGRESS + active files |
| Deep dive | ~30K | All docs + relevant source |
| Full context | ~50K+ | Complete project load |
Turn-by-turn progressive loading to minimize initial token cost.
| Turn | Load | Tokens | Purpose |
|---|---|---|---|
| 1 | CLAUDE.md (lean) + MANIFEST.md + CONTEXT.md | ~1.1K | User intent detection |
| 2 | Task-specific files (TASKS.md row + source) | +2-3K | Intent-scoped work |
| 3+ | On-demand (SPEC, PRD, etc.) | +varies | As referenced |
Note:
auto_load_phase_docsdeferred until Turn 2.
| User Intent | Load Target | Budget Impact |
|---|---|---|
| Phase 작업 | +phase docs (SPEC, TASKS) | +2-3K |
| 아키텍처 논의 | +TECH-SPEC + adjacent phase SPECs | +5-8K |
| 코드 리뷰 | +PRD (requirements) + CHECKLIST | +3-5K |
| 전체 현황 | +PROGRESS + all phase summaries | +5-10K |
{
"context-optimizer": {
"loading_strategy": "incremental",
"initial_budget": 800,
"expansion_triggers": {
"phase_work": "+phase_docs",
"architecture": "+tech_spec+adjacent_phases",
"review": "+prd+checklist",
"full_status": "+progress+all_phase_summaries"
}
}
}
| Level | Scope | Tokens | Contents |
|---|---|---|---|
| 1: Task-Active (default) | Single task | ~3-6K | CONTEXT.md + Task row + source + direct deps |
| 2: Phase-Full | Full phase | ~8-10K | Level 1 + SPEC.md + CHECKLIST.md |
| 3: Cross-Phase | Architecture | ~12-15K | Level 2 + adjacent SPEC.md + TECH-SPEC sections |
## Task T2-03 Context
Source: docs/phases/phase-2/TASKS.md (row 3 only)
| ID | Task | Status | Priority | Est |
|----|------|--------|----------|-----|
| T2-03 | API endpoint 구현 | 🔄 | P0 | 3h |
Related files: server/src/routes/api.ts, server/src/models/schema.ts
Dependencies: T2-01 (DB schema), T2-02 (auth middleware)
Auto-save on context threshold (>80% budget) for seamless recovery.
# Session Checkpoint
Date: {{TIMESTAMP}}
Phase: {{CURRENT_PHASE}} | Task: {{TASK_ID}} | Progress: {{PHASE_PROGRESS}}%
## State
- Working on: {{TASK_DESCRIPTION}}
- Modified: {{FILE_LIST}}
## Decisions / Next / Resume Command
Recovery cost: CLAUDE.md (300) + checkpoint (1.6K) = ~2K for instant recovery.
Location: .claude/templates/CLAUDE.lean.md
| Variable | Source | Example |
|---|---|---|
{{PROJECT_NAME}} | DISCOVERY.md | "Resumely" |
{{TECH_STACK}} | DISCOVERY.md | "Next.js+Supabase" |
{{CURRENT_PHASE}} | PROGRESS.md | "2" |
{{PHASE_PROGRESS}} | TASKS.md calc | "60" |
| Format | Tokens | Savings |
|---|---|---|
| Standard CLAUDE.md | ~1,700+ | baseline |
| Lean CLAUDE.md | ~300 | ~82% |
Location: .claude/agents/MANIFEST.md
1. MANIFEST.md loaded (~500 tokens, 25 agents × 1 row)
2. Match user intent → keyword lookup
3. Load only matched agent file (~1-3K tokens)
Total: ~1.5-3.5K per invocation vs. ~38K if all loaded
| # | Practice |
|---|---|
| 1 | Start Lean: Ultra-lean CLAUDE.md + MANIFEST only |
| 2 | Task-First: Scope to current Task, not entire Phase |
| 3 | Expand Incrementally: Add files only when referenced |
| 4 | Structured > Prose: Use tables/key-value over paragraphs |
| 5 | Checkpoint Often: Auto-save before context exhaustion |
| 6 | Phase-Scoped Loading: Load current Phase docs by default |
| 7 | Summarize Adjacents: Only SPEC.md from neighboring Phases |
| 8 | Budget Awareness: Match loading strategy to session type |
Bidirectional Notion sync for project management data via Notion MCP. Sync tasks, documents, sprints, gates, and activity logs. "notion", "노션", "동기화", "sync", "notion push", "notion pull", "notion status" 키워드에 반응.
Office document automation -- PPTX presentations, DOCX documents, XLSX spreadsheets, PDF reports. "office", "pptx", "docx", "xlsx", "pdf", "프레젠테이션", "presentation", "보고서", "report", "엑셀", "워드" 키워드에 반응.
Programmatic video generation with React using Remotion. JSX/CSS compositions rendered to MP4. "remotion", "video", "비디오", "MP4", "animation", "워크스루", "demo video", "영상" 키워드에 반응.
Google Labs stitch-skills integration for design systems, React components, multi-page websites, prompt enhancement, walkthrough videos, and shadcn/ui. "stitch", "stitch-skills", "design system", "react components", "stitch-loop", "enhance-prompt", "shadcn" 키워드에 반응.
Synchronize all agile development artifacts in one command. Updates CHANGELOG, README stats, progress tracking, and validates documentation completeness.
Claude Code와 Codex CLI의 토큰 사용량 및 비용 리포트를 조회하는 스킬. "ccusage", "토큰 사용량", "usage report", "codex usage", "claude usage", "비용 리포트", "token usage", "cost report" 키워드에 반응.