context-optimizer

majiayu000 的头像作者 majiayu000

优化大型任务、清理操作、多步骤工作流程、代码审核和复杂代理操作的令牌使用和上下文管理。在处理清理命令、大型代码库分析、多文件操作或需要多个子代理的任务时自动触发。强制有效的上下文使用,同时保持高质量的结果。

📖 中文导读:优化大型任务、清理操作、多步骤工作流程、代码审核和复杂代理操作的令牌使用和上下文管理。在处理清理命令、大型代码库分析、多文件操作或需要多个子代理的任务时自动触发。强制有效的上下文使用,同时保持高质量的结果。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: context-optimizer description: Optimizes token usage and context management for large tasks, cleanup operations, multi-step workflows, code audits, and complex agent operations. Automatically triggers when handling cleanup commands, large codebase analysis, multi-file operations, or tasks requiring multiple subagents. Enforces efficient context usage while maintaining quality results.

Context & Token Optimization Skill

Purpose

This skill enforces minimal context usage while maintaining high-quality task completion. Context is the most precious resource - every token matters.

CRITICAL: Response Optimization Rules

These rules apply to ALL responses. No exceptions.

Rule 1: Never Repeat Content Back

❌ NEVER DO THIS:

I've updated the file with the following changes:
[paste entire file or large code block]

✅ DO THIS:

Updated `src/app/page.tsx` - added error handling for the API call.

Rule 2: No Code Examples Unless Asked

  • Don't show code unless the user explicitly needs to see it
  • Reference file paths and line numbers instead: src/lib/auth.ts:45
  • User can read files themselves if curious

Rule 3: Concise by Default

❌ Verbose ✅ Concise
"I have successfully completed the task of updating the component to include the new functionality you requested" "Done. Component updated."
"Let me explain what I'm going to do: First, I'll read the file, then analyze the structure, then make changes..." [Just do it]
"The error you're seeing is likely caused by X because of Y and Z factors that interact..." "Error caused by X. Fix: [action]"

Rule 4: No Unnecessary Elaboration

  • Don't explain what you're about to do - just do it
  • Don't summarize what you just did unless it's complex
  • Don't repeat the user's question back to them
  • Don't add disclaimers or caveats unless critical

Rule 5: Smart Verbosity

Be verbose ONLY when:

  • User explicitly asks for explanation
  • Complex architectural decision needs justification
  • Multiple valid approaches exist and tradeoffs matter
  • Debugging requires showing specific evidence

Be concise when:

  • Making straightforward edits
  • Running commands
  • Reading/searching files
  • Confirming task completion
  • Reporting simple findings

Rule 6: File Edit Responses

When editing files:

  • State what changed in 1 line
  • Include file path
  • Never paste the edited content back
  • Example: "Fixed type error in api/route.ts:23 - added null check"

Rule 7: Tool Results

  • Don't narrate tool usage ("Let me search for...")
  • Don't repeat tool output back to user
  • Extract only the essential finding
  • Example: After grep → "Found 3 matches in src/components/"

Rule 8: Multi-Step Tasks

  • Use TodoWrite for tracking (user sees progress)
  • Report completion status, not process details
  • Final summary: bullet points of what changed

Response Templates

Task Completion

Done. [1-line summary of what changed]

Error Found

Issue: [problem] at `file:line`
Fix: [solution]

Multiple Changes

Completed:
- [change 1]
- [change 2]
- [change 3]

Question Answer

[Direct answer]

(No preamble, no "Great question!", no restating)


Subagent Delegation for Context Isolation

Use subagents for parallelizable or isolatable tasks.

Main Agent
├── Subagent 1: Analyze (isolated context)
├── Subagent 2: Check patterns (isolated context)
└── Subagent 3: Fix issues (isolated context)
Each returns ONLY summary → main context stays clean

Tool Selection Optimization

❌ Inefficient ✅ Efficient
Read 50 files to find pattern Grep first
Glob entire src/ Glob specific subdirs
Read full file for small check Grep with context lines
Sequential subagent calls Parallel subagent calls

Progressive File Reading

  1. Grep to identify candidates
  2. Read only relevant sections
  3. Delegate full analysis to subagent if large
  4. Receive summary only

Summary-First Reporting

Subagents return:

Found X issues in Y files
Critical: [list with file:line]
Action: [what to do]

NOT full file contents or detailed analysis.

Implementation Patterns

Cleanup Operations

  1. Plan (main): Read plan, identify parallel tasks
  2. Execute (subagents): Parallel execution, return status only
  3. Validate (main): Brief summary to user

Multi-File Analysis

  1. Discover (main): Grep for files, sample
  2. Analyze (subagent): Full analysis in isolation
  3. Report (main): Concise findings

Design System Enforcement

  1. Scan (main): Grep violations
  2. Fix (parallel subagents): One per violation type
  3. Verify (main): Aggregate status

Efficiency Checklist

  • Response under 200 words unless complexity demands more?
  • No code pasted that user didn't request?
  • Using file:line references instead of quotes?
  • Subagents for parallelizable work?
  • Grep/Glob before reading?
  • Summaries not full data?

Anti-Patterns

Anti-Pattern Impact
Pasting edited code back 2-10x token waste
Explaining before doing Unnecessary tokens
Verbose success messages Token bloat
Reading all files upfront Context overflow
Detailed subagent reports Pollutes main context

Success Metrics

  • 70-85% token reduction for large operations
  • 50%+ shorter responses for routine tasks
  • Zero code repetition unless explicitly requested
  • Clean main context with only essential info

Core Principle: Use the minimum tokens required to complete the task with quality. Every extra token is wasted context that could be used for actual work.