菜单
Launch multiple specialized Claude agents simultaneously to maximize productivity. Achieves 3-5x speedup on independent tasks like benchmarking, documentation, and analysis. Use when you have multiple independent tasks that each take more than 1 minute.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Drive performance optimizations through a disciplined measure-optimize-measure cycle. Write benchmarks first, optimize with data, and validate improvements empirically. Use when building performance-critical code paths, lock-free structures, or when performance requirements are quantified.
Prevent false sharing in concurrent Rust code by aligning data structures to cache-line boundaries. Critical for lock-free data structures and multi-threaded hot paths. Use when building multi-threaded code with shared atomics or lock-free structures.
Document decisions, challenges, and solutions immediately after completing each phase. Captures context that would be lost later, creates invaluable reference for future work. Use for complex projects with non-obvious design decisions or any project taking more than 1 week.
Test and validate after each component/phase instead of waiting until the end. Catches bugs early when they're cheap to fix, prevents cascade failures. Use when building complex systems with multiple components where components depend on each other.
Measure sub-microsecond latencies using architecture-specific timestamp counters (TSC) and wait-free histograms. Essential for real-time systems, HFT, and latency-sensitive applications. Use when sub-microsecond latency requirements exist or when P99/P95 percentile analysis matters.
Layer different test types (unit, property, concurrency, stress, benchmark) to catch different bug classes. Each level catches what others miss. Use for complex systems, concurrent/parallel code, or performance-critical applications.
| name | parallel-subagent-orchestration |
| description | Launch multiple specialized Claude agents simultaneously to maximize productivity. Achieves 3-5x speedup on independent tasks like benchmarking, documentation, and analysis. Use when you have multiple independent tasks that each take more than 1 minute. |
Launch multiple specialized Claude agents simultaneously to maximize productivity. Achieves 3-5x speedup on independent tasks like benchmarking, documentation, and analysis.
Good candidates:
Bad candidates:
Available agents:
| Agent Type | Best For | Max Concurrent |
|---|---|---|
Explore | Codebase analysis, file searches | 2-3 |
general-purpose | Benchmarks, examples, audits, docs | 3-4 |
Bash | Git operations, command execution | 1-2 |
Plan | Architecture design, planning | 1 |
Selection guide:
Codebase analysis? → Explore agent
Running commands? → general-purpose agent
Each prompt must:
CRITICAL: Use one message with multiple Task tool calls for true parallelism.
Wait for all agents to complete, then review results.
After agents complete:
Scenario: Need to validate codebase architecture, performance, examples, and security
Agents launched (4 in parallel):
Explore Agent - Codebase architecture analysis
General-Purpose Agent - Run benchmarks
General-Purpose Agent - Generate examples
General-Purpose Agent - Security audit
Results:
Scenario: Compare 3 different queue implementations
Agents launched (3 in parallel):
Agent 1: Analyze crossbeam-queue Agent 2: Analyze tokio mpsc Agent 3: Analyze custom lock-free queue
Each agent produces:
Result: Comparison table in 10 minutes vs 30+ minutes sequential
Scenario: Need README, API docs, examples, and architecture docs
Agents launched (4 in parallel):
Agent 1: Write README.md (getting started, install, basic usage) Agent 2: Generate API documentation from code Agent 3: Create examples/ directory with 5 examples Agent 4: Write ARCHITECTURE.md (system design, data flow)
Result: Complete documentation suite in 15 minutes
Wrong:
Message 1: Task tool call for Agent 1
[wait for result]
Message 2: Task tool call for Agent 2
[wait for result]
Correct:
Message 1: Task tool calls for Agent 1, 2, 3, 4 (all in one message)
[all run in parallel]
Wrong:
Agent 1: Analyze codebase and save to /tmp/analysis.txt
Agent 2: Read /tmp/analysis.txt and write report
Correct:
Agent 1: Analyze codebase and write ANALYSIS.md
Agent 2: Run benchmarks and write BENCHMARKS.md
(No dependency between them)
Wrong:
"Look at the code and tell me about performance"
Correct:
"Run all benchmarks in benches/ directory:
1. Execute each with cargo bench
2. Extract P50/P99 latencies
3. Compare against targets in README
4. Create BENCHMARKS.md with results table"
Agent 1 (Explore): Analyze existing authentication system
Agent 2 (Explore): Find all security vulnerabilities
Agent 3 (general-purpose): Draft secure auth implementation plan
Then (after review): Implement based on findings
Agent 1: Create unit tests for module A
Agent 2: Create unit tests for module B
Agent 3: Create integration tests
Agent 4: Create property tests
Result: Full test suite in fraction of time
Agent 1: Build and test on Linux
Agent 2: Build and test on macOS
Agent 3: Build and test on Windows
Agent 4: Run cross-compilation tests
(Note: Requires appropriate build environments)
Before submitting PR:
Agent 1: Run all tests + generate coverage report
Agent 2: Run linters + format checks
Agent 3: Run benchmarks + compare to main
Agent 4: Generate changelog from commits
Results ready in minutes instead of running sequentially
Parallel agents can pre-validate before pushing:
Agent 1: Security scan
Agent 2: Performance regression check
Agent 3: Documentation check
Agent 4: License compliance
Push only if all pass
Skill Version: 1.0 Last Updated: 2025-01-06