accelerate

分类: 工具与效率 | 上传者: sharpdeveye | 下载: 0 | 版本: v1.0（最新）

当工作流程太慢、太昂贵或两者兼而有之，并且需要延迟、成本或令牌使用优化时使用。

更新日志: Source: GitHub https://github.com/sharpdeveye/maestro

目录结构

当前层级: tree/main/source/skills/accelerate/

📄 SKILL.md 3.0 KB

SKILL.md

---
name: accelerate
description: "Use when the workflow is too slow, too expensive, or both and needs latency, cost, or token usage optimization."
argument-hint: "[target metric]"
category: enhancement
version: 1.3.1
user-invocable: true
---

## MANDATORY PREPARATION

Invoke {{command_prefix}}agent-workflow — it contains workflow principles, anti-patterns, and the **Context Gathering Protocol**. Follow the protocol before proceeding — if no workflow context exists yet, you MUST run {{command_prefix}}teach-maestro first.
Consult the context-management reference in the agent-workflow skill for window optimization and budget strategies.


---

Make the workflow faster and cheaper without sacrificing quality. Measure before and after.

### Performance Audit

Measure current performance:

```text
Current metrics:
  Latency (p50): ___ms
  Latency (p95): ___ms
  Cost per request: $___
  Token usage (avg): ___ input / ___ output
  Error rate: ___%
```

### Acceleration Strategies

**Reduce Token Usage**

- Shorten system prompts (remove redundant instructions)
- Compress few-shot examples to minimum viable length
- Use structured output schemas instead of verbose text
- Summarize context instead of passing raw documents
- Reduce output length requirements

**Model Cascading**

- Route simple tasks to cheaper/faster models
- Escalate only complex tasks to capable models
- Use classification to determine complexity

**Caching**

- Cache responses for identical or near-identical inputs
- Cache tool results with appropriate TTL
- Cache embeddings for frequently-queried documents
- Use semantic caching for similar (not identical) queries

**Parallelization**

- Run independent tool calls in parallel
- Run independent agent steps in parallel
- Use streaming to start processing before full response

**Context Optimization**

- Retrieve less, retrieve better (improve retrieval precision)
- Use context compression techniques
- Implement sliding window for long conversations

### Acceleration Report

For each optimization:

1. **What changed**: Specific modification
2. **Before**: Latency/cost/tokens before
3. **After**: Latency/cost/tokens after
4. **Quality impact**: Any quality change (verify with golden tests)
5. **Trade-off**: What was sacrificed for the improvement

### Acceleration Checklist

- [ ] Baseline metrics recorded before any changes
- [ ] Each optimization measured with before/after comparison
- [ ] Quality impact verified (golden tests still pass)
- [ ] Trade-offs documented for each change
- [ ] Cost/latency improvements quantified

### Recommended Next Step

After optimization, run `{{command_prefix}}evaluate` to verify quality didn't degrade, or `{{command_prefix}}iterate` to set up continuous monitoring.

**NEVER**:

- Optimize without measuring first (you need a baseline)
- Sacrifice quality for speed without explicit user approval
- Cache outputs that depend on real-time data
- Skip the quality check after optimization
- Optimize prematurely (make it correct first, then make it fast)

登录后下载/点赞/收藏 ❤ 76 | ★ 0

accelerate

目录结构

SKILL.md

举报内容

提示