AgentSkillsCN

improve

运行改进循环——分析失败原因，提炼持久化的经验教训。

SKILL.md
--- frontmatter
name: improve
description: Run the improvement loop - analyze failures and generate persistent learnings.

Run Improvement Loop

Analyze failed queries from an evaluation run and generate improvements.

Usage

code
/improve [run_id]

Process

  1. •Load Failures: Read failed evaluations from evals/results/<run_id>/
  2. •For Each Failure:
    • •Spawn improver agent with failure context
    • •Agent diagnoses root cause
    • •Agent generates appropriate improvement
    • •Agent writes to knowledge/
  3. •Verify: Check that improvements were written correctly
  4. •Report: Summarize what was learned

Workflow

code
/improve run_001
  │
  ├─> Load failures from evals/results/run_001/
  │
  ├─> For each failure:
  │     ├─> Analyze: What went wrong?
  │     ├─> Categorize: Example? Function? Doc?
  │     └─> Write: Update knowledge/
  │
  └─> Report: "Added 3 examples, 1 function, 2 doc entries"

Output

markdown
# Improvement Report: [run_id]

## Summary
- Failures analyzed: N
- Improvements generated: M
  - Examples: X
  - Functions: Y
  - Documentation: Z

## Changes Made

### knowledge/examples.md
- Added: [pattern name] for query [q_id]

### knowledge/functions.py
- Added: `function_name()` - [description]

### knowledge/schema.md
- Added: [data fact or edge case discovery]

## Next Steps
- Re-run evaluation to verify improvements
- Command: `/run-baseline test`

Files Read

  • •evals/results/<run_id>/*.json - Failed evaluation results
  • •knowledge/* - Existing knowledge (to avoid duplicates)

Files Written

  • •knowledge/examples.md - Worked examples
  • •knowledge/functions.py - Helper functions
  • •knowledge/schema.md - Schema/data discoveries