knowledge-distillation

name	knowledge-distillation
description	Extracts reusable patterns and lessons from execution history and project experience. Generates structured knowledge entries for the knowledge base.
tags	["knowledge","distill","extract","lesson","pattern","experience","postmortem"]
stage	evaluation
type	delivery-skill
category	KNOWLEDGE
scope	PLATFORM
version	1.0
author	forge-team

Knowledge Distillation Skill

Purpose

Analyzes execution history, session summaries, and project artifacts to extract reusable patterns, lessons learned, and knowledge entries. This corresponds to the asset-extractor component in the Forge learning loop architecture.

When to Use

• User says "总结经验" / "提炼 pattern" / "写入知识库"
• After a series of bug fixes to identify recurring patterns
• When a complex task is completed and lessons should be captured
• For postmortem analysis after incidents

Distillation Process

Step 1: Gather Source Material

1. Use get_session_history to retrieve recent session summaries
2. Search workspace memory for accumulated facts
3. Identify failure→fix chains in execution history
4. Read any relevant project docs or changelogs

Step 2: Pattern Recognition

1. Recurring Problems: Same type of error appearing multiple times
2. Effective Solutions: Patterns that consistently resolve issues
3. Convention Gaps: Implicit rules not yet documented
4. Tool Usage Patterns: Which tools work best for which tasks
5. Architectural Insights: Design decisions and their outcomes

Step 3: Structure the Knowledge

Experience Record Template

## [Title]: [Brief description of the pattern/lesson]

### Context
- When does this apply?
- What conditions trigger this?

### Problem
- What goes wrong without this knowledge?
- What are the symptoms?

### Solution
- Step-by-step resolution
- Code examples if applicable

### Prevention
- How to avoid this in the future
- Suggested Skill/Baseline updates

### Evidence
- Session(s) where this was observed
- Number of occurrences

ADR (Architecture Decision Record) Template

## ADR-XXX: [Decision Title]

### Status
Accepted | Proposed | Deprecated

### Context
[What situation prompted this decision?]

### Decision
[What was decided?]

### Consequences
[Positive and negative impacts]

### Alternatives Considered
[What else was evaluated?]

Runbook Template

## Runbook: [Operation Name]

### Prerequisites
- [Required access/tools]

### Steps
1. [Step with exact commands]
2. [Step with expected output]

### Troubleshooting
| Symptom | Cause | Fix |
|---------|-------|-----|
| ... | ... | ... |

### Rollback
[How to undo if something goes wrong]

Step 4: Suggest Improvements

1. Skill Updates: New rules or patterns to add to Foundation Skills
2. Baseline Additions: New quality checks to automate
3. Knowledge Base Entries: New docs to create
4. CLAUDE.md Updates: New "known pitfalls" to record

Output

Present the distilled knowledge in chat as structured Markdown. If the user requests, save to:

• Knowledge base: knowledge-base/[category]/[name].md
• Workspace docs: docs/[type]/[name].md

Integration

• Profile: evaluation-profile (primary)
• Tools used: get_session_history, search_knowledge, workspace_read_file
• Learning Loop: Feeds into skill-feedback-analyzer for continuous improvement