Experience Synthesis — Learning from Claude's Past

Insight Taxonomy

When analyzing conversations, extract insights in these categories (ordered by durability — values first):

1. Learned Values

Comparative preferences and priority orderings — "X is better than Y".

Signal sources:

• User states "prefer X over Y", "X is better than Y", "X matters more than Y"
• User confirms a comparative statement from Claude ("yes, readability > cleverness")
• Trade-off discussions that resolve into a clear preference
• Statements with "prioritize", "choose X over Y", "the most important thing is"

What to capture: The value choice, both sides of the comparison, and the reasoning. Values are the most durable type of memory — they survive codebase rewrites.

2. Decisions

What was chosen, why, and what alternatives were rejected.

Signal sources:

• AskUserQuestion tool calls + the user's response in the next tool_result
• Plan mode content (planContent field on user records, ExitPlanMode tool calls)
• Thinking blocks where Claude weighs options
• Assistant text containing "I'll use X instead of Y because..."

What to capture: The decision, the rationale, the alternatives considered, and the context (what problem it solved).

3. Mistakes & Corrections

What went wrong, root cause, and how it was fixed.

Signal sources:

• Tool results with is_error: true
• Bash results containing: error, failed, FAIL, exit code 1, stack trace, traceback, Exception
• User corrections: "no, that's wrong", "revert that", "that broke X"
• Retry patterns: same tool called 2+ times on the same target with different inputs
• Reverted file edits (same file edited, then edited back)

What to capture: What failed, why it failed, what fixed it, how to avoid it next time.

4. Effective Patterns

Approaches that worked well and could be reused.

Signal sources:

• Successful test runs (Bash results with: passed, PASS, success, 0 errors)
• Successful builds (built in, compiled, no errors)
• PR creation (pr-link records)
• Git commits (successful completion of work)
• User satisfaction signals: "perfect", "great", "exactly what I needed"

What to capture: The approach, when it applies, why it worked.

5. Anti-patterns

Approaches that failed, were abandoned, or caused problems.

Signal sources:

• Sequences where multiple attempts fail before finding a working solution
• Sessions with many compactions (long, possibly struggling sessions)
• Tool calls that were user-interrupted
• Files edited many times in one session (version count > 3 in file-history-snapshot)

What to capture: What was tried, why it failed, what worked instead.

6. User Preferences

The user's preferred tools, styles, and workflows.

Signal sources:

• Tool usage frequency across sessions (which tools does the user/Claude use most?)
• Model choices in sessions (opus vs sonnet vs haiku patterns)
• Permission mode patterns
• Common first prompts (workflow entry points)
• Branch naming conventions

What to capture: The preference, evidence across sessions, strength of pattern.

7. Architecture Knowledge

System design decisions, component relationships, tech stack details.

Signal sources:

• Early messages in sessions (problem descriptions, requirement discussions)
• Plan mode content (architectural plans)
• File paths touched across sessions (reveals project structure)
• Dependencies installed or configured

What to capture: Component, its role, relationships, key decisions about it.

8. Recurring Problems

Issues that keep coming back across sessions.

Signal sources:

• Similar error messages appearing in different sessions
• Same files being edited for fixes repeatedly
• Similar user prompts ("fix X again", "the Y bug is back")

What to capture: The problem, frequency, root cause pattern, whether it has a permanent fix.

9. Performance & Cost Patterns

Token usage trends, session efficiency, cost optimization opportunities.

Signal sources:

• sd-recall.py stats output across sessions: token totals, compaction counts
• Turn duration from system records with subtype: turn_duration
• Cache hit ratios (cache_read vs cache_creation tokens)
• Model selection patterns (when opus vs sonnet is used)

What to capture: Trends, outliers, optimization opportunities.

Synthesis Methodology

When extracting insights, follow this process:

1. Gather: Use sd-recall.py sessions to identify relevant sessions, then sd-recall.py search <keyword> for detail
2. Trend check: Run /trend to see if a pattern is increasing/decreasing over time
3. Identify: Look for signals from the taxonomy above
4. Cross-reference: Check if a pattern appears in multiple sessions (stronger signal). Use /optimize for automated pain-point detection
5. Contextualize: Combine with git history when available (what code resulted from the decision/mistake?)
6. Synthesize: Produce actionable insights, not just observations

Output Format

Present insights in a structured format:

## [Category]: [Brief Title]
**Sessions:** [session IDs or dates]
**Context:** [What was happening]
**Insight:** [The key learning]
**Evidence:** [Specific data points]
**Action:** [What to do differently / what to keep doing]

Insight Quality Criteria

Good insights are:

• Specific: "Use --follow with git log when tracing renamed files" not "git is useful"
• Actionable: Something you can apply in future sessions
• Evidenced: Backed by data from multiple sessions when possible
• Contextual: Includes when the insight applies (and when it doesn't)

Skip insights that are:

• Generic programming advice (not specific to this user/project)
• One-time occurrences that aren't likely to recur
• Already documented in the project's CLAUDE.md or README

上游（0.9.18 起优先）

下游协作

DO NOT

• 从零创建新 skill → skill-creator
• 需要解析 JSONL 原始数据 → jsonl-core（synthesis 只做提炼，不做解析）
• 需要管理 MEMORY.md 文件 → memory-management