Blueprint Optimize - Dual-Layer Memory Crystallization

Overview

Complete the iteration cycle by syncing status, committing properly, and permanently capturing hard-won lessons in the dual-layer memory system:

•./postmortem/ = Deep case history (detailed analysis)
•./CLAUDE.md = High-frequency operating rules (concise directives)

Core Principle

code

./postmortem/ = 深度病历库 (Deep Case History)
                → 完整的事故分析、root cause、时间线、详细修复
                → 无行数限制，越详细越好

./CLAUDE.md   = 高频手术准则 (High-Frequency Operating Rules)
                → 只有指令和警告，无解释
                → 严格 ≤ 100 行

Rule: 先写病历 (postmortem)，再提炼准则 (CLAUDE.md)

When to Use

•PR has been merged (after blueprint-review)
•Ready to close out the iteration
•User mentions "上线", "merge complete", "close ticket", or "记录经验"
•After any incident, bug, or significant learning

Do NOT use when:

•PR still under review (use blueprint-review)
•Still implementing (use blueprint-develop)

Prerequisites

Before invoking this skill, verify:

• PR merged to main branch
• All tests passing on main
• Linear ticket exists and is linked to PR

Workflow

dot

digraph blueprint_optimize {
    rankdir=TB;
    node [shape=box];

    start [label="Start: PR Merged" shape=ellipse];
    sync [label="1. Status Sync\n(Linear MCP)"];
    commit [label="2. Final Commit\n(commit-commands)"];
    pm_decision [label="3. PM Decision\n(Need New Postmortem?)" shape=diamond style=filled fillcolor=lightyellow];
    pm_write [label="3a. Write Postmortem\n(./postmortem/)" style=filled fillcolor=lightyellow];
    claude_preview [label="4. CLAUDE.md Preview\n(Extract from PM)" style=filled fillcolor=lightyellow];
    confirm [label="5. Human Confirms" shape=diamond];
    claude_write [label="6. Write to CLAUDE.md"];
    done [label="Iteration Complete" shape=ellipse];

    start -> sync;
    sync -> commit;
    commit -> pm_decision;
    pm_decision -> pm_write [label="Yes: Incident/Bug Found"];
    pm_decision -> claude_preview [label="No: Minor Learnings Only"];
    pm_write -> claude_preview;
    claude_preview -> confirm;
    confirm -> claude_write [label="Approved"];
    confirm -> claude_preview [label="Revise"];
    claude_write -> done;
}

Phase 1: Status Sync

Goal: Mark Linear ticket as Done and ensure all links are in place.

Step 1.0: Gather Context (if needed)

When syncing status or analyzing merged changes:

code

Use Skill tool: context7:context7
Purpose: Gather context about merged PR, related tickets, and impact
Focus: PR changes, related code, dependencies, PM context

Update Linear Ticket

code

Tool: mcp__plugin_linear_linear__update_issue

Parameters:
- id: [ticket_id]
- state: "Done"

Verification Checklist

Item	Action
Ticket state → Done	`update_issue` with state
PR link attached	Verify via `get_issue`
Sub-tickets closed	Update each child ticket
Labels updated	Add "shipped" or equivalent

Comment Summary

Add closing comment to ticket:

code

Tool: mcp__plugin_linear_linear__create_comment

Body:
✅ Strategy shipped to main
- PR: [PR link]
- Backtest Sharpe: X.XX
- PM: [PM-XXX if created, or "None"]
- CLAUDE.md: [Updated/No change]

Output: Linear ticket marked Done with closing comment.

Phase 2: Final Commit (if needed)

Goal: Ensure any final changes follow commit conventions.

Invoke commit-commands Skill

code

Use Skill tool: commit-commands:commit
Purpose: Generate standardized commit message

Commit Message Format

code

<type>(<scope>): <subject>

<body>

<footer>

Co-Authored-By: Claude Code <noreply@anthropic.com>

Output: Clean commit following project conventions.

Phase 3: Postmortem Decision & Creation

Goal: Determine if a new postmortem is needed, and create if yes.

Decision Matrix

Condition	Create PM?	PM Severity
Bug caused incorrect trades	YES	CRITICAL
API issue discovered	YES	HIGH
Performance degradation found	YES	MEDIUM
Edge case hit in backtest	YES	MEDIUM
Code worked but pattern was suboptimal	NO	-
Simple refactor, no issues	NO	-

Postmortem Template

Create file: ./postmortem/PM-[NNN]-[brief-description].md

markdown

# PM-[NNN]: [Incident Title]

## Metadata
- **Date:** [YYYY-MM-DD]
- **Severity:** CRITICAL | HIGH | MEDIUM | LOW
- **Strategy:** [Strategy Name]
- **Linear Ticket:** [LIN-XXX]
- **Author:** [Name/Claude Code]

## Summary
[One paragraph describing what happened]

## Timeline
- [HH:MM] [Event 1]
- [HH:MM] [Event 2]
- [HH:MM] [Detection/Discovery]
- [HH:MM] [Resolution]

## Root Cause Analysis

### What Happened
[Detailed technical explanation]

### Why It Happened
[5 Whys analysis or similar]

### Contributing Factors
- [Factor 1]
- [Factor 2]

## Impact
- **Financial:** [Estimated loss/gain, if any]
- **Operational:** [Service disruption, if any]
- **Data:** [Data corruption/loss, if any]

## Resolution
[What was done to fix the immediate issue]

## Mitigations Implemented

### Code Changes
```python
# BEFORE (anti-pattern)
[code snippet]

# AFTER (correct pattern)
[code snippet]

Tests Added

•tests/pm_regression/test_pm_[NNN].py

Configuration Changes

•[Config change 1]

Lessons Learned

What Went Well

•[Item 1]

What Went Wrong

•[Item 1]

Action Items

• [Action 1] - Owner: [Name]
• [Action 2] - Owner: [Name]

Keywords

[keyword1] [keyword2] [keyword3]

•Previous: PM-[XXX] (if related)
•See also: [Link to docs/code]

code


### PM Naming Convention

PM-001-order-duplication.md PM-002-api-timeout-cascade.md PM-003-liquidity-trap-election.md

code


**Output:** Postmortem file created (or decision to skip documented).

### 🔄 Ralph Loop: Iterative PM Refinement (Optional)

**When to Activate:** Complex incidents requiring thorough root cause analysis.

For significant incidents, use Ralph Loop to iteratively refine the postmortem quality:

Skill: ralph-loop:ralph-loop

Prompt Template: "Refine Postmortem PM-[NNN] for quality. Incident: [brief description]

Quality Checklist:

• Root cause identified (not just symptoms)
• 5 Whys analysis complete
• Code snippets include before/after
• Action items are specific and assigned
• Keywords comprehensive for future search

Current gaps: [list] Output <promise>PM REFINED</promise> when all checklist items complete."

Options: --max-iterations 5 --completion-promise "PM REFINED"

code


### Ralph Loop PM Refinement Cycle

```dot
digraph ralph_pm {
    rankdir=LR;
    node [shape=box];

    start [label="Draft PM\nCreated" shape=ellipse];
    check [label="Quality Check"];
    gaps [label="Gaps Found?" shape=diamond];
    refine [label="Refine Section"];
    done [label="<promise>\nPM REFINED" shape=ellipse style=filled fillcolor=lightgreen];

    start -> check;
    check -> gaps;
    gaps -> refine [label="Yes"];
    refine -> check;
    gaps -> done [label="No"];
}

PM Quality Checklist for Ralph Loop

Section	Quality Criteria
Root Cause	Specific, technical, not "human error"
5 Whys	Chain leads to systemic fix, not blame
Code Snippets	Actual code, not pseudocode
Timeline	Timestamped, includes detection method
Action Items	SMART: Specific, Measurable, Assigned
Keywords	≥5 relevant search terms

When to Use Ralph Loop for PM

Situation	Use Ralph Loop?
CRITICAL severity incident	✅ Yes - thorough analysis needed
First incident of its type	✅ Yes - establish good template
Similar to existing PM	❌ No - reference existing, add delta
Simple config error	❌ No - straightforward PM sufficient

Phase 4: CLAUDE.md Preview Generation

Goal: Extract high-frequency rules from postmortem (or learnings) for CLAUDE.md.

Extraction Rules

From Postmortem	To CLAUDE.md
Root Cause	⚠️ WARNING (one line)
Code Fix Pattern	📋 DIRECTIVE (one line)
Parameter Value	🔢 PARAMETER (one line)
Timeline Detail	❌ OMIT
5 Whys Analysis	❌ OMIT
Impact Assessment	❌ OMIT

CRITICAL: Human Review Required

DO NOT write directly to CLAUDE.md. Generate preview first.

Preview Template

markdown

---
## 📝 Proposed Addition to CLAUDE.md

### Source: PM-[NNN] - [Title] ([Date])

#### ⚠️ WARNINGS
- [Extracted from Root Cause - ONE LINE, specific values]

#### 📋 DIRECTIVES
- [Extracted from Resolution - ONE LINE, actionable]

#### 🔢 PARAMETERS
- [Parameter]: [Value] ([brief reason])

---

**Estimated CLAUDE.md impact:** +X lines (current: Y, after: Z)
**Total after update:** Z/100 lines

⚡ Please review the above. Reply "确认" to write, or provide edits.

Quality Checklist for Preview

Check	Example Pass	Example Fail
Specific values	"429 after 50 req/min"	"API has rate limits"
Actionable	"Use decimal.Decimal"	"Be careful with floats"
One line max	"Check market.active before orders"	[Multi-line explanation]
No prose	Just the rule	"Because we found that..."

Output: Markdown preview displayed to user, awaiting confirmation.

Phase 5: Human Confirmation

Goal: Ensure human approves before modifying project memory.

Wait for User Response

Response	Action
"确认" / "confirm" / "yes"	Proceed to Phase 6
Edits provided	Regenerate preview with changes
"取消" / "cancel"	Skip CLAUDE.md update

Revision Handling

If user provides edits:

•Incorporate feedback
•Regenerate preview
•Show updated preview
•Wait for confirmation again

Output: User confirmation received.

Phase 6: Write to CLAUDE.md

Goal: Append high-frequency rules to project memory file.

Pre-Write Line Count Check

bash

wc -l CLAUDE.md
# MUST check before writing

100 Line Limit Enforcement

If exceeding 100 lines after addition:

markdown

⚠️ CLAUDE.md will exceed 100 lines after this addition.

Current: 92 lines
Adding: 15 lines
Total: 107 lines (7 over limit)

### Pruning Candidates (Oldest First)

1. **Lines 5-12:** Mean-Reversion v1 (2023-06-01) - 8 lines
   - Superseded by PM-015? [Y/N]

2. **Lines 13-20:** Early API Learnings (2023-07-15) - 8 lines
   - Still relevant? [Y/N]

### Options:
1. Prune selected entries
2. Condense new content further
3. Move old entries to postmortem archive
4. Proceed anyway (NOT RECOMMENDED)

Which entries should be pruned?

Write Rules

Rule	Implementation
Append only	Add at end, never delete without approval
100 line HARD limit	Must prune if exceeding
Preserve structure	Match existing formatting
Date stamp	Always include date for each entry
PM Reference	Always link to source postmortem

Write Action

code

Tool: Edit or Write (append mode)
File: CLAUDE.md
Content: [Approved preview content]

Post-Write Verification

bash

# Verify line count
wc -l CLAUDE.md
# MUST be ≤ 100

Output: CLAUDE.md updated, line count confirmed.

Quick Reference

Phase	Tool/Skill	Key Action
1. Context	`context7:context7` (if needed)	Gather context about merged PR
1. Sync	`mcp__plugin_linear_linear__update_issue`	Mark ticket Done
2. Commit	`commit-commands:commit`	Standardized commit
3. PM	`Write` to `./postmortem/`	Create if incident found
3b. 🔄 Ralph	`ralph-loop:ralph-loop`	Iterative PM refinement (optional)
4. Preview	Manual generation	Extract rules from PM
5. Confirm	`AskUserQuestion` or wait	Human approval required
6. Write	`Edit` to `CLAUDE.md`	Append, enforce 100 line limit

Dual-Layer Memory Architecture

code

┌─────────────────────────────────────────────────────────────┐
│                    INCIDENT OCCURS                           │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│  ./postmortem/PM-XXX.md (深度病历库)                          │
│  ─────────────────────────────────────────────────────────  │
│  • Full timeline                                            │
│  • Root cause analysis (5 Whys)                             │
│  • Code snippets (before/after)                             │
│  • Impact assessment                                        │
│  • Action items                                             │
│  • NO LINE LIMIT - be thorough                              │
└─────────────────────────────────────────────────────────────┘
                            │
                            │ EXTRACT
                            ▼
┌─────────────────────────────────────────────────────────────┐
│  ./CLAUDE.md (高频手术准则)                                   │
│  ─────────────────────────────────────────────────────────  │
│  • ⚠️ WARNING: [one line]                                   │
│  • 📋 DIRECTIVE: [one line]                                 │
│  • 🔢 PARAMETER: [value]                                    │
│  • STRICT 100 LINE LIMIT                                    │
│  • Reference: PM-XXX                                        │
└─────────────────────────────────────────────────────────────┘

Common Mistakes

Mistake	Prevention
Skipping postmortem for incidents	If there was a bug/issue, PM is MANDATORY
Verbose CLAUDE.md entries	Extract only: warning, directive, parameter
Exceeding 100 lines	Prune old entries, ask user to choose
No PM reference in CLAUDE.md	Always link source: "See: PM-XXX"
Writing CLAUDE.md without preview	ALWAYS generate preview, wait for "确认"
Missing specifics	"API is slow" ❌ → "API 429s after 50 req/min" ✅
Forgetting Linear sync	First action is always status update

Exit Criteria

Before marking iteration complete:

• Linear ticket state = Done
• Closing comment added with summary
• Final commit follows conventions
• Postmortem created (if incident occurred)
• Lesson preview shown to user
• User explicitly confirmed ("确认")
• CLAUDE.md updated with PM reference
• Line count ≤ 100 verified

Example Invocation

Standard Mode

code

User: "PR merged, close out the mean-reversion iteration. We hit a rate limit bug during testing."

Claude:
1. [Sync] Updating Linear ticket LIN-123 → Done ✅
   Added closing comment

2. [Commit] No pending changes, main branch clean ✅

3. [PM Decision] Rate limit bug found → Creating postmortem

   Creating: ./postmortem/PM-007-rate-limit-429.md
   - Severity: HIGH
   - Root Cause: No delay between API calls
   - Resolution: Added 1.2s minimum delay
   - Tests Added: tests/pm_regression/test_pm_007.py ✅

4. [Preview] Extracting rules from PM-007...

---
## 📝 Proposed Addition to CLAUDE.md

### Source: PM-007 - Rate Limit 429 (2024-01-15)

#### ⚠️ WARNINGS
- Polymarket API returns 429 after 50 req/min without warning header

#### 📋 DIRECTIVES
- Implement minimum 1.2s delay between all API calls
- Parse X-RateLimit-Remaining header when available

#### 🔢 PARAMETERS
- API_MIN_DELAY: 1.2s (discovered empirically, see PM-007)

---

**Estimated CLAUDE.md impact:** +8 lines (current: 52, after: 60)
**Total after update:** 60/100 lines ✅

⚡ Please review the above. Reply "确认" to write, or provide edits.

User: "确认"

Claude:
5. [Write] Appending to CLAUDE.md... ✅
   Line count verified: 60/100 ✅

Iteration complete:
- Linear: LIN-123 Done
- Postmortem: PM-007-rate-limit-429.md created
- CLAUDE.md: +8 lines (total: 60/100)
- Ready for next strategy

🔄 Ralph Loop Mode (PM Refinement)

code

User: "PR merged. 有个严重的 order cascade 事故需要详细分析"

Claude:
1-2. [Standard Phases Complete] ✅

3. [PM Decision] CRITICAL severity → Creating postmortem
   Creating: ./postmortem/PM-008-order-cascade-failure.md

   Initial draft quality check:
   - Root cause: ❌ Too vague ("system overload")
   - 5 Whys: ❌ Incomplete (only 2 levels)
   - Code snippets: ⚠️ Missing before/after
   - Action items: ❌ Not specific

   CRITICAL incident → Activating Ralph Loop for thorough analysis

🔄 [Ralph Loop: PM Refinement]
   Invoking: /ralph-loop "Refine PM-008 for quality.
   Incident: Order cascade causing $X loss
   Gaps: root cause vague, 5 whys incomplete, no code snippets
   Output <promise>PM REFINED</promise> when all sections complete."
   --max-iterations 5
   --completion-promise "PM REFINED"

   Iteration 1: Root cause refined → "Unbounded retry loop in error handler"
   Iteration 2: 5 Whys complete → leads to "missing circuit breaker"
   Iteration 3: Code snippets added with before/after
   Iteration 4: Action items made SMART, owners assigned

   <promise>PM REFINED</promise>

🔄 [Ralph Loop Complete]
   PM Quality: All checklist items ✅
   - Root cause: Specific technical issue identified
   - 5 Whys: Complete chain to systemic fix
   - Code: Actual snippets with fix
   - Actions: 4 SMART items with owners

4-6. [Continuing to CLAUDE.md preview and write...]

Output: High-quality PM-008 created via iterative refinement

Blueprint Optimize - Dual-Layer Memory Crystallization

Overview

Core Principle

When to Use

Prerequisites

Workflow

Phase 1: Status Sync

Step 1.0: Gather Context (if needed)

Update Linear Ticket

Verification Checklist

Comment Summary

Phase 2: Final Commit (if needed)

Invoke commit-commands Skill

Commit Message Format

Phase 3: Postmortem Decision & Creation

Decision Matrix

Postmortem Template

Tests Added

Configuration Changes

Lessons Learned

What Went Well

What Went Wrong

Action Items

Keywords

Related

PM Quality Checklist for Ralph Loop

When to Use Ralph Loop for PM

Phase 4: CLAUDE.md Preview Generation

Extraction Rules

CRITICAL: Human Review Required

Preview Template

Quality Checklist for Preview

Phase 5: Human Confirmation

Wait for User Response

Revision Handling

Phase 6: Write to CLAUDE.md

Pre-Write Line Count Check

100 Line Limit Enforcement

Write Rules

Write Action

Post-Write Verification

Quick Reference

Dual-Layer Memory Architecture

Common Mistakes

Exit Criteria

Example Invocation

Standard Mode

🔄 Ralph Loop Mode (PM Refinement)