Session Retrospective

A structured methodology for extracting actionable improvements from any Claude Code session. Based on After Action Review principles combined with AI-specific metacognitive learning patterns.

When to Use This Skill

• After any session longer than 30 minutes
• When significant friction was encountered
• After completing a complex multi-step task
• When a new pattern or approach was discovered
• Periodically (weekly/monthly) for continuous improvement

Core Principles

1. Ground Truth: Those who experienced the session directly are best able to understand its significance
2. Forward Focus: Think about what happened in context of what will happen next time
3. Iteration: Keep retrospectives short and actionable—blur the line between learning and doing
4. Lesson Learned ≠ Lesson Observed: If there's no change, there's been no learning

---

Phase 0: Scope Selection

Before beginning, clarify what type of improvements to focus on.

Present to user:

What would you like to improve from this session?

1. PROCESS   - How we work together (applicable to all future sessions)
2. PROJECT   - The specific project/skill/codebase we worked on
3. BOTH      - Extract both process and project improvements

Select [1-3]:

Store selection for Phase 5 filtering.

---

Phase 1: Session Review (Ground Truth)

Objective: Capture what actually happened while memories are fresh.

1.1 Successes

Ask the user (or infer from conversation if user prefers):

What went WELL in this session?
- Tasks completed successfully
- Efficient approaches discovered
- Good collaboration moments
- Problems solved elegantly

(List 3-5 items, or say "infer from conversation")

1.2 Friction Points

What caused FRICTION in this session?
- Repeated attempts needed
- Confusion or miscommunication
- Slow or tedious operations
- Permission prompts or interruptions
- Context limits hit
- Errors encountered

(List 3-5 items, or say "infer from conversation")

1.3 Surprises

What was UNEXPECTED (good or bad)?
- Behaviors that differed from expectations
- Discoveries made along the way
- Assumptions that proved wrong

(List 1-3 items, or say "infer from conversation")

---

Phase 2: Evidence Gathering

Objective: Collect concrete evidence to support observations.

2.1 Conversation Analysis

Review the session for:

• Tool calls that failed or required multiple attempts
• User messages expressing frustration or confusion
• Repeated patterns (good or bad)
• Time-consuming operations
• Successful approaches worth capturing

2.2 Artifact Review

Check for relevant outputs:

• Error messages and stack traces
• Tool output that was verbose or truncated
• Files created or modified
• Commands that were blocked or required confirmation

2.3 Categorize Findings

Organize evidence into:

Category	Evidence Type	Example
Efficiency	Time sinks, redundant operations	"Ran same search 3 times"
Quality	Errors, incomplete results	"Missing validation caused bug"
UX	Friction, confusion, interruptions	"12 permission prompts"
Knowledge	Missing info, wrong assumptions	"Didn't know API existed"
Architecture	Structural issues, design problems	"Sequential when parallel possible"

---

Phase 3: External Research

Objective: Gather best practices and prior art to inform recommendations.

3.1 Identify Research Topics

Based on friction points and categories, identify 2-4 research topics:

• If efficiency issues → search for optimization patterns
• If quality issues → search for testing/validation approaches
• If UX issues → search for permission/confirmation patterns
• If knowledge gaps → search for documentation/learning resources
• If architecture issues → search for design patterns

3.2 Execute Research

Use the ai-dev-research skill or direct web searches:

Research: [topic] best practices 2025
Focus on: production patterns, lessons learned, authoritative sources

3.3 Synthesize Findings

For each research topic, capture:

• Key insight with citation
• How it applies to observed friction
• Concrete recommendation derived from research

---

Phase 4: Cross-Cutting Analysis

Objective: Find overlaps, dependencies, and unified improvements.

4.1 Map Relationships

For each potential improvement, ask:

• Does this overlap with another improvement?
• Does this depend on another improvement?
• Does this enable other improvements?
• Is this the same idea at a different level (process vs. project)?

4.2 Cluster Related Items

Group improvements that:

• Address the same root cause
• Must be implemented together
• Form a coherent "system" (e.g., parallel execution + artifact storage + detailed prompts)

4.3 Identify Dependencies

Create dependency graph:

A ──depends on──> B
C ──enables──> D
E ──same as──> F (different levels)

---

Phase 5: Prioritization

Objective: Rank improvements by combined impact.

5.1 Scoring Criteria

Criterion	Weight	Description
Impact	40%	How much improvement if implemented?
Frequency	25%	How often will this help?
Effort	20%	How hard to implement? (inverse)
Dependencies	15%	Does this enable other improvements?

5.2 Scope Filtering

Based on Phase 0 selection:

• PROCESS: Include only user-level improvements (applicable to any project)
• PROJECT: Include only project-specific improvements
• BOTH: Include all, but tag each with scope

5.3 Produce Ranked List

Create unified top N list (typically 5-10 items):

Rank	Improvement	Scope	Impact	Effort	Dependencies
1	...	...	...	...	...

---

Phase 6: Actionable Output

Objective: Produce implementation-ready recommendations.

6.1 Document Structure

Create a document with:

# Session Retrospective: [Date] - [Topic/Project]

## How to Use This Document
[Instructions for future agent sessions]

## Executive Summary
[2-3 sentence overview]

## Context
- Session date: [date]
- Duration: [approximate]
- Primary task: [what was being done]
- Scope: [PROCESS/PROJECT/BOTH]

## Key Findings
### Successes
- [item with evidence]

### Friction Points
- [item with evidence]

## Research Insights
[Key findings from Phase 3 with citations]

## Unified Recommendations (Ranked)
[Table from Phase 5]

## Implementation Menu
[Selectable items with effort estimates and dependencies]

## Implementation Specifications
[For each menu item: files to modify, steps, acceptance criteria]

6.2 Implementation Specifications

For each recommendation, include:

• Files to modify/create
• Implementation steps (numbered, specific)
• Acceptance criteria (checkboxes)
• Dependencies (which other items must come first)

6.3 Save Location

Ask user:

Where should I save the retrospective document?

1. ~/Documents/retrospectives/[date]-[topic].md
2. [Project]/.claude/retrospectives/[date].md
3. Custom path
4. Display only (don't save)

Select [1-4]:

---

Phase 7: Verification

Objective: Ensure the retrospective will lead to actual change.

7.1 Actionability Check

For each recommendation, verify:

• Specific enough to implement without re-research?
• Clear acceptance criteria?
• Dependencies identified?
• Effort estimated?

7.2 User Confirmation

I've identified [N] improvements ranked by impact.

Top 3:
1. [name] - [scope] - [effort]
2. [name] - [scope] - [effort]
3. [name] - [scope] - [effort]

Would you like to:
1. Review full document
2. Save and implement top item now
3. Save for later implementation
4. Revise (specify what to change)

Select [1-4]:

7.3 Backlog Management

After implementation decisions, handle deferred items:

For each item NOT implemented:

1. Add to improvement_backlog.md with:
   • Unique ID (P-XXX for PROCESS, J-XXX for PROJECT)
   • Source link to this retrospective
   • Problem, proposed solution, acceptance criteria
   • Priority and effort from scoring

Backlog location: ~/Library/CloudStorage/GoogleDrive-sabreace33@gmail.com/My Drive/Obsidian/my-notes/ai_plans/

Template for backlog entry:

#### [P/J]-XXX: [Improvement Name]

**Added:** [date]
**Source:** [[retrospective-filename#Item N]]
**Priority:** [High/Medium/Low]
**Effort:** [Low/Medium/High]
**Dependencies:** [list or None]

**Problem:** [What friction or issue was encountered]

**Proposed Solution:** [What should be done]

**Acceptance Criteria:**
- [ ] [criterion 1]
- [ ] [criterion 2]

7.4 Update Improvement Log

After implementing improvements:

1. Add entry to improvement_log.md with full details
2. Update retrospective document to mark items as completed
3. Move any completed backlog items to the archive section

---

Anti-Patterns to Avoid

Anti-Pattern	Why It's Bad	Better Approach
Vague recommendations	"Improve error handling" → no action	"Add try/catch to X function with specific error types"
No evidence	Opinions without support	Cite specific moments from session
No research	Reinventing known solutions	Check existing best practices first
No prioritization	Everything seems important	Force-rank by impact × frequency
No implementation spec	Future self won't know what to do	Include files, steps, acceptance criteria
No follow-through	"Lesson observed, not learned"	Offer to implement top item immediately

---

Example Invocations

# After a skill development session
"Let's do a session retrospective focused on the process"

# After a debugging session
"What did we learn from this debugging session? Extract project improvements."

# Periodic review
"Run a retrospective on our last few sessions - what patterns should we change?"

# Immediate action
"Retrospective with immediate implementation of top improvement"

---

Integration with Other Skills

This skill works well with:

• ai-dev-research: For Phase 3 external research
• research-verification: Apply verification checklist during research
• parallel-first-design-guide: If architecture issues identified
• skill-creator: If creating new skills based on learnings

Related Documents

Maintain these companion documents:

Location: ~/Library/CloudStorage/GoogleDrive-sabreace33@gmail.com/My Drive/Obsidian/my-notes/ai_plans/

Document	Purpose
improvement_log.md	Running log of completed improvements with full context
improvement_backlog.md	Queue of deferred improvements awaiting implementation

Workflow:

1. Retrospective identifies improvements
2. Implemented items → improvement_log.md
3. Deferred items → improvement_backlog.md
4. Future sessions check backlog for relevant items

---

Metadata

Version: 1.1.0
Created: 2025-01-18
Updated: 2025-01-20

Changelog:

• 1.1.0 (2025-01-20): Added backlog management (7.3), improvement log integration (7.4), related documents section, additional triggers
• 1.0.0 (2025-01-18): Initial release

Based on:

• After Action Review methodology (US Army, 1980s)
• Metacognitive learning framework (arXiv:2506.05109)
• HealthFlow self-evolving agent patterns (arXiv:2508.02621)
• Zimmerman's self-regulation model (planning, execution, reflection)

Sources:

• BetterEvaluation: After Action Review
• Beam AI: Self-Learning Agents
• Monday.com: After Action Reports 2025
• Anthropic: Multi-Agent Research System