after-action-review - SKILL.md Agent Skill

name: after-action-review description: Structured post-goal reflection that extracts lessons, evaluates quality, and feeds the self-improvement loop author: Writing Secrets version: 1.0.0 triggers: - "after action review" - "review goal" - "post mortem" - "what went well" - "what went wrong" - "retrospective" - "goal review" - "debrief" permissions: - file:read - file:write

After-Action Review — Core Skill

A structured reflection process that runs after every completed goal. Extracts concrete lessons, evaluates output quality, identifies what worked and what didn't, and feeds everything into the self-improvement loop.

When It Runs

Automatically after any goal completes (all steps done)
On request when the user says "review goal" or "what went well"
Periodically as part of a weekly self-assessment (if autonomous mode is enabled)

The Review Process

Step 1: Gather Context

Collect all relevant data about the completed goal:

Goal title, type, description
Number of steps planned vs. actually executed
Time taken per step and total
AI providers used and their costs
Which skills were triggered
Any errors or retries that occurred
User feedback received during execution

Step 2: Quality Assessment

Rate the overall output on 5 dimensions:

After-Action Review: "Plan my time travel novel"
═══════════════════════════════════════════════════

Quality Assessment:
┌─────────────────────────────────┬───────┐
│ Completeness                    │ 9/10  │
│ Did we accomplish the goal?     │       │
├─────────────────────────────────┼───────┤
│ Quality                         │ 7/10  │
│ How good was the output?        │       │
├─────────────────────────────────┼───────┤
│ Efficiency                      │ 6/10  │
│ Did we use resources well?      │       │
├─────────────────────────────────┼───────┤
│ User Satisfaction               │ ?/10  │
│ (Awaiting user rating)          │       │
├─────────────────────────────────┼───────┤
│ Reusability                     │ 8/10  │
│ Can this approach work again?   │       │
└─────────────────────────────────┴───────┘

Overall Score: 7.5/10

Step 3: What Went Well

Identify and document successes:

✅ WHAT WENT WELL
─────────────────
1. Dynamic AI planning produced a coherent 7-step plan
   → The AI planner correctly identified this as a "planning" goal
   → Steps were logically ordered (premise → characters → world → outline)

2. Gemini handled planning steps efficiently at zero cost
   → All 4 planning steps used free-tier Gemini
   → Quality was sufficient for brainstorming/outlining

3. Character profiles were detailed and interconnected
   → AI naturally created relationships between characters
   → Motivations tied directly to the central conflict

4. User accepted the outline without major revisions
   → Strong signal that the structure was sound

Step 4: What Needs Improvement

Identify failures, inefficiencies, and areas for growth:

⚠️ WHAT NEEDS IMPROVEMENT
──────────────────────────
1. World-building step was too generic
   → Setting description lacked sensory specificity
   → Lesson: Add "include 3+ sensory details per location" to world-building prompts

2. Step 5 (review) was redundant with step 4 (outline)
   → Could have been combined into a single step
   → Lesson: For planning goals, combine review into the outline step

3. Total execution time: 8 minutes for 7 steps
   → Steps 2 and 3 could have run in parallel
   → Lesson: Character and world-building don't depend on each other — parallelize

4. Cost: $0.00 (all Gemini free tier)
   → Good for planning, but creative writing would need a better model
   → Lesson: Use Gemini for planning, switch to Claude/DeepSeek for prose

Step 5: Extract Lessons

Convert observations into structured lessons for the improvement log:

[
  {
    "category": "worldbuild",
    "lesson": "Always include 3+ sensory details (sight, sound, smell, touch, taste) per location description",
    "confidence": 0.75,
    "source": "after_action_review"
  },
  {
    "category": "task_execution",
    "lesson": "For planning goals, character profiles and world-building can run in parallel (no dependency)",
    "confidence": 0.8,
    "source": "after_action_review"
  },
  {
    "category": "task_execution",
    "lesson": "Combine 'review and refine' into the preceding step for planning goals to reduce redundancy",
    "confidence": 0.7,
    "source": "after_action_review"
  },
  {
    "category": "task_execution",
    "lesson": "Use Gemini free tier for planning/outlining tasks. Reserve Claude/DeepSeek for creative prose.",
    "confidence": 0.85,
    "source": "after_action_review"
  }
]

Step 6: User Feedback Request

Ask the user for their assessment:

📋 Goal Complete: "Plan my time travel novel"

I've completed my self-review. Quick questions:

1. Overall, how would you rate the output? (1-10)
2. What specifically did you like most?
3. What would you change for next time?

(Or just say "looks good" and I'll note that as positive feedback!)

Review Storage

Reviews are saved to workspace/memory/reviews/:

workspace/memory/reviews/
├── 2026-02-24-plan-time-travel-novel.md
├── 2026-02-24-research-medieval-weapons.md
└── 2026-02-25-write-chapter-1.md

Each review file contains the full structured assessment in Markdown format, readable by both humans and the AI.

Aggregate Reviews

Over time, reviews accumulate into patterns:

review performance this week

Shows:

Goals completed: 5
Average quality score: 7.8/10
Most common improvement area: "Prose specificity"
Lessons extracted: 12 (8 high confidence)
User satisfaction trend: Improving ↑

Integration with Self-Improvement

The After-Action Review feeds directly into the self-improvement loop:

Lessons extracted here are written to improvement-log.jsonl
Next time a similar goal runs, those lessons are injected into context
The review itself checks whether previous lessons were applied
Creates a measurable improvement trajectory over time

Commands

after action review — Run a review on the most recently completed goal
review goal [id] — Review a specific goal
review performance — Aggregate performance metrics
what went well — Quick summary of recent successes
what went wrong — Quick summary of recent failures
rate last goal [1-10] — Provide a user rating for the last goal