system-review

name	system-review
description	Performs a meta-level review of how well an implementation followed its plan, classifying divergences and recommending AI-Layer improvements. Use after an execution report exists to find bugs in the process, not the code.

System Review

Perform a meta-level analysis of how well the implementation followed the plan and identify process improvements.

Purpose

System review is NOT code review. You're not looking for bugs in the code - you're looking for bugs in the process.

Your job:

• Analyze plan adherence and divergence patterns
• Identify which divergences were justified vs problematic
• Surface process improvements that prevent future issues
• Suggest updates to AI-Layer assets (CLAUDE.md, plan templates, skills)

Philosophy:

• Good divergence reveals plan limitations → improve planning
• Bad divergence reveals unclear requirements → improve communication
• Repeated issues reveal missing automation → create skills

Context & Inputs

You will analyze four key artifacts:

Plan Skill: Read this to understand the planning process and what instructions guide plan creation. .claude/skills/plan-feature/SKILL.md

Generated Plan: Read this to understand what the agent was SUPPOSED to do. Plan file: $1

Execute Skill: Read this to understand the execution process and what instructions guide implementation. .claude/skills/execute/SKILL.md

Execution Report: Read this to understand what the agent ACTUALLY did and why. Execution report: $2

Analysis Workflow

Step 1: Understand the Planned Approach

Read the generated plan ($1) and extract:

• What features were planned?
• What architecture was specified?
• What validation steps were defined?
• What patterns were referenced?

Step 2: Understand the Actual Implementation

Read the execution report ($2) and extract:

• What was implemented?
• What diverged from the plan?
• What challenges were encountered?
• What was skipped and why?

Step 3: Classify Each Divergence

For each divergence identified in the execution report, classify as:

Good Divergence ✅ (Justified):

• Plan assumed something that didn't exist in the codebase
• Better pattern discovered during implementation
• Performance optimization needed
• Security issue discovered that required different approach

Bad Divergence ❌ (Problematic):

• Ignored explicit constraints in plan
• Created new architecture instead of following existing patterns
• Took shortcuts that introduce tech debt
• Misunderstood requirements

Step 4: Trace Root Causes

For each problematic divergence, identify the root cause:

• Was the plan unclear, where, why?
• Was context missing, where, why?
• Was validation missing, where, why?
• Was manual step repeated, where, why?

Step 5: Generate Process Improvements

Based on patterns across divergences, suggest:

• CLAUDE.md updates: Universal patterns or anti-patterns to document
• Plan skill updates: Instructions that need clarification or missing steps
• New skills: Manual processes that should be automated
• Validation additions: Checks that would catch issues earlier

Output Format

Save your analysis to: .claude/system-reviews/[feature-name]-review.md

Report Structure:

Meta Information

• Plan reviewed: [path to $1]
• Execution report: [path to $2]
• Date: [current date]

Overall Alignment Score: __/10

Scoring guide:

• 10: Perfect adherence, all divergences justified
• 7-9: Minor justified divergences
• 4-6: Mix of justified and problematic divergences
• 1-3: Major problematic divergences

Divergence Analysis

For each divergence from the execution report:

divergence: [what changed]
planned: [what plan specified]
actual: [what was implemented]
reason: [agent's stated reason from report]
classification: good ✅ | bad ❌
justified: yes/no
root_cause: [unclear plan | missing context | etc]

Pattern Compliance

Assess adherence to documented patterns:

• Followed codebase architecture
• Used documented patterns (from CLAUDE.md)
• Applied testing patterns correctly
• Met validation requirements

System Improvement Actions

Based on analysis, recommend specific actions:

Update CLAUDE.md:

• Document [pattern X] discovered during implementation
• Add anti-pattern warning for [Y]
• Clarify [technology constraint Z]

Update Plan Skill ($1):

• Add instruction for [missing step]
• Clarify [ambiguous instruction]
• Add validation requirement for [X]

Create New Skill:

• A new skill for [manual process repeated 3+ times]

Update Execute Skill:

• Add [validation step] to execution checklist

Key Learnings

What worked well:

• [specific things that went smoothly]

What needs improvement:

• [specific process gaps identified]

For next implementation:

• [concrete improvements to try]

Important

• Be specific: Don't say "plan was unclear" - say "plan didn't specify which auth pattern to use"
• Focus on patterns: One-off issues aren't actionable. Look for repeated problems.
• Action-oriented: Every finding should have a concrete asset update suggestion
• Suggest improvements: Don't just analyze - actually suggest the text to add to CLAUDE.md or skills