| name | thorough-review |
| description | Fan-out all review skills via parallel agents, then synthesize findings into one report |
Thorough Review
Run all review skills against a target by spawning parallel agents, then synthesize their findings into a single consolidated report.
Invocation
/thorough-review [target]
Target can be a file path (e.g., docs/architecturemud.md) or a concept name (e.g., "phase 1 design").
Workflow
1. Determine Target and Name
- If target is a file path: name = filename without path/extension (e.g.,
architecturemud)
- If target is a concept: slugify it (e.g.,
phase_1_design)
- If ambiguous, ask the user what to call this review
2. Read the Target
Read the target document fully. Extract a 2-3 sentence project context summary to include in each agent prompt.
3. Create Temp Directory
mkdir -p /tmp/reviews/{name}
4. Spawn Agent Batches
Spawn up to 3 agents in parallel per batch. Each agent gets:
Agent prompt template:
You are reviewing a software architecture document for [specific concern].
Project: Concept MRI is a research tool for studying attractor basin dynamics in MoE language models. It has a Python/FastAPI backend, React frontend, and uses Claude Code as the analysis runtime. The design under review is for adding a MUD-based interface layer.
Read the file at: [target path]
Then analyze it against these review questions:
[paste the Questions section from the relevant review skill]
Rules:
- Every finding must be concrete and specific to THIS design — reference specific sections
- If you can't point to a specific section or decision, the finding isn't real — drop it
- Generic advice that could apply to any project is not useful
Write your findings to: /tmp/reviews/{name}/{skill}.md
Use this output format:
[paste the Output Format section from the relevant review skill]
Batching order:
- Batch 1 (parallel): review-onboarding, review-interfaces, review-best-practices
- Batch 2 (parallel): review-risks, review-scope, review-evolution
- Batch 3 (parallel): review-deliverability, review-consistency, review-trace
- Batch 4 (parallel): devils-advocate (on overall approach), competitive-design (on key decisions)
- Skip review-drift if no implementation exists to compare against
5. Read All Findings
After all batches complete, read every file in /tmp/reviews/{name}/.
6. Synthesize
Produce a consolidated report:
# Thorough Review: {name}
Reviewed: {date}
Target: {target path or description}
Skills run: {list}
## Critical Findings
[Findings rated Critical from any skill, deduplicated]
## Important Findings
[Findings rated Important, grouped by theme rather than by skill]
## Minor Findings
[Brief list]
## Cross-Cutting Patterns
[Themes that appeared across multiple skills — these are the real insights]
## Strengths
[What the design does well — important for context, not just a list of problems]
## Recommended Actions
[Prioritized list: what to fix first, what can wait, what to accept]
7. Write Report and Clean Up
Write to: docs/scratchpad/review_{name}.md
rm -rf /tmp/reviews/{name}
8. Present Summary
Give the user a brief summary of top findings and point them to the full report.
Notes
- This takes several minutes (4 agent batches, each reading and analyzing 800+ lines)
- First run on a new target is the calibration — expect to adjust agent prompts based on output quality
- Re-running overwrites the previous report. Git tracks history.
- Individual review skills can also be run standalone for a focused review (output is inline, no file)