quality-review
Skill6 starsChanged 29 days agoMITView on GitHub
Structured 3-agent quality review using parallel simplifier, adversarial-reviewer, and chaos-engineer agents. Use when evaluating plans, designs, PRDs, architecture decisions, or any proposal before implementation. Three non-overlapping lenses — KISS/YAGNI, evidence/reasoning, robustness/edge cases.
9 other projects use a version of this.
Adopted by
Most recently taken up by
- PicsArt/media-studiotheir own version, first seen Nov 16, 2026
- Gamehu/reinstheir own version, first seen Nov 8, 2026
- TerryFYL/ai-research-armytheir own version, first seen Nov 8, 2026
- calven-ai/calven-mcptheir own version, first seen Nov 1, 2026
- PicsArt/gen-ai-skillstheir own version, first seen Nov 1, 2026
- prinx/agentstheir own version, first seen Oct 20, 2026
Other versions9 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 9
- TerryFYL/ai-research-army★ 9same name
- PicsArt/gen-ai-skills★ 7same name
- calven-ai/calven-mcp★ 3same name
- gener8v/gener8v.claude-skills★ 3same name
- simota/quality-skills★ 2same name
- PicsArt/media-studio★ 1same name
- Gamehu/reins★ 0same name
- kaiohenricunha/dotbabel★ 0same name
- prinx/agents★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository30 other files it gives its agents.
- adversarial-re-readSkill
- adversarial-reviewSkill
- build-estimateSkill
- build-evaluateSkill
- build-executeSkill
- build-planSkill
- build-polishSkill
- build-prdSkill
- build-probeSkill
- build-validateSkill
- excel-connector-governanceSkill
- five-whysSkill
- insight-builderSkill
- insight-crystallizerSkill
- insight-promotionSkill
- multi-sessionSkill
- notion-governanceSkill
- oodaSkill
- prd-writerSkill
- promote-to-toolingSkill
- ralph-implementSkill
- ralph-loopSkill
- search-governanceSkill
- session-outputSkill
- skill-validatorSkill
- task-planSkill
- verifySkill
- voice-profile-testSkill
- why-diagnosticSkill
- writing-qualitySkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
16 rules. 0 are common, 15 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 7 edited versions we read keep these.
15
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 24
Before implementing a new phase, feature, or architecture change
- Distinctive: line 25
When making irreversible decisions (schema design, MCP server architecture, skill interfaces, API contracts)
- Distinctive: line 26
After a research team or planning agent produces recommendations — before accepting them
- Distinctive: line 27
When the user asks "is this right?" or "should I do this?"
- Distinctive: line 28
Before publishing public-facing copy or signing off on a deliverable
- Distinctive: line 29
Before merging a structural PR
- Distinctive: line 69
When the user provides new context that changes the framing (e.g., "this is a template, not a one-off" or "this is throwaway code"), …
- Distinctive: line 88
When spawning each review agent, always provide:
- Distinctive: line 95
Keep the prompt under 300 words.
- Distinctive: line 105
All accepted recommendations should be integrated into the source artifact (PRD, design doc, ROADMAP) immediately, not left as floating …
- Distinctive: line 109
Don't run quality-review on trivial changes
- Distinctive: line 110
Don't run more than 2 rounds without genuinely new context
- Distinctive: line 111
Don't treat agent findings as authoritative — they inform user judgment, they don't override it
- Distinctive: line 112
Don't let review become a substitute for building — the goal is better decisions, not more documents
- Distinctive: line 113
Don't skip the chaos-engineer when the proposal "feels safe." Safety bias is exactly when chaos catches the most.
- Very specific: line 117
After Round 1 gathers findings from the three agents in parallel, invoke knowledge-synthesizer to combine the outputs rather than …
“Common” is measured: kept by at least 6 in 10 of the 7 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Quality Review
A structured interrogation methodology using three review agents in parallel to stress-test proposals before committing to them. Each agent attacks a different axis with no overlap.
The Three Lenses
| Agent | Model | Lens | Asks |
simplifier | Sonnet | KISS / YAGNI — complexity, abstractions, duplication | Is this the simplest way? |
adversarial-reviewer | Opus | Evidence, assumptions, logical gaps | Is this the right thing? |
chaos-engineer | Sonnet | Robustness, edge cases, failure modes | What breaks this? |
Why three: the simplifier prevents over-engineering, the adversarial reviewer prevents under-thinking, the chaos engineer prevents under-stress-testing. Together they catch design failures across complexity, correctness, and robustness.
Why Opus for adversarial: adversarial reasoning needs to find non-obvious logical gaps and challenge framing itself. The other two work in well-defined search spaces (complexity patterns, failure patterns) and run fine on Sonnet.
When to Use
- Before implementing a new phase, feature, or architecture change
- When making irreversible decisions (schema design, MCP server architecture, skill interfaces, API contracts)
- After a research team or planning agent produces recommendations — before accepting them
- When the user asks "is this right?" or "should I do this?"
- Before publishing public-facing copy or signing off on a deliverable
- Before merging a structural PR
When NOT to Use
- Trivial changes (single-file edits, config tweaks, typo fixes)
- Pure information requests (quality-review is for proposals, not questions)
- Code review of already-implemented code where the design is locked — use a code-review pass instead
The Pattern
Round 1 — Initial Review
Spawn all three agents in parallel against the proposal:
Agent(simplifier): "Review this proposal. Find over-engineering and simpler alternatives."
Agent(adversarial-reviewer): "Review this proposal. Challenge assumptions and find weaknesses."
Agent(chaos-engineer): "Review this proposal. Find edge cases, failure modes, and robustness gaps."Feed each agent:
- The proposal / plan / PRD being reviewed (full text or file path)
- Relevant context (existing PRDs, CLAUDE.md, current project state)
- The specific decision being made, if there is one
Collect: numbered findings grouped by priority (Critical / High / Medium / Low) from each agent.
Round 2 — Synthesis
Synthesize all three agents' findings into a unified report:
- Where do they agree? (high-confidence findings — implement these)
- Where do they disagree? (needs human judgment — surface the tension, don't resolve it)
- What's the combined recommendation? (specific structural changes to PRD, design, or implementation plan)
Present to the user for decision.
Round 3 — Re-Review (if context changes)
When the user provides new context that changes the framing (e.g., "this is a template, not a one-off" or "this is throwaway code"), re-engage the relevant agents with the updated context:
Agent(<lens>): "You previously found X. The user clarifies Y. Re-evaluate."This round often produces the strongest insights — agents refine their positions with better information. Don't run more than 2 re-review rounds unless genuinely new information arrives.
Round 4 — Convergence
Final synthesis. Typically you have:
- Findings all three agents agree on → implement
- Findings resolved by new context → document the resolution
- Remaining disagreements → user decides
- Specific structural recommendations → integrate into PRD / design doc / ROADMAP immediately, don't leave as floating review notes
Prompt Structure
When spawning each review agent, always provide:
- Role reminder — "You are the Simplifier / Adversarial Reviewer / Chaos Engineer."
- The artifact — full text of what's being reviewed, or a file path with explicit instruction to read it
- Context — existing PRD, project state, user constraints (≤200 words, only what's relevant)
- Specific question — what the user is actually deciding (not just "review this")
Keep the prompt under 300 words. Over-specified prompts constrain the agent's judgment and waste context.
Outputs
Quality review produces:
- Findings table — numbered, prioritized, with all three agents' positions
- Structural recommendations — specific changes to the PRD or design
- Decision items — things only the user can resolve
All accepted recommendations should be integrated into the source artifact (PRD, design doc, ROADMAP) immediately, not left as floating review notes.
Anti-Patterns
- Don't run quality-review on trivial changes
- Don't run more than 2 rounds without genuinely new context
- Don't treat agent findings as authoritative — they inform user judgment, they don't override it
- Don't let review become a substitute for building — the goal is better decisions, not more documents
- Don't skip the chaos-engineer when the proposal "feels safe." Safety bias is exactly when chaos catches the most.
Synthesis via knowledge-synthesizer
After Round 1 gathers findings from the three agents in parallel, invoke knowledge-synthesizer to combine the outputs rather than hand-rolling the merge. The synthesizer:
- Identifies findings 2+ agents independently surfaced (high confidence)
- Surfaces unique findings from a single agent (still data, not consensus)
- Preserves disagreements explicitly (does not force resolution)
- Flags coverage gaps
For Round 2 re-engagement after context shifts, run the affected agents in parallel via multi-agent-coordinator (so a failure in one does not stop the others), then re-synthesize via knowledge-synthesizer.
Parallel spawn discipline
The three review agents run truly in parallel — spawn them via multi-agent-coordinator, do not serialize. The coordinator tracks state, handles partial failures, and returns the combined results to the synthesizer.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

