Review a technical plan before implementation.
Bring an architecture proposal, requirements, and rollout plan. Ask AI reviewers to look for failure cases, missing trade-offs, and assumptions that need testing.
Two free Refinement reviews each month · No credit card required
A design review needs more than the happy path.
Include dependencies, expected load, failure tolerance, and migration constraints. Without them, a reviewer may criticize choices that are necessary or miss risks that matter.
Ask findings to name a failure mechanism, likely impact, and a way to test the concern. The engineering team decides which findings become changes or release requirements.
What to look for in the review.
Test assumptions
Question capacity, dependencies, and the conditions the design relies on.
Review the rollout
Ask what happens during migration, partial failure, and rollback.
Prepare the human review
Bring a focused list of questions and proposed checks to the team.
From design review comments to an adversarial RFC audit
A strong review covers the system lifecycle, not only the proposed steady state. Use the table as a minimum issue map for the panel and the human reviewers who follow it.
| RFC surface | Weak review question | Adversarial review question |
|---|---|---|
| Capacity | Will this scale? | Which resource saturates first and how will we observe it? |
| Failure | Is it highly available? | What happens during partial dependency failure and retry storms? |
| Data | Is migration covered? | How are dual writes, backfill, reconciliation, and rollback proven? |
| Security | Is authentication included? | Where do trust boundaries, confused deputies, and privilege escalation appear? |
| Operations | Do we have monitoring? | Which signal detects each failure before customers report it? |
| Delivery | Can we roll back? | Which state changes make rollback unsafe or impossible? |
| Ownership | Who maintains it? | Who decides, mitigates, and responds when the accepted risk occurs? |
Prepare a useful review.
Give reviewers a clear question and enough context to examine it.
Attack the capacity model
Require concrete traffic assumptions, resource multipliers, queue behavior, and degradation modes.
Walk every state transition
Inspect create, update, delete, replay, retry, partial success, cancellation, and recovery.
An AI reviewer cannot observe your production system from the RFC alone.
The panel only knows the architecture, constraints, metrics, and incidents you provide. Missing organizational context can make a plausible critique irrelevant or conceal the real risk. Include traffic shape, service dependencies, data classification, deployment model, operational maturity, and explicit non-goals. Then have system owners validate the findings.
Do not paste secrets, credentials, exploitable internal details, or restricted customer data into a review without authorization. Use sanitized diagrams and least-necessary context. Security, privacy, and reliability owners remain responsible for approval and for deciding which findings require testing before launch.
What to know before you run the panel
Which debate format works best for an RFC review?
Red Team is the natural default for failure modes and security. Socratic works well when the design rests on unclear assumptions. Devil’s Advocate is useful when a preferred architecture already has momentum. Steelman helps when the rejected alternative has not received a fair comparison.
What should I include with the RFC?
Include goals, non-goals, constraints, diagrams, data flows, load assumptions, dependencies, security boundaries, migration steps, rollback, observability, alternatives, and known risks. State which decision the review must make and what evidence would block approval.
Does agreement mean the answer is correct?
No. Models can repeat the same mistake or rely on the same weak source. Read the objections and check important claims against evidence.
Why use other models instead of asking one to critique itself?
Self-critique can help. Other models and specific review roles provide additional responses to compare. The format determines whether reviewers begin independently or see earlier answers.
Does this replace human review?
No. Use the findings to improve a draft or prepare questions for a responsible reviewer. The models do not approve a decision or verify every claim.
Bring your next draft.
Create a free account to draft or import your work. Choose the reviewers and settings before starting a review.