Red/Blue/Green Review

Simple Definition

Red/Blue/Green Review is a three-stage AI review method where each stage has a distinct job: Red identifies risks and failure modes in a piece of work, Blue ranks which of those risks actually matter, and Green revises the work to fix the important ones while preserving what was already working.

How the Three Stages Work

Red looks for what could be wrong: missing context, weak assumptions, security or privacy risks, confusing wording, loopholes, potential misuse, unintended effects, and edge cases. Red’s job is to find problems, not fix or judge them yet.

Blue takes Red’s findings and triages them by likelihood, severity, reversibility, strength of evidence, urgency, and cost to fix. Blue’s output is a short, prioritized list, not a restatement of everything Red found.

Green revises the work to address what Blue ranked as important, preserving the original goal, adding safeguards where warranted, and documenting any remaining risk that wasn’t worth fully resolving.

Why One-Stage Self-Review Is Weak

A model reviewing work it just produced tends to inherit the same context, assumptions, and mild bias toward agreement that shaped the original answer. A single self-review pass often reads more like polish than genuine scrutiny. Splitting the process into three distinct jobs, find, rank, fix, forces each one to actually happen instead of collapsing into a quick “looks fine to me.”

Example

Before sending an important client proposal, a team runs a Red pass that flags six potential issues: an unclear pricing term, a missing timeline caveat, and four smaller wording concerns. A Blue pass ranks the pricing term and timeline caveat as the two that genuinely matter, since both could cause a real dispute later, and flags the rest as minor. A Green pass fixes those two issues specifically, leaving the rest of the proposal, which was already solid, unchanged.

Best Use Cases

This workflow earns its extra process on work with real consequences if it’s wrong: client-facing communication, policy documents, code changes touching production, marketing claims, hiring materials, financial models, and anything published or shipped publicly. It’s overkill for casual drafts, reversible low-stakes tasks, and brainstorming.

Limitations

This workflow doesn’t guarantee correctness. Blue’s prioritization is still a judgment call that can rank a real risk as unimportant, and Green’s fix is only as good as Blue’s ranking. Running all three stages in the same conversation as the original work weakens the independence that makes the method useful, a fresh chat or separate agent call for each stage produces a more genuinely independent review. And, like any AI review process, it supports human judgment rather than replacing it, especially for anything with serious consequences if it’s wrong.

AI Red Teaming, the adversarial-testing tradition this method’s “Red” stage draws from
Verification Gate, the evidence-based checkpoint that should still follow Green’s revision
Blind-Spot Audit, a lighter, single-step way to surface AI uncertainty
Decision Audit, reviewing the assumptions behind finished work after the fact
Human-in-the-Loop, the person who should still review the final result

For the full workflow, a comparison table, and a copy-paste prompt, see The Red, Blue, Green AI Review Loop: Find Risks, Rank Them, Then Fix Them.