Reviewer Loop
Work product passes through a distinct review pass (separate skill, separate agent, or separate role prompt) that hunts for defects against a rubric before the work is accepted. Findings route back to the producer for fixes, bounded by iteration count.
When to use
Code review, content editorial, plan review: anywhere producer bias hides defects.
Steps
| Step | What happens | Verification |
|---|---|---|
| Produce | Producer completes work to its own definition of done. | Producer's own checks pass first; review is not a substitute for self-verification. |
| Review | Reviewer evaluates against a written rubric, reporting findings with severity and evidence. | Each finding cites the specific artifact location and failure scenario. |
| Resolve | Producer fixes confirmed findings or documents disagreement; loop repeats up to the iteration bound. | Re-review of changed areas only; iteration count enforced. |
Specification
| Verification method | The rubric. Reviewer quality is measurable via seeded-defect recall. |
|---|---|
| Recovery behavior | bounded-iteration-then-escalate |
| Observability | Findings log with severities and resolutions. |
| Inputs | Work product, Review rubric, Iteration bound |
| Outputs | Accepted work, Findings log |
| Risk level / maturity | low / established |
Failure modes
- Reviewer rubber-stamps (same context as producer)
- Unbounded ping-pong
- Rubric drift: findings not tied to written criteria
Assessment
Context separation is what makes this work: a reviewer sharing the producer's full reasoning inherits its blind spots. Fresh-context review finds what self-review cannot.