agent-acceptance-review

name	agent-acceptance-review
description	Use when an agent has produced a task version, implementation, document, plan, research result, or other deliverable that must be independently accepted, rejected, scored, or sent back for iteration.

Agent Acceptance Review

Overview

Use this skill as an acceptance gate for agent work. Review the submitted version with neutral evidence, score it on a 0-100 scale, and accept only when it reaches 90 or higher with no blocker issues.

This is not a replacement for tests, linters, builds, domain checks, or user approval. Treat deterministic verification as the first layer and judge scoring as the second layer.

Core Rules

• Preserve reviewer independence. Do not reveal the worker agent's self-assessment, preferred conclusion, or expected score.

• Score evidence, not effort. A long attempt, many tool calls, or confident summary does not improve the score unless the result satisfies the task.

• Require concrete verification before acceptance. If expected tests or checks were not run, cap the score at 89 unless the task is clearly non-testable.

• Accept only if all are true: final score >= 90, no blocker, critical requirements satisfied, and verification evidence is adequate.

• If score < 90, return actionable iteration feedback instead of declaring completion.

• For high-risk work, prefer full mode and include multi-agent-decision-review.

• Never invent evaluation results from a platform, test suite, or tool that was not actually run.

Review Modes

Light Mode

Use light mode for small, reversible, low-risk tasks.

• Use one independent judge pass.

• Inspect the task request, relevant artifacts, changes, and verification evidence.

• Output score, blocker status, reasons, and next actions.

Full Mode

Use full mode when the work is broad, risky, user-facing, expensive to undo, or likely to need multiple perspectives.

• Use multi-agent-decision-review when available.

• Ask reviewers to score independently before synthesis.

• Use 3-5 reviewer perspectives selected for failure modes, such as correctness, user value, maintainability, risk, and first-principles skepticism.

• Surface disagreements instead of averaging them away.

Acceptance Workflow

1. Frame the submitted version:

   • Original request and acceptance criteria.

   • Claimed deliverable.

   • Changed files, created artifacts, or produced content.

   • Verification already performed.

   • Known constraints, skipped checks, or unresolved questions.

2. Run deterministic checks first:

   • Tests, build, lint, typecheck, format, smoke test, visual check, schema validation, citation check, or other task-specific checks.

   • If a check cannot be run, record why and whether this should cap the score.

3. Choose review mode:

   • Light mode for narrow, low-risk work.

   • Full mode for architecture, product behavior, user-facing UI, data, legal/compliance, security, research claims, or multi-file changes.

4. Score with the rubric:

   • Requirement satisfaction: 35

   • Correctness and verification: 25

   • Maintainability and integration: 15

   • Risk control: 15

   • Efficiency and focus: 10

5. Apply gates:

   • Any blocker means reject, regardless of numeric score.

   • Missing required verification usually caps score at 89.

   • Scope drift, fabricated evidence, or unreviewed destructive changes cap score at 79 or lower.

6. Decide:

   • Accepted: score >= 90 and no blockers.

   • Needs iteration: score 70-89 or non-blocking but material gaps remain.

   • Rejected: score < 70, blocker present, or result does not satisfy the core task.

7. If not accepted, give the worker the smallest useful next iteration:

   • Confirmed problems.

   • Evidence for each problem.

   • Required fixes.

   • Checks to rerun.

   • What would raise the score above 90.

Output Shape

Use this shape for most reviews:

**Acceptance**
Decision: Accepted | Needs iteration | Rejected
Score: NN/100
Blockers: none | ...
Confidence: low | medium | high

**Score**
| Dimension | Points | Reason |
| --- | ---: | --- |
| Requirement satisfaction | NN/35 | ... |
| Correctness and verification | NN/25 | ... |
| Maintainability and integration | NN/15 | ... |
| Risk control | NN/15 | ... |
| Efficiency and focus | NN/10 | ... |

**Evidence**
- ...

**Required Next Iteration**
- ...

Omit Required Next Iteration only when accepted.

Reference Files

• Read references/acceptance-prompts.md before writing prompts for judge agents, subagents, or simulated reviewer passes.