Prompt Reviewer

Prompt Reviewer

Purpose

Review prompts against quality, safety, and reliability criteria; produce findings with severity and fix recommendations.

Input: Prompt specification or raw prompt text, use case context, risk level Output: Prompt Review Report with scored dimensions, findings, and revised sections Examples: See [examples.md](examples.md) for worked input/output.

---

Workflow

Step 1: Establish review context

Confirm:

• Prompt purpose and use case
• Risk level (internal / customer-facing / regulated)
• Model target (if known)
• Expected output format

Step 2: Score review dimensions

| Dimension | Score (1-5) | Notes | |-----------|-------------|-------| | Clarity | | Instructions unambiguous | | Completeness | | All needed context and rules | | Output contract | | Format enforceable | | Safety | | Injection resistance, refusal behavior | | Efficiency | | Token usage reasonable | | Testability | | Can verify output programmatically | | Maintainability | | Variables, versioning, modularity |

Step 3: Run safety checklist

• [ ] Input delimiter boundaries defined
• [ ] Instruction/data separation clear
• [ ] Out-of-scope refusal behavior
• [ ] No secrets or credentials in template
• [ ] PII handling rules if applicable
• [ ] Jailbreak resistance for customer-facing

Step 4: Document findings

| ID | Severity | Category | Finding | Fix | |----|----------|----------|---------|-----|

Severity: Critical | High | Medium | Low

Step 5: Provide revised sections

For Critical/High findings, rewrite affected prompt sections.

Step 6: Validate

Run Validation checklist.

---

Decision Rules

| Condition | Action | |-----------|--------| | Customer-facing with no refusal rules | Critical finding; block deploy | | Instructions contradict examples | High finding; fix examples or instructions | | Ambiguous output format | High finding; specify exact format | | Role Play detected | Medium finding; rewrite to task-based instructions | | No test cases provided | Recommend prompt-engineer test cases |

---

Validation

• [ ] All 7 dimensions scored with notes
• [ ] Safety checklist completed
• [ ] ≥3 findings documented (or explicit "no issues" with evidence)
• [ ] Critical/High findings have fix recommendations
• [ ] Revised sections provided for Critical/High
• [ ] Overall deploy recommendation: Approve / Approve with fixes / Block
• [ ] No Role Play introduced in revisions

---

Anti-patterns

• Rubber stamp approval — all 5s without analysis.
• Style-only feedback — word choice vs structural issues.
• Missing safety on external prompts — ignoring injection risks.
• Vague fixes — "make it clearer" without rewrite.
• Ignoring examples — examples that teach wrong behavior.

---

Best Practices

• Test mentally with adversarial inputs.
• Check instruction-example consistency.
• Verify output format is parseable.
• Align safety level with use case risk.
• Reference OWASP LLM Top 1 (prompt injection) for external prompts.

---

Output Structure

# Prompt Review: [Prompt Name]
**Date:** YYYY-MM-DD
**Risk level:** [level]
**Recommendation:** Approve | Approve with fixes | Block

## Dimension Scores
| Dimension | Score | Notes |
|-----------|-------|-------|

## Safety Checklist
- [x] or [ ] [item]

## Findings
| ID | Severity | Finding | Fix |
|----|----------|---------|-----|

## Revised Sections
### [Section name]
[Rewritten content]

## Required Actions Before Deploy
1. [Action]

---

Next Skills

| Outcome | Recommended Skill | |---------|-------------------| | Optimize approved prompt | ai/prompt-optimizer | | Major rewrite needed | ai/prompt-engineer | | Context budget issues | ai/context-engineering | | AI security review | security/prompt-injection-detector |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: patonkikh
• Source: patonkikh/APES
• License: MIT

Install and usage instructions live in the source repository linked above.