coding workflow

PR Review Pack

Return a severity-ranked review in which every finding is actionable and evidence-backed.

Setup: 20 minutes
Difficulty: Intermediate
Tested: 2026-07-26

# PR Review Pack

Review a change against its intent, reproduce important findings, and prepare focused fixes.

**Outcome:** Return a severity-ranked review in which every finding is actionable and evidence-backed.

**Category:** coding

**Setup:** 20 minutes · Intermediate

## Requirements

- The change diff and base revision
- The approved plan or acceptance criteria
- Relevant test commands

**Tools:** claude-code, openai-codex

## Expected output

A review with verified findings, file references, and explicit residual risk

## Steps

### 1. Lock the review boundary

List the requested behavior, changed files, and verification commands before judging the code.

**Prompt**

```text
Summarize the change contract and identify any requirement that cannot be verified from the provided context.
```

### 2. Inspect for behavioral risk

Trace changed inputs, state transitions, errors, and downstream consumers.

**Prompt**

```text
Find correctness, security, data-loss, and regression risks. Ignore style unless it obscures a defect.
```

### 3. Reproduce each finding

Run or write the smallest check that proves each suspected issue before reporting it.

**Prompt**

```text
For every candidate finding, provide evidence, impact, exact location, and a reproducible verification step.
```

### 4. Deliver the review

Rank only confirmed issues, then state missing tests and residual risk separately.

**Prompt**

```text
Output findings first by severity, followed by open questions and a two-sentence summary.
```

## Limitations

An agent review cannot replace domain-owner judgment for business rules or high-risk releases.

## Privacy

Use repository-scoped permissions and do not expose secrets, private dependencies, or production data to unapproved services.

## Sources

- https://www.anthropic.com/product/claude-code
- https://openai.com/index/unrolling-the-codex-agent-loop/

Tested 2026-07-26 · Updated 2026-07-26

# PR Review Pack

Review a change against its intent, reproduce important findings, and prepare focused fixes.

**Outcome:** Return a severity-ranked review in which every finding is actionable and evidence-backed.

**Category:** coding

**Setup:** 20 minutes · Intermediate

## Requirements

- The change diff and base revision
- The approved plan or acceptance criteria
- Relevant test commands

**Tools:** claude-code, openai-codex

## Expected output

A review with verified findings, file references, and explicit residual risk

## Steps

### 1. Lock the review boundary

List the requested behavior, changed files, and verification commands before judging the code.

**Prompt**

```text
Summarize the change contract and identify any requirement that cannot be verified from the provided context.
```

### 2. Inspect for behavioral risk

Trace changed inputs, state transitions, errors, and downstream consumers.

**Prompt**

```text
Find correctness, security, data-loss, and regression risks. Ignore style unless it obscures a defect.
```

### 3. Reproduce each finding

Run or write the smallest check that proves each suspected issue before reporting it.

**Prompt**

```text
For every candidate finding, provide evidence, impact, exact location, and a reproducible verification step.
```

### 4. Deliver the review

Rank only confirmed issues, then state missing tests and residual risk separately.

**Prompt**

```text
Output findings first by severity, followed by open questions and a two-sentence summary.
```

## Limitations

An agent review cannot replace domain-owner judgment for business rules or high-risk releases.

## Privacy

Use repository-scoped permissions and do not expose secrets, private dependencies, or production data to unapproved services.

## Sources

- https://www.anthropic.com/product/claude-code
- https://openai.com/index/unrolling-the-codex-agent-loop/

Tested 2026-07-26 · Updated 2026-07-26

{
  "slug": "pr-review-pack",
  "title": "PR Review Pack",
  "description": "Review a change against its intent, reproduce important findings, and prepare focused fixes.",
  "outcome": "Return a severity-ranked review in which every finding is actionable and evidence-backed.",
  "category": "coding",
  "tags": [
    "pull-request",
    "review",
    "testing"
  ],
  "setupMinutes": 20,
  "difficulty": "Intermediate",
  "featured": true,
  "testedAt": "2026-07-26",
  "updatedAt": "2026-07-26",
  "tools": [
    "claude-code",
    "openai-codex"
  ],
  "inputs": [
    "The change diff and base revision",
    "The approved plan or acceptance criteria",
    "Relevant test commands"
  ],
  "expectedOutput": "A review with verified findings, file references, and explicit residual risk",
  "steps": [
    {
      "title": "Lock the review boundary",
      "instruction": "List the requested behavior, changed files, and verification commands before judging the code.",
      "prompt": "Summarize the change contract and identify any requirement that cannot be verified from the provided context."
    },
    {
      "title": "Inspect for behavioral risk",
      "instruction": "Trace changed inputs, state transitions, errors, and downstream consumers.",
      "prompt": "Find correctness, security, data-loss, and regression risks. Ignore style unless it obscures a defect."
    },
    {
      "title": "Reproduce each finding",
      "instruction": "Run or write the smallest check that proves each suspected issue before reporting it.",
      "prompt": "For every candidate finding, provide evidence, impact, exact location, and a reproducible verification step."
    },
    {
      "title": "Deliver the review",
      "instruction": "Rank only confirmed issues, then state missing tests and residual risk separately.",
      "prompt": "Output findings first by severity, followed by open questions and a two-sentence summary."
    }
  ],
  "limitations": "An agent review cannot replace domain-owner judgment for business rules or high-risk releases.",
  "privacy": "Use repository-scoped permissions and do not expose secrets, private dependencies, or production data to unapproved services.",
  "sources": [
    "https://www.anthropic.com/product/claude-code",
    "https://openai.com/index/unrolling-the-codex-agent-loop/"
  ],
  "related": [
    "bug-report-cleaner"
  ]
}

Requirements

Bring these inputs.

The change diff and base revision
The approved plan or acceptance criteria
Relevant test commands

Output contract

Know what done looks like.

A review with verified findings, file references, and explicit residual risk

Execution

Run the playbook.

01
Lock the review boundary
List the requested behavior, changed files, and verification commands before judging the code.
```
Summarize the change contract and identify any requirement that cannot be verified from the provided context.
```
02
Inspect for behavioral risk
Trace changed inputs, state transitions, errors, and downstream consumers.
```
Find correctness, security, data-loss, and regression risks. Ignore style unless it obscures a defect.
```
03
Reproduce each finding
Run or write the smallest check that proves each suspected issue before reporting it.
```
For every candidate finding, provide evidence, impact, exact location, and a reproducible verification step.
```
04
Deliver the review
Rank only confirmed issues, then state missing tests and residual risk separately.
```
Output findings first by severity, followed by open questions and a two-sentence summary.
```

Field note

How to use this route.

This workflow treats review as verification, not opinion generation. A finding is not ready until the reviewer can show how it fails or why the behavior violates the contract.

Bring these inputs.

Know what done looks like.

Run the playbook.

Lock the review boundary

Inspect for behavioral risk

Reproduce each finding

Deliver the review

How to use this route.

Keep moving through the atlas