# Manage Coding Agents With Clear Scope and Acceptance Criteria

Coding agents drift, expand scope, and skip verification — this skill provides a lightweight PM operating loop with templates for briefing, reviewing, and shipping their work. It gives founders and product engineers a repeatable system to delegate, review, and verify agent output without losing control.

## Install

```bash
npx skillstore add bridge ai labs/kiwi-phantomworks-coding-agent-pm
```

## Metadata

- - Slug: kiwi-phantomworks-coding-agent-pm
- - Version: 1.0.1
- - Author: Bridge AI Labs
- - GitHub username: kiwi-phantomworks
- - License: MIT
- - Repository: https://github.com/kiwi-phantomworks/claude-code-pm-starter-pack/tree/main/
- - Ref: main
- - Supported tools: Claude, Codex, Claude Code
- - Risk level: safe
- - Risk factors: external\_commands, network, filesystem
- - Quality score: 81
- - Quality tier: silver
- - Public page: https://skillstore.pages.dev/skills/kiwi-phantomworks-coding-agent-pm
- - Manifest: https://skillstore.pages.dev/api/skills/kiwi-phantomworks-coding-agent-pm/manifest

## Capabilities

- Define a six-step operating loop for assigning and reviewing coding-agent work
- Generate a structured PM agent brief with outcome, scope, context, and acceptance criteria
- Create task plans with ownership boundaries, forbidden actions, and verification steps
- Provide reusable prompt snippets for planning, implementation, review, drift recovery, and delegation
- Supply a review checklist covering scope, behavior, code quality, verification, and security/privacy
- Track first-week agent workflow results with a scoring rubric for time saved and rework

## Use Cases

- Delegate a feature task to a coding agent: A founder uses the PM agent brief template to define outcome, scope, and acceptance criteria before assigning a feature to Claude Code or Codex.
- Review agent-submitted pull requests systematically: A product engineer applies the review checklist to evaluate an agent's diff for scope drift, missing tests, security issues, and verification quality.
- Recover when a coding agent drifts off-scope: A PM uses the drift-recovery prompt snippet to stop the agent, map changes back to acceptance criteria, and re-establish boundaries.

## Prompt Templates

### Create a PM agent brief

```
Use the pm-agent-brief template. Fill in: (1) the one-sentence shipped outcome, (2) in-scope and out-of-scope files and behaviors, (3) project context and conventions, (4) acceptance criteria that can be checked directly, (5) review questions about scope drift and secret exposure.
```

### Plan a coding task with boundaries

```
Use the task-plan-template. Define project name, goal, owner, current state, target state, and the work plan steps. Then fill in the agent assignment prompt with owned files, excluded files, acceptance criteria, and validation commands.
```

### Review agent output with a checklist

```
Use the review-checklist template. Walk through scope, behavior, code quality, verification, and security/privacy sections. For each unchecked item, note whether the agent's final response addressed it or not.
```

### Recover from scope drift and delegate cleanly

```
First, use the recovery prompt snippet: tell the agent to stop, return to the original acceptance criteria, and list any out-of-scope changes. Then, use the delegation snippet to split remaining work into independent tasks with disjoint write scopes, each with owned files, expected output, and validation commands.
```

## Limitations

- Does not execute or run coding agents — it is a documentation and template skill only
- Does not provide automated code review — it gives human reviewers a structured checklist
- Does not integrate with any external project management or version control tools
- Requires the user to manually fill in project-specific details in all templates

## Best Practices

- Write the shipped outcome as a single concrete sentence before assigning any work
- State explicitly which files the agent owns and which it must not touch
- Require the agent to report changed files, commands run, results, and remaining risks in its final response
- Review the diff before accepting work, prioritizing behavioral bugs and security issues over style

## Anti Patterns

- Assigning a vague task without scope boundaries or acceptance criteria
- Accepting agent output without inspecting the diff or running validation
- Letting the agent revert unrelated user work or invent new abstractions when existing patterns suffice

## Security Audit

- - Safe to publish: true
- - Audited at: 2026-06-10T04:57:24.486\+00:00
- - Summary: The static analyzer flagged 61 potential issues, but all are false positives on manual review. The skill contains only Markdown documentation and templates — no executable code. The 'Ruby/shell backtick execution' findings are Markdown code fences \(\`\`\`\), the 'weak cryptographic algorithm' matches are review checklist text about security, the 'hidden file in home directory' matches are install path instructions \(~/.claude/skills/\), and hardcoded URLs are the project's GitHub and Gumroad pages. No malicious intent or harmful behavior was found.

## Stats

- - Views: 0
- - Downloads: 2
- - Favorites: 0
- - Popularity score: 0