---

name: prompt-writing description: Creates effective prompts for AI coding agents and subagent delegation. Use PROACTIVELY when writing prompts for Task tool delegation, creating agent instructions, or designing multi-step workflows. MUST BE USED when delegating complex tasks to subagents or writing system prompts. allowed-tools: ["Read", "Write", "Edit", "Grep", "Glob"]

Prompt Writing Skill

Overview

Core Principle: Context completeness over brevity. Agents cannot infer what you don't provide.

Target Users:

• Claude Code delegating to subagents via Task tool
• UI agents that need to generate prompts
• Engineers writing system prompts or agent instructions

---

The CLEAR Framework

Use this mnemonic for every prompt you write:

C - Context

Complete background: project state, environment, history. Include file paths, versions, prior decisions.

L - Limitations

Explicit boundaries: DO NOT modify X, use only Y, forbidden patterns, compliance requirements.

E - Expectations

Exact output format: markdown structure, JSON schema, where to save, success criteria checklist.

A - Actions

Numbered steps, validation checkpoints, decision points, error handling guidance.

R - Resources

Required reading (file paths), available tools, reference implementations, examples.

Full guide with examples: reference/prompt-anatomy.md

---

Context Patterns

Critical Insight

Subagents have NO conversation history. Each invocation is completely fresh. They cannot infer context.

Three Strategies

State Transfer Pattern

CopyAgent 1 → Save results to `.claude/state/results/agent-{timestamp}.md`
Agent 2 → Read from that file path

Full patterns: reference/context-patterns.md

---

Output Format Quick Reference

Delivery Patterns

• State directory: .claude/state/results/[agent]-[timestamp].md
• Project output: output/[task-id]/[artifact].md
• Inline: Return directly (use for summaries only)

Full templates: reference/output-specifications.md

---

Quality Checklist

Before sending any prompt, verify:

Context (C)

• All background information included?
• File paths explicit and complete?
• No assumptions about prior knowledge?

Limitations (L)

• Constraints clearly stated?
• Prohibited actions explicit?
• Scope boundaries defined?

Expectations (E)

• Output format specified exactly?
• Success criteria measurable?
• Delivery location defined?

Actions (A)

• Steps numbered and sequential?
• Validation checkpoints included?
• Error handling guidance provided?

Resources (R)

• Required files listed with paths?
• Available tools mentioned?
• Examples provided where helpful?

Subagent-Specific

• No implicit context assumptions?
• State transfer mechanism defined?
• Output format prevents overload?

---

Anti-Patterns Quick Reference

Full examples with solutions: reference/anti-patterns.md

---

Templates

Ready-to-use templates for common scenarios:

---

Quick Start

1. Choose template from templates/ matching your task type
2. Fill CLEAR sections systematically
3. Run checklist to verify completeness
4. Execute the prompt
5. Iterate if results don't match expectations

Remember: More context is almost always better. When in doubt, include it.

---

Troubleshooting

---

Reference Documentation

• reference/prompt-anatomy.md - Full CLEAR examples by language
• reference/context-patterns.md - State transfer strategies
• reference/output-specifications.md - Output format templates
• reference/anti-patterns.md - Common mistakes and fixes