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:

Letter	Component	Purpose	Key Question
C	Context	Background information	What does the agent need to know?
L	Limitations	Constraints, prohibitions	What should NOT be done?
E	Expectations	Output format, success criteria	What exactly should be returned?
A	Actions	Step-by-step instructions	What steps to follow?
R	Resources	Files, tools, references	What materials are available?

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

Strategy	When to Use	How
File-Based	Multi-step workflows	Save to file, pass path to next agent
Embedded	Small contexts	Include all info directly in prompt
Layered	Large contexts	Summary + references for deep-dive

State Transfer Pattern

Agent 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

Pattern	Use Case	Key Feature
Structured Markdown	Comprehensive results	Tables, sections, test output
Binary Verdict	Review/validation	PASS/FAIL first, then details
Checklist	Multi-requirement	Progress tracking with status
Condensed	Subagent responses	Max 3 findings, no verbose output

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

Anti-Pattern	Problem	Solution
Context Amnesia	Assumes agent remembers	Embed all context explicitly
Vague Expectations	"Give me a good result"	Specify exact format
Implicit Constraints	"Follow best practices"	List specific requirements
Unbounded Scope	"Improve the code"	Define precise boundaries
Missing Resources	"Check the docs"	Provide file paths
Output Overload	No format limits	Request condensed response

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

Templates

Ready-to-use templates for common scenarios:

Template	Purpose
templates/subagent-delegation.md	Task tool delegation
templates/code-task.md	Implementation tasks
templates/research-task.md	Research/exploration
templates/review-task.md	Code review/validation

Quick Start

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

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

Troubleshooting

Problem	Likely Cause	Solution
Irrelevant results	Insufficient context	Add explicit file references, examples
Wrong format	Vague specification	Provide template with placeholders
Unwanted changes	Missing constraints	Add explicit DO NOT statements
Incomplete task	Unclear success criteria	Add specific acceptance criteria
Agent asks questions	Ambiguous requirements	Make all decisions upfront

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

prompt-writing