---
name: prompt-writing
description: Write effective prompts for AI systems — system prompts, agent instructions, skills, or any LLM prompt. Use when creating or improving prompts.
---
# Prompt Writing
Good prompts are clear and concise, not long and verbose. Longer prompts cause attention dilution — the model pays less attention to each part.
## What Improves Prompts
### Remove Redundancy
Same instruction said multiple ways dilutes attention.
**Before:** "You must always verify each item. Do not skip items. Every item needs to be checked. Make sure you don't miss any items."
**After:** "Verify each item individually. No skipping."
### Remove Noise
Don't teach the model things it already knows. State your specific requirements.
**Before:** 40-line example of what a code walkthrough looks like
**After:** "Write a detailed walkthrough: what changed, line numbers, analysis of the flow, data transformations, dependencies."
### Sharpen Instructions
"Do X" is clearer than "You should consider doing X because..."
**Before:** "It's important that you remember to always navigate to the project root directory before starting any work, as this ensures that all file paths will be correct..."
**After:** `cd "$CLAUDE_PROJECT_DIR"`
### Keep Load-Bearing Content
These must stay:
- Workflow steps and their order
- Quality criteria
- Critical rules and constraints
- Output format requirements (if parsed programmatically)
- Behavioral guardrails
### Write Clean Prose
Write as if it was always this way — not "correction" style.
**Before:** "Remember that you should do TWO interviews, not just one. The first interview is for business requirements, and then after discovery you should do a second interview..."
**After:**
```
### 0. Business Requirements Interview
Interview the user to understand what needs to be built.
### 2. Technical Interview
With discovery complete, interview the user about implementation details.
```
## Structure
Good prompts have clear sections. Adapt this pattern to your use case:
```
[1-2 sentence role/purpose]
## Core Concept or Approach
[Key principle guiding behavior]
## Workflow / Steps
[What to do, in order]
## Rules / Constraints
[Non-negotiable requirements]
## Quality Criteria
[What good looks like, what to avoid]
## Output Format
[Expected structure if applicable]
```
For procedural agents, number the steps clearly. For guidance prompts, use descriptive sections.
## Common Patterns
**Workflow Steps:** Numbered, clear action, what to produce, when to proceed.
**Quality Criteria:** Two lists — DO (what matters) and DON'T (what to ignore). Prevents false positives and negatives.
**Output Formats:** If parsed programmatically, include exact schema. If human-readable, describe what to include.
## Checklist
- [ ] Every instruction serves a purpose (no redundancy)
- [ ] No verbose examples of things the model knows
- [ ] Instructions are direct, not hedged
- [ ] Load-bearing content preserved
- [ ] Written in clean prose, not "correction" style
