WritePrompt

Framework for creating, modifying, and reviewing agentic prompts following the 7-level prompt hierarchy.

Core Principle

The prompt is THE fundamental unit of agentic engineering.

Every prompt should have the right level of complexity for its use case — no more, no less.

Workflow Routing

Workflow	Trigger	File
Write	"create prompt", "new command", "modify prompt", "update workflow", "edit prompt", "improve prompt"	`Workflows/Write.md`
Review	"review this prompt", "check prompt quality", "audit prompt"	`Workflows/Review.md`
Harden	"harden prompt", "add validation", "add hooks", "self-validation", "add self-validation"	`Workflows/Harden.md`

The 7 Prompt Levels

Level	Name	Key Feature	When to Use
1	High Level	Static, reusable	Simple one-shot tasks
2	Workflow	Sequential steps with I/O	Multi-step processes
3	Control Flow	Conditions and loops	Dynamic branching logic
4	Delegate	Spawns subagents	Parallel or distributed work
5	Higher Order	Accepts prompt files as input	Composable prompt systems
6	Template Metaprompt	Creates new prompts	Prompt generation
7	Self-Improving	Updates its own expertise	Learning systems

Quick Section Reference

Essential sections (most prompts need these):

•Metadata — YAML frontmatter (allowed-tools, description, argument-hint, model)
•# Title — Clear, action-oriented name
•## Purpose — What it does and when to use it
•## Variables — Dynamic ($1, $ARGUMENTS) and static values
•## Workflow — Numbered steps to execute
•## Report — How to present results

Advanced sections (for complex prompts):

•## Instructions — Rules, constraints, edge cases
•## Relevant Files — Files to read/modify
•## Codebase Structure — Directory layout context
•## Expertise — Accumulated domain knowledge (Level 7)
•## Template — Output format for metaprompts (Level 6)
•## Examples — Usage demonstrations

Examples

Example 1: Create a new slash command

code

User: "Create a command to summarize git history"
→ Invokes Write workflow (create mode)
→ Asks clarifying questions (output format? date range?)
→ Determines Level 2 (Workflow) is appropriate
→ Creates .claude/commands/git-summary.md with proper sections

Example 2: Modify existing prompt

code

User: "Add a verification step to the Check workflow"
→ Invokes Write workflow (modify mode)
→ Reads existing prompt, identifies what needs to change
→ Adds the new step, renumbers subsequent steps
→ Updates the prompt file

Example 3: Review prompt quality

code

User: "Review this prompt and suggest improvements"
→ Invokes Review workflow
→ Identifies current level and sections
→ Suggests missing sections or level upgrade
→ Provides concrete recommendations (does NOT modify)

Example 4: Complex automation prompt

code

User: "Create a command that processes multiple files in parallel"
→ Invokes Write workflow (create mode)
→ Recognizes need for Level 4 (Delegate)
→ Includes Task tool delegation pattern
→ Adds proper agent configuration variables

Example 5: Add self-validation to existing agent

code

User: "Harden the normalize-csv command with validation hooks"
→ Invokes Harden workflow
→ Reads the command, identifies CSV output + Edit/Write tool usage
→ Recommends Stop hook with csv-validator.py
→ Writes validator script, adds hooks YAML to frontmatter
→ Verifies the validator runs without errors

References

•PromptLevels — Detailed breakdown of all 7 levels with examples
•PromptSections — Complete section reference with usage guidelines
•SelfValidationHooks — Hook types, JSON protocol, validator architecture, design patterns