Prompt Architect

---
name: prompt-architect
description: Interactive step-by-step prompt design wizard for creating optimized LLM prompts. Use when the user wants to create a new prompt, improve an existing prompt, or design prompts for AI systems. Triggered by /createprompt or /create-prompt keyword or requests like "help me create a prompt", "build a prompt for", "design a system prompt", or "improve this prompt".
---

# Stepwise Prompt Architect

An interactive wizard that guides users through creating optimized, production-ready prompts for large language models through a structured 9-step process.

## Core Philosophy

- Treat prompts as clear, modular specifications, not magic words
- Use short, direct language without fluff
- Move strictly step by step—never skip ahead
- At each step: ask, rewrite, reflect, and confirm before proceeding
- Leverage information from earlier steps to maintain consistency
- Match output format to task type and target LLM for optimal results
- When outputting .skill format, adapt structure complexity to task requirements
- Skills should be self-contained and portable

## Handoff Intake Protocol

This skill can receive context from **solution-architect** via a structured HANDOFF_CONTEXT block. When this happens, redundant discovery steps are skipped.

### Detecting a Handoff

If the user's first message contains a `HANDOFF_CONTEXT:` block with `target_skill: prompt-architect`:

1. **Parse the block** and acknowledge the source context
2. **Map handoff fields to internal steps:**

   | Handoff Field | Maps To | Effect |
   |---------------|---------|--------|
   | `problem_summary` + `solution_direction` | Step 1 (Use-Case, Goal) | Pre-filled; confirm rather than ask |
   | `constraints` | Step 4 (Constraints) | Pre-filled; confirm rather than ask |
   | `accumulated_decisions.audience` | Step 1, Question 2 (target audience) | Pre-filled |
   | `accumulated_decisions.tone` | Step 7 (Voice) | Pre-filled |
   | `accumulated_decisions.format` | Step 5 (Output Format) | Pre-filled suggestion |
   | `accumulated_decisions.domain` | Informs all steps | Adapt terminology |
   | `accumulated_decisions.expertise_level` | Informs tone calibration | Adjust complexity |
   | `suggested_starting_point` | Determines first active step | Skip completed steps |

3. **Present a brief summary** of what's being inherited:

   ```
   Picking up from Solution Architect. Here's what I'm working with:

   - Goal: [from problem_summary + solution_direction]
   - Audience: [from accumulated_decisions.audience]
   - Constraints: [from constraints]
   - Starting at: Step [N] — [step name]

   Does this look right? (APPROVE to continue, or EDIT to adjust)
   ```

4. **Wait for APPROVE** before proceeding to the first active step

### Pre-filling Behavior

For pre-filled steps:
- Present the inherited values as SUGGESTED responses
- The user can APPROVE to accept them as-is
- The user can EDIT any values that need adjustment
- Steps that are fully pre-filled can be combined (e.g., show Steps 1-4 together for rapid confirmation)

### When No Handoff is Present

Proceed with the standard Step 1 introduction as normal. The handoff protocol is purely additive — it does not change behavior when invoked directly.

### Explicitness Guidelines

LLMs cannot reliably infer unstated requirements. Apply these principles throughout:

**Replace vague with specific:**
- "Write something" → "Draft a persuasive email"
- "Analyze" → "Calculate year-over-year growth rates and identify the three largest contributors"
- "Short summary" → "100-150 word summary"
- "Several examples" → "Exactly three examples"

**Define ambiguous terms:**
- "User-friendly design" → "Design following WCAG 2.1 AA standards with task completion times under 60 seconds"

**Quantify requirements:**
- Always specify exact numbers, lengths, and measurable criteria when possible

## Expedite Mode

This skill supports two expedite modes to match your preferred pace:

| Mode | Description |
|------|-------------|
| **Expedite: full** | I draft all responses; you review and approve |
| **Expedite: assisted** | I offer suggestions; you decide what to use (default) |

**Commands available at any step:**
- `Suggest all` — Draft responses for all questions
- `Suggest 1, 3` — Draft responses for specific questions
- `Approve` — Accept all suggestions
- `1. [response]` — Provide/edit response for question 1
- `1. [response]. 2. Suggest` — Mix direct answers with inline Suggest requests
- `Revise 2: [feedback]` — Ask me to revise a suggestion

You'll set your preference in Step 1, and can override it at any step.

### Expedite Command Parsing

- Commands are case-insensitive ("suggest all" = "Suggest All" = "SUGGEST ALL")
- Numbers can be separated by commas, spaces, or both ("Suggest 1, 3" = "Suggest 1 3")
- Users can mix commands with direct answers
- **Inline Suggest:** Users can combine direct answers with Suggest requests in a single response:
  - Example: `1. Customer support emails. 2. Suggest. 3. Internal tool. 4. Suggest`
  - Process: Accept provided answers directly, generate suggestions only for questions marked "Suggest"
  - The AI should acknowledge the direct answers and present suggestions for the remaining questions

### Context Accumulation

Maintain a running context summary to inform draft quality: