Prompt author

Prompt Author

Creates high-quality, reusable prompts for AI workflows. This is the core skill — everything else (layouts, exports) flows from prompts.

When to Use

Trigger phrases (exact match):

• "write the master prompt"
• "draft the homepage prompt"
• "make a revision prompt"
• "create a prompt"
• "write a prompt for"
• "compose a prompt"
• "author a prompt"
• "generate a prompt"

Also triggers if:

• User describes a task and says "what prompt would do this?"
• User needs a prompt rewritten or optimized
• User says "this prompt isn't working, fix it"

Not a trigger if:

• User provides a prompt and says "run this" (that's execution, not authoring)
• User asks for the output the prompt would produce (that's generation, not authoring)

---

Before Starting

Ask (or infer) these:

1. Deliverable type: Image prompt, text prompt, or workflow prompt?
2. Subject: What is the prompt about?
3. Audience: Who will use this prompt? (affects complexity and specificity)
4. Platform: Which AI will run this? (Claude, Midjourney, DALL-E, GPT-4, etc.)
   • This affects syntax and constraints
5. Format: Single prompt or template with variables?
6. Brand file: Is there a BRAND.md to reference? (if not, use defaults)

Default if not specified:

• Type: text prompt (most common)
• Platform: Claude (default for opencode)
• Format: single prompt (no variables)
• Brand: technical + direct

---

Prompt Types and Anatomy

Type 1: Image Generation Prompt

Anatomy:

[Subject] + [Style] + [Composition] + [Technical specs] + [Mood/Atmosphere]

Example:

A modern SaaS dashboard interface, clean minimal design, light mode,
showing data visualization with line charts and bar graphs,
8K resolution, soft shadows, Figma-style UI design,
professional corporate aesthetic, morning light, screenshot quality

Platform-specific syntax:

Platform	Syntax notes
Midjourney	Use --ar for aspect ratio, --style for style codes
DALL-E 3	Natural language, no special syntax needed
Stable Diffusion	Use权重 weights (1.0), negative prompts for exclusions

---

Type 2: Text Generation Prompt

Anatomy:

[Role/Identity] + [Context] + [Task] + [Constraints] + [Output format] + [Tone]

Template:

You are a {role} with expertise in {domain}.
{Context: what the user already knows, what situation they're in}
{Task: what they need you to produce}
{Constraints: what to avoid, what must be included}
{Output format: how to structure the response}
{Tone: how to sound}

Example (blog post intro):

You are a senior engineer writing for a technical blog.
The reader is a fellow engineer evaluating whether to adopt a new library.
Write the first 3 paragraphs of a blog post about {topic}.
Start with a concrete problem, not a generic introduction.
No corporate language. Use "we" for Lightrains perspective.
Include one specific example with real metrics.
End with a hook that makes the reader want the next section.

---

Type 3: Workflow Prompt (Chain of Tasks)

Anatomy:

[Goal] + [Steps] + [Inputs] + [Outputs] + [Error handling]

Example:

Create a workflow that:
1. Takes a topic as input
2. Generates 3 related subtopics
3. For each subtopic, writes one paragraph
4. Combines into a coherent article
5. Outputs in markdown with ## headings

Inputs: {topic}
Outputs: {markdown article}
On failure: retry step, log error, continue if unsolvable

---

Step-by-Step Process

Step 1: Identify Prompt Type

Match the request to one of the three types:

Request	Type
"I need a prompt for a hero image"	Image
"Write a prompt for generating blog outlines"	Text
"Create a prompt that does X, then Y, then Z"	Workflow
Ambiguous	Ask: "Is this generating an image, text, or a multi-step process?"

---

Step 2: Extract Key Information

Ask (or infer) these for every prompt:

Required:

• What is the output? (image, text, action)
• Who is the audience?
• What is the context/situation?
• What must the output include?
• What must the output exclude?

Nice to have:

• Brand rules (from BRAND.md)
• Platform constraints
• Example of ideal output
• Anti-example of what to avoid

---

Step 3: Draft the Prompt

Follow this structure (in order):

1. Role assignment (if applicable)

   • Who is the AI acting as?
   • What expertise does it have?

2. Context provision

   • What situation are we in?
   • What does the user already know?

3. Task statement

   • What needs to be produced?
   • Use action verbs: "write", "generate", "create", "analyze"

4. Constraints

   • What to avoid (negative constraints)
   • What must be included (positive constraints)
   • Any rules or guardrails

5. Output format

   • Structure: markdown, JSON, list, paragraph
   • Length: short, medium, long
   • Any required elements

6. Tone guidance

   • How should it sound?
   • Any voice rules from brand file

Write prompts as if giving instructions to a capable junior — specific but not condescending.

---

Step 4: Optimize for Platform

Claude-specific optimizations:

• Use XML tags for structure: <context>, <task>, <output>
• Break complex tasks into numbered steps
• Include "think step by step" for reasoning tasks
• Add output validation instructions

Midjourney-specific optimizations:

• Start with subject
• End with style descriptors
• Use artist names for style references
• Include aspect ratio

DALL-E specific optimizations:

• Natural language, conversational
• Can include more detail than other platforms
• Avoid negative prompts (not supported)

---

Step 5: Add Variables (if template)

For reusable prompts, add variables:

Variable syntax:

{variable_name}

Example:

Write a blog post intro about {topic}
for {audience}
in {tone} tone

Variable types:

Type	Example	Notes
String	{topic}	Most common
Number	{count}	For iterations
Enum	{format: json/markdown}	Limited options
Boolean	{include_examples: true}	Binary choices

---

Step 6: Test the Prompt

For text prompts: Run the prompt and evaluate:

• Does it produce what was requested?
• Are constraints respected?
• Is the format correct?
• Is the tone appropriate?

For image prompts: Visualize mentally or run through:

• Are style descriptors coherent?
• Is composition clear?
• Are technical specs reasonable?

For workflow prompts: Trace through each step:

• Are dependencies correct?
• Can any step fail silently?
• Is the final output what's expected?

---

Step 7: Iterate Based on Test

Common fixes:

Problem	Fix
Too vague	Add specific constraints or examples
Too rigid	Remove unnecessary constraints
Wrong tone	Add tone guidance or reference brand file
Missing edge cases	Add conditional logic or error handling
Output wrong format	Add explicit format instructions

---

Output Formats

Single Prompt File

Save as prompts/{name}.md:

# {Prompt Name}

**Type:** {image/text/workflow}
**Platform:** {platform}
**Author:** {your name or 'agent'}

---

{Prompt text here}

---

## Variables

| Variable | Type | Description | Default |
|----------|------|-------------|---------|
| `{var1}` | string | What this controls | - |

## Example Usage

{Example with variables filled in}

## Notes

{Any caveats, tips, or context}

---

Prompt Collection File

For multiple related prompts, use a collection:

# {Collection Name}

## Prompts

### {Prompt 1}
...content...

### {Prompt 2}
...content...

---

## Usage Guide

{How to use these prompts together}

---

Quality Rubric

Rate each prompt on these dimensions:

Dimension	Score	Criteria
Clarity	1-5	Does the prompt say exactly what it wants? No ambiguity?
Specificity	1-5	Are constraints concrete, not vague?
Completeness	1-5	Does it include all needed context, format, and constraints?
Correctness	1-5	Is it syntactically correct for the target platform?
Reusability	1-5	Are variables clearly named? Is it templated appropriately?

Minimum passing scores:

• Clarity: 4+
• Completeness: 4+
• All others: 3+

If any score is below minimum:

• Identify the specific problem
• Fix it
• Retest

---

Common Pitfalls and Fixes

Pitfall 1: Vague Role Assignment

Bad:

You are a writer.

Good:

You are a senior backend engineer with 10 years of experience in distributed systems.
You have debugged production incidents at scale and understand operational trade-offs.

Pitfall 2: Missing Output Format

Bad:

Write about microservices.

Good:

Write a 2-paragraph explanation of microservices architecture.
Format:
- Paragraph 1: What it is (1 sentence) + why it matters (2 sentences)
- Paragraph 2: Main benefit (1 sentence) + trade-off (2 sentences)

Pitfall 3: Conflicting Constraints

Bad:

Write something short but comprehensive.

Fix:

Write exactly 3 sentences that cover the 3 most important points.
Prioritize: clarity > completeness > length.

Pitfall 4: Platform Mismatch

Bad for Claude (too verbose):

As a language model with extensive training data, I would like you to please
provide a detailed response regarding the topic of...

Good for Claude:

Provide a detailed response about {topic}.
Include: key points, examples, and a conclusion.

Pitfall 5: Variables Too Broad

Bad:

Write about {subject}.

Good:

Write a 200-word introduction about {topic} for {audience}.
Focus on {angle}: practical benefits, not theory.

---

Before/After Examples

Example 1: Image Prompt

Before:

Dashboard design for a SaaS app.

After:

A modern SaaS analytics dashboard, dark mode, featuring:
- Line chart showing 30-day user growth (upward trend)
- Bar chart showing revenue by plan (3 bars)
- KPI cards with percentage changes (+12%, +8%, -3%)
- User avatar cluster showing recent signups

Style: Clean Figma interface, soft shadows, subtle gradients
Colors: Dark navy background (#0F172A), white text, accent blue (#3B82F6)
Layout: Left sidebar (collapsed), main content area with 2-column grid
Quality: 8K, photorealistic, screenshot aesthetic
Mood: Professional, data-driven, confident

---

Example 2: Text Prompt

Before:

Write a blog post intro.

After:

Write the opening 3 paragraphs of a technical blog post.

Role: Senior engineer writing for peers
Audience: Developers evaluating a new tool or approach
Opening: Start with a specific problem, not a generic observation
Structure:
  - Paragraph 1: The problem (concrete, not abstract)
  - Paragraph 2: Why common solutions fail (specific reasons)
  - Paragraph 3: What we're about to show (the hook)

Constraints:
  - No "In this article..."
  - No corporate language
  - One specific example with metrics if possible
  - End with a question or challenge to the reader

Format: Markdown, ~200 words total

---

Example 3: Workflow Prompt

Before:

Generate content for a landing page.

After:

Create a landing page copy workflow that:

1. INPUT: {product_name}, {target_audience}, {primary_benefit}

2. Generate headlines (5 options):
   - Pattern A: Problem-first
   - Pattern B: Benefit-first
   - Pattern C: Question-based
   - Pattern D: Bold claim
   - Pattern E: WRYDWTD (What Race You Want to Die Doing)

3. For each headline, generate:
   - Subheadline (1 sentence)
   - 3 bullet points (benefit-focused)
   - CTA text (action-oriented, max 5 words)

4. OUTPUT FORMAT:
   ## Headline A: {headline}
   Subheadline: {text}
   Bullets:
   - {bullet 1}
   - {bullet 2}
   - {bullet 3}
   CTA: {text}

5. ERROR HANDLING:
   - If input missing: ask for {product_name} at minimum
   - If output too long: trim bullets to 2
   - If tone off: regenerate with "more {adjective}" instruction

---

Done Checklist

Before declaring complete:

• Prompt type identified
• Key information extracted (output, audience, constraints)
• Prompt drafted following anatomy template
• Optimized for target platform
• Variables defined (if template)
• Tested or mentally traced
• Iterated based on test results
• Saved to correct location (prompts/{name}.md or skill file)
• Quality rubric scores all pass minimums
• Example usage provided
• Notes/caveats documented

---

Cross-Skill Orchestration

Called by:

• User directly
• repo-bootstrap — if skills need prompts written
• bento-layout-director — for prompt templates

Feeds into:

• bento-layout-director — image prompts become layout concepts
• packaging-and-export — prompts are packaged as deliverables
• repo-bootstrap — prompts can be part of skill definitions

Depends on:

• brand-system-builder — for tone guidance (if brand file exists)

---

Edge Cases

Situation	Handling
Request is ambiguous	Ask one clarifying question: "Is this for images, text, or a workflow?"
User provides example output	Reverse-engineer the prompt from the example
Platform not specified	Default to Claude syntax, note the assumption
Prompt works but feels weak	Add specificity — more constraints, better examples
Multiple valid approaches	Show both, let user pick, explain trade-offs
User says "this isn't working"	Diagnose: wrong type? Missing constraints? Platform mismatch?
Prompt needs iteration	Show before/after, explain the improvement