PromptUp Skill

Purpose: Transform user's vague requirements into production-grade prompts using evidence-based principles, not hardcoded templates.

Core Philosophy

Prompts are communication, not magic formulas.

A good prompt is clear communication with a capable assistant. There are principles, not templates.

When NOT to Use Prompt Engineering

Before writing any prompt, evaluate if prompt engineering is the right approach.

In Claude Code environment, you have these capabilities:

Signal	Better Approach	How in Claude Code
Need external/recent knowledge	Retrieval	Use Context7 MCP (official docs), Exa MCP (search), or Read (local files)
Need persistent behavior rules	Configuration	Define in CLAUDE.md or create a skill
Need real-time data	Tool use	Use WebSearch, WebFetch, or MCP tools
Need domain-specific knowledge	Retrieval + Config	Combine MCP lookup with skill-defined rules

If any signal matches: Use the appropriate tool/config instead of trying to solve with prompt alone.

---

Invocation Flow

Step 1: Understand the Task

Use AskUserQuestion to gather:

Question 1: Task Description
Header: "Task"
Options:
- "Generate content (articles, code, creative)"
- "Extract/classify data from text"
- "Analyze/reason about a problem"
- "Build a chatbot/assistant persona"
- "Create an agent with tools"

Follow-up based on answer:

• Ask for a concrete example of desired output
• Ask for constraints (length, format, tone)
• Ask what the prompt will be used for (one-off vs API integration)

Step 2: Check Boundaries

Evaluate if prompt engineering is sufficient:

Question	If Yes →
Does this need external knowledge?	Use Context7/Exa MCP or Read tool first
Does this need persistent behavior?	Suggest defining in CLAUDE.md or skill
Is this a one-time prompt for user's project?	Proceed with prompt engineering

If boundary crossed, use the appropriate capability instead of forcing a prompt-only solution.

Step 3: Apply Principles

Build the prompt using these evidence-based principles (from Anthropic official docs):

Principle 1: Be Explicit

BAD:  "Write something about AI"
GOOD: "Write a 500-word blog post explaining how transformers work to a non-technical audience"

Principle 2: Provide Context and Motivation

BAD:  "Never use ellipses"
GOOD: "Your response will be read by a text-to-speech engine, so never use ellipses since it won't know how to pronounce them"

The model generalizes from understanding WHY.

Principle 3: Say What TO Do, Not What NOT to Do

BAD:  "Don't use markdown"
GOOD: "Write in flowing prose paragraphs without formatting"

Principle 4: Structure with XML Tags

<context>
Background information here
</context>

<task>
What you want the model to do
</task>

<constraints>
- Length: 500 words
- Tone: Professional
- Format: Prose paragraphs
</constraints>

<output_format>
How the response should be structured
</output_format>

Principle 5: Examples Must Match Desired Behavior

If using few-shot, examples must be:

• Consistent in format
• Representative of edge cases
• Free of behaviors you don't want

BAD:  One example shows JSON, another shows plain text
GOOD: All examples show identical JSON structure

Principle 6: For Reasoning, Use Chain-of-Thought

<task>
Analyze this problem step by step.
</task>

<thinking_format>
Work through this systematically:
1. Identify key information
2. Break into sub-problems
3. Analyze each component
4. Form conclusion with reasoning
</thinking_format>

Step 4: Construct and Deliver

Output format:

## Prompt

[Complete prompt with XML structure]

---

## Usage Notes

- **Variables**: `{variable}` placeholders to replace
- **When to use**: [Specific scenarios]
- **Limitations**: [What this prompt can't do]
- **Iteration hints**: [How to improve if results aren't good]

Step 5: Offer Iteration

Use AskUserQuestion:

Question: Next Steps
Header: "Action"
Options:
- "Test this prompt and report results"
- "Modify the prompt"
- "Explain the design decisions"
- "Done"

If user reports issues, diagnose:

Symptom	Likely Cause	Fix
Too verbose	No length constraint	Add explicit length
Wrong format	No format example	Add output_format tag
Hallucinating	Task too vague	Add constraints, context
Inconsistent	Examples inconsistent	Fix example format
Ignoring instructions	Instructions buried	Move to top, use XML

---

Quality Checklist

Before delivering, verify:

• Task is explicit (no guessing required)
• Context explains WHY, not just WHAT
• Instructions are positive (do X, not don't Y)
• Structure uses XML tags for separation
• Examples (if any) are consistent
• Output format is specified
• Limitations are acknowledged

---

Anti-Patterns

Pattern	Problem	Fix
"Be creative"	Too vague	Define what creative means
"Don't hallucinate"	Negative instruction	"Only use information from the provided context"
"Think step by step" alone	No structure	Provide thinking format
Multiple roles in one prompt	Conflicting behaviors	Split into separate prompts
Prompt > 2000 tokens	Too complex	Simplify or use RAG

---

Advanced Techniques

Grounded Generation (RAG Pattern)

<context>
{retrieved_documents}
</context>

<grounding_rules>
- Only use information from the context
- Cite sources using [source_id]
- If not in context, say "I don't have this information"
</grounding_rules>

<question>
{user_question}
</question>

Agent Pattern (Tool Use)

<tools>
{tool_definitions}
</tools>

<behavior>
1. Determine if tools are needed
2. Select appropriate tool with correct parameters
3. Wait for results before proceeding
4. Use results to formulate response
</behavior>

<output_format>
When using tools: Action: {tool} / Input: {params}
When answering: Final Answer: {response}
</output_format>

Persona Pattern (Chatbot)

<persona>
You are {name}, a {role} at {org}.
- Personality: {traits}
- Expertise: {areas}
- Communication style: {style}
</persona>

<capabilities>
You can help with: {list}
</capabilities>

<rules>
- Always: {required}
- Never: {prohibited}
- When asked about {sensitive}: {handling}
</rules>

---

Evaluation Criteria

A good prompt should score well on:

Criterion	Question
Clarity	Can someone else understand what this prompt wants?
Specificity	Are edge cases and constraints defined?
Structure	Is information organized logically?
Completeness	Does it include all necessary context?
Minimalism	Is every part necessary? (No bloat)

---

Error Recovery

If user says "the prompt doesn't work":

1. Ask for specifics: What output did you get? What did you expect?
2. Check boundaries: Is this actually a prompt engineering problem?
3. Diagnose: Match symptoms to causes table above
4. Iterate: Make one change at a time, test

---

References

• Anthropic Claude 4 Best Practices (platform.claude.com)
• Anthropic Prompt Engineering Interactive Tutorial
• DAIR.AI Prompt Engineering Guide