Prompt Creation
Skill for creating high-performance prompts following 2025 best practices and Anthropic conventions.
Documentation
-
techniques.md - Structuring techniques (CoT, Few-shot, etc.)
-
templates.md - Reusable templates
Anthropic Official Structure (9 Elements)
Each system prompt should cover these 9 aspects:
- Task Context
Task Context
You are an expert [ROLE] specializing in [DOMAIN]. Your goal is to [PRIMARY_OBJECTIVE].
- Tone Context
Tone
- Tone: [formal/casual/technical/friendly]
- Length: [concise/detailed/adaptive]
- Format: [prose/lists/tables/mixed]
- Task Description + Rules
Task Description
Here are important rules:
- Always [MANDATORY_BEHAVIOR]
- Never [FORBIDDEN_BEHAVIOR]
- If [CONDITION] then [ACTION]
If unsure, say "I don't have enough information to answer."
- Examples (Few-shot)
Examples
<example> Input: [STANDARD_CASE_INPUT] Output: [EXPECTED_OUTPUT] </example>
<example> Input: [EDGE_CASE_INPUT] Output: [EXPECTED_OUTPUT] </example>
- Input Data
Input Data
<document> [USER_PROVIDED_CONTENT] </document>
- Immediate Task
Immediate Task
Now, [SPECIFIC_ACTION] the above [document/code/data].
- Precognition (Scratchpad)
Precognition
First, in <scratchpad> tags, analyze the key points. Then provide your final answer in <answer> tags.
- Output Formatting
Output Format
Analysis
[Problem understanding]
Solution
[Main response]
Recommendations
[Additional suggestions]
- Prefill (Assistant Turn)
Prefill
<scratchpad>
Emphasis Techniques
In order of increasing effectiveness:
Level Syntax Usage
1 Please do X
Suggestion
2 You should do X
Recommendation
3 Always do X
Standard rule
4 YOU MUST do X
Strong rule
5 IMPORTANT: Do X
Critical rule
6 CRITICAL - ZERO TOLERANCE: Do X
Absolute rule
Creation Workflow
- DEFINE the objective (1 sentence) ↓
- IDENTIFY target audience ↓
- LIST constraints ↓
- CHOOSE techniques (CoT? Few-shot?) ↓
- WRITE with 9-element structure ↓
- ADD guardrails ↓
- TEST mentally (edge cases) ↓
- ITERATE if necessary
Forbidden
-
Never create prompts without clear objective
-
Never write ambiguous instructions
-
Never forget error cases
-
Never ignore target model context