intelli-verse-x Prompt Design Skill
Purpose
Create prompts that are clear, effective, reproducible, and maintainable.
Prompt Structure
## Role
You are a [specific role]. You [context and constraints].
## Task
[Specific, imperative instruction]
## Input
[Format and content of input]
## Constraints
- MUST: [hard requirements]
- MUST NOT: [prohibitions]
- FORMAT: [output structure]
## Examples
[2-3 few-shot examples]
## Output
[Exact output format, schema, or template]
Principles
- Specificity: Precise instructions beat vague ones
- Examples: Few-shot examples for complex tasks
- Constraints: Explicit boundaries (do/don't)
- Format: Structured output specification
- Brevity: Every token should earn its place
CF Prompt Registry
File Structure
prompt_registry/
├── _index.yaml # Prompt catalog + versions
├── _schema.json # Validation schema
└── learning_series/
├── script_writer.yaml
├── visual_director.yaml
└── voice_selector.yaml
YAML Format
id: learning_series.script_writer
version: "1.2.0"
description: "Writes educational scripts for learning series"
template: |
You are an educational script writer. Write a {minutes}-minute script about {topic}...
variables:
- name: minutes
type: int
min: 1
max: 60
- name: topic
type: str
Code Usage
from prompt_registry import get_prompt_or_default
INLINE_FALLBACK = "Write a {minutes}-minute script about {topic}..."
template = get_prompt_or_default("learning_series.script_writer", default=INLINE_FALLBACK)
prompt = template.format(minutes=5, topic="Inertia")
Optimization
- A/B Test: Version A vs B on eval dataset
- Tokens: Compress without losing clarity
- Position: Important instructions at start and end
- Context: Reference external docs by ID, don't inline
Anti-Patterns
- ❌ Hardcoded in Python strings
- ❌ Reading from disk in hot loops
- ❌ > 4000 tokens without compression
- ❌ Vague: "be creative"
- ❌ Mixing system and user content
Evaluation
- Success rate on eval set
- Token usage per call
- Latency (time to first token)
- Human rating (1-5 scale)