Centralized Prompts System

ARTEMIS uses a centralized prompt management system for versioning, A/B testing, and maintainability. All prompts are stored in artemis/prompts/ and accessed through a unified API.

Why Centralized Prompts?

• Versioning: Test new prompt versions without code changes
• A/B Testing: Compare prompt performance across versions
• Maintainability: Single source of truth for all prompts
• Rollback: Quickly revert to previous prompt versions

Usage

Basic Usage

from artemis.prompts import get_prompt, list_prompts

# Get a prompt by key
prompt = get_prompt("hdag.strategic_instructions")

# Get prompt with formatting variables
prompt = get_prompt(
    "evaluation.user",
    topic="AI Safety",
    level="strategic",
    content="...",
    round=1,
    total_rounds=3,
    position="pro",
    prev_count=2,
)

# List all available prompts
prompts = list_prompts()
# Returns: {'hdag': ['STRATEGIC_INSTRUCTIONS', ...], 'evaluation': [...], ...}

Prompt Key Format

Prompts are accessed using dot notation: module.prompt_name

• hdag.strategic_instructions - H-L-DAG strategic level prompt
• evaluation.system - Evaluation system prompt
• extraction.causal_extraction - Causal link extraction prompt
• jury.analytical_perspective - Analytical juror perspective

Available Modules

Module	Description	Key Prompts
hdag	H-L-DAG argument generation	STRATEGIC_INSTRUCTIONS, TACTICAL_INSTRUCTIONS, OPERATIONAL_INSTRUCTIONS, OPENING_STATEMENT, CLOSING_STATEMENT
evaluation	LLM-based evaluation	SYSTEM, USER, CRITERIA_DEFINITIONS, DEFAULT_WEIGHTS
extraction	Evidence/causal extraction	CAUSAL_EXTRACTION, EVIDENCE_EXTRACTION
jury	Jury perspectives	ANALYTICAL_PERSPECTIVE, ETHICAL_PERSPECTIVE, PRACTICAL_PERSPECTIVE, ADVERSARIAL_PERSPECTIVE
benchmark	Benchmark evaluation	ARGUMENT_QUALITY, DECISION_ACCURACY, REASONING_DEPTH

Versioning

Prompts are organized by version (v1, v2, etc.) for controlled experimentation.

Default Version

The default version is v1. All get_prompt() calls use this unless specified otherwise.

from artemis.prompts import get_prompt_version

current = get_prompt_version()  # Returns "v1"

Switching Versions

Global Version Switch

from artemis.prompts.loader import set_prompt_version

# Switch all prompts to v2
set_prompt_version("v2")

# Now all get_prompt() calls use v2
prompt = get_prompt("hdag.strategic_instructions")  # Uses v2

Per-Call Version

from artemis.prompts import get_prompt

# Use v2 for this specific call
prompt = get_prompt("hdag.strategic_instructions", version="v2")

# Default version unchanged for other calls
other_prompt = get_prompt("evaluation.system")  # Still uses v1

Adding New Prompts

1. Add to Existing Module

Edit the appropriate file in artemis/prompts/v1/:

# artemis/prompts/v1/hdag.py

# Add new constant (uppercase)
MY_NEW_PROMPT = """Your prompt content here.

Variables use {curly_braces} for formatting:
Topic: {topic}
"""

Access with:

prompt = get_prompt("hdag.my_new_prompt", topic="AI Safety")

2. Create New Version

Copy the v1/ directory to v2/ and modify prompts:

artemis/prompts/
├── v1/
│   ├── hdag.py
│   └── ...
└── v2/          # New version
    ├── hdag.py  # Modified prompts
    └── ...

Register in artemis/prompts/loader.py:

PROMPT_MODULES = {
    "v1": {...},
    "v2": {
        "hdag": "artemis.prompts.v2.hdag",
        # ... other modules
    },
}

Prompt Module Structure

Each prompt module follows this structure:

# artemis/prompts/v1/example.py
"""
Example Prompts - Version 1

Description of what these prompts are for.
"""

# Prompts are uppercase constants
SYSTEM_PROMPT = """You are an assistant..."""

USER_PROMPT = """Given the following:
Topic: {topic}
Context: {context}

Please provide your analysis."""

# Supporting constants (also accessible)
DEFAULT_CONFIG = {
    "temperature": 0.7,
    "max_tokens": 1000,
}

Integration Example

Here's how the prompt system integrates with ARTEMIS components:

from artemis.prompts import get_prompt
from artemis.core.types import Message

# In LLM evaluation
system_prompt = get_prompt("evaluation.system")
user_prompt = get_prompt(
    "evaluation.user",
    topic=context.topic,
    level=argument.level.value,
    content=argument.content[:2000],
    round=context.current_round,
    total_rounds=context.total_rounds,
    position=position,
    prev_count=len(context.transcript),
)

messages = [
    Message(role="system", content=system_prompt),
    Message(role="user", content=user_prompt),
]

response = await model.generate(messages=messages)

Best Practices

1. Use descriptive names: STRATEGIC_INSTRUCTIONS not PROMPT1
2. Document variables: Comment which {variables} the prompt expects
3. Keep prompts focused: One purpose per prompt
4. Test before deploying: Verify new versions in benchmarks
5. Version incrementally: Small changes per version for easier debugging