Custom Rules - Flower Engine

Overview

Rules are modular system prompts that modify AI behavior during sessions. They allow you to dynamically adjust tone, content restrictions, narrative style, and other behavioral aspects without editing world or character files.

Why Use Rules?

Rules provide flexibility:

Modular: Apply different rulesets to the same character/world combination
Reusable: Use the same rules across multiple sessions
Dynamic: Activate or deactivate rules mid-session
Layered: Stack multiple rules for combined effects

Rules are injected into the system prompt at runtime. They modify AI behavior without altering the underlying world lore or character persona.

Rule File Structure

Rules are defined as YAML files in the assets/rules/ directory.

Required Fields

string

required

Unique identifier for the rule. Used in commands to activate/deactivate.

name

string

required

Display name shown in the TUI rules selector.

prompt

string

required

The instruction text that modifies AI behavior. This is injected into the system prompt.

Basic Example

From assets_example/rules/gritty.yaml:

id: "gritty"
name: "Gritty Realism"
prompt: "The world is dark, dangerous, and unforgiving. Actions have heavy consequences."

Create the YAML file

touch assets/rules/my_rule.yaml

Define the rule

Add the required fields with your rule content.

Restart the engine

Rules are loaded on startup:

python -m engine.main

Activate the rule

/rules add my_rule

Advanced Examples

Tone Modifiers

horror.yaml

id: "horror"
name: "Horror Atmosphere"
prompt: |
  Emphasize dread, tension, and the unknown. Use sensory details to create
  unease: flickering shadows, distant sounds, things glimpsed in peripheral
  vision. Build slowly. The most terrifying things are those not fully seen.
  
  Avoid:
  - Jump scares or sudden gore
  - Explaining everything (mystery is key)
  - Making the protagonist helpless (fear is strongest when there's a chance)

Content Restrictions

family_friendly.yaml

id: "family_friendly"
name: "Family Friendly"
prompt: |
  Keep all content appropriate for ages 13+. No graphic violence, sexual content,
  excessive profanity, or disturbing themes. Handle conflict and danger in an
  adventure-appropriate way. Focus on heroism, problem-solving, and positive
  character growth.

Narrative Style

noir_style.yaml

id: "noir"
name: "Noir Narration"
prompt: |
  Narrate in classic noir style:
  - Use hardboiled, cynical language
  - Employ similes and metaphors ("She was trouble in a red dress")
  - Describe environments atmospherically (rain, shadows, neon, smoke)
  - Keep dialogue snappy and loaded with subtext
  - Everyone has secrets, nothing is as it seems
  
  Example tone: "The rain came down like it had a grudge. In this city, even
  the weather knew you were guilty of something."

Mechanical Rules

combat_detailed.yaml

id: "combat_detailed"
name: "Detailed Combat"
prompt: |
  During combat scenes:
  - Describe actions in specific physical detail
  - Track wounds and fatigue realistically
  - Consider environment, positioning, and tactics
  - Make combat risky and consequential
  - Victory should require strategy, not just power
  
  Example: Instead of "You hit him", write "Your fist connects with his jaw,
  snapping his head back. He staggers, spitting blood, but stays on his feet."

Content Expansion

nsfw.yaml

id: "nsfw"
name: "NSFW Mode"
prompt: |
  You are in NSFW mode. You may include:
  - Adult sexual content and explicit descriptions
  - Graphic violence and gore
  - Dark, disturbing, or taboo themes
  - Strong profanity
  
  Always maintain:
  - Consent between characters in intimate scenes
  - Narrative coherence and character consistency
  - Quality writing (explicit ≠ low quality)
  
  The user is an adult who has opted into this content.

NSFW rules remove content restrictions. Only use with appropriate models and in private sessions.

Rule Loading Process

Rules are loaded from YAML files at startup:

# From engine/state.py:58-61
available_rules = [
    {"id": d["id"], "name": d.get("name", d["id"])}
    for d in load_yaml_assets("assets/rules/*.yaml")
]

When activated, rules are injected into the system prompt:

# From engine/llm.py:83-86
for rule_id in state.ACTIVE_RULES:
    with open(f"assets/rules/{rule_id}.yaml", "r", encoding="utf-8") as f:
        data = yaml.safe_load(f)
        rule_prompts.append(data.get("prompt", ""))

All active rule prompts are concatenated and included in the system message sent to the LLM.

Activating Rules

Add a Rule

/rules add gritty

From engine/commands.py:76-89:

if len(parts) >= 3 and parts[1] == "add":
    rule_id = parts[2].strip()
    if os.path.exists(f"assets/rules/{rule_id}.yaml"):
        if rule_id not in state.ACTIVE_RULES:
            state.ACTIVE_RULES.append(rule_id)
        state.save_state()

The rule is immediately active for all subsequent messages.

Clear All Rules

/rules clear

Removes all active rules, returning to base behavior.

Quick NSFW Toggle

/nsfw

Toggles the nsfw rule on/off (if the rule file exists).

Rules are saved to persist.json and persist across engine restarts.

Stacking Multiple Rules

You can activate multiple rules simultaneously:

/rules add noir
/rules add combat_detailed
/rules add gritty

All rule prompts are injected into the system message in the order they were added. This allows layered behavior modification:

noir: Sets narrative style
combat_detailed: Defines how combat is handled
gritty: Adds consequence and tone

The result: noir-style narration with detailed, consequential combat.

Conflicting rules may produce unpredictable results. For example, combining “horror” and “comedy” rules might confuse the AI.

Rule Categories

Organize rules by purpose:

Tone Rules

Modify emotional atmosphere:

horror.yaml: Dread and tension
lighthearted.yaml: Fun and comedic
epic.yaml: Grand and heroic
melancholic.yaml: Sad and reflective

Content Rules

Control what appears:

family_friendly.yaml: Age-appropriate content
nsfw.yaml: Adult content
low_violence.yaml: Minimal combat
high_violence.yaml: Graphic action

Style Rules

Define narrative approach:

noir.yaml: Hardboiled detective style
poetic.yaml: Lyrical descriptions
terse.yaml: Minimal, clipped sentences
verbose.yaml: Rich, detailed prose

Mechanical Rules

Specify how systems work:

combat_detailed.yaml: Blow-by-blow combat
dialogue_focused.yaml: Emphasize conversation
investigation.yaml: Clue-based mysteries
sandbox.yaml: Player-driven exploration

Best Practices

Keep rules focused

Each rule should do one thing well. Don’t create mega-rules that try to control everything.

Use clear, directive language

Write rule prompts as instructions, not descriptions. “Emphasize X” is better than “X is important.”

Provide examples

Show the AI what you mean with concrete examples in the rule prompt.

Test rule combinations

Before using rules in important sessions, test how they interact with your characters and worlds.

Version your rules

If you update a rule significantly, consider creating a new version (e.g., combat_v2.yaml).

Rule Template

template.yaml

id: "rule_identifier"
name: "Human Readable Name"
prompt: |
  Clear, directive instructions for the AI.
  
  What to do:
  - Specific behavior 1
  - Specific behavior 2
  - Specific behavior 3
  
  What to avoid:
  - Unwanted behavior 1
  - Unwanted behavior 2
  
  Example of desired output:
  [Concrete example here]

Runtime Rule Management

View Active Rules

The TUI displays active rules in the state panel. Programmatically:

# From engine/state.py:48
ACTIVE_RULES = []  # List of active rule IDs

Persisted State

Active rules are saved to persist.json:

{
  "world_id": "cyberpunk_city",
  "character_id": "detective",
  "session_id": "abc123def456",
  "model": "gpt-4o-mini",
  "rules": ["noir", "combat_detailed"]
}

Rules reactivate automatically when restarting the engine.

System Prompt Injection

Rules are injected into the system prompt before each LLM call:

# Simplified from engine/llm.py
system_parts = [
    base_system_prompt,
    character.persona,
    world.system_prompt,
    *rule_prompts,  # All active rules
]

system_message = "\n\n".join(system_parts)

Order of precedence:

Base engine instructions
Character persona
World system prompt
Active rules (in activation order)

Later instructions can override earlier ones, so rules have high priority.

Creating Rule Collections

Create themed rule sets for different play styles:

Investigation Pack

assets/rules/investigation/
├── mystery_focus.yaml
├── clue_based.yaml
└── cerebral_tone.yaml

Activate all at once:

/rules add mystery_focus
/rules add clue_based
/rules add cerebral_tone

Combat Pack

assets/rules/combat/
├── tactical.yaml
├── detailed_combat.yaml
└── high_stakes.yaml

assets/rules/social/
├── dialogue_focused.yaml
├── political_intrigue.yaml
└── relationship_dynamics.yaml

The engine loads rules from subdirectories if you update the glob pattern to assets/rules/**/*.yaml.

Troubleshooting

Rule not found

If /rules add fails:

Verify the rule file exists in assets/rules/
Check that the ID matches the filename (without .yaml)
Ensure the YAML is valid and has an id field
Restart the engine to reload rule assets

Rule not affecting behavior

Check that the rule prompt is clear and directive
Verify the rule is in the active rules list
Test with a stronger model (some models ignore weak instructions)
Ensure other rules aren’t contradicting it
Review the system prompt to see if rule injection is working

Rules conflicting

Clear all rules and add them one at a time to identify conflicts
Reorder rule activation (later rules override earlier ones)
Rewrite conflicting rules to be more compatible
Use fewer, more focused rules instead of many broad ones

Command Reference

Command	Description
`/rules add <rule_id>`	Activate a rule
`/rules clear`	Deactivate all rules
`/nsfw`	Toggle NSFW mode (if rule exists)

Get Started

Core Concepts

Guides

Advanced

​Overview

​Why Use Rules?

​Rule File Structure

​Required Fields

​Basic Example

​Advanced Examples

​Tone Modifiers

​Content Restrictions

​Narrative Style

​Mechanical Rules

​Content Expansion

​Rule Loading Process

​Activating Rules

​Add a Rule

​Clear All Rules

​Quick NSFW Toggle

​Stacking Multiple Rules

​Rule Categories

​Tone Rules

​Content Rules

​Style Rules

​Mechanical Rules

​Best Practices

​Rule Template

​Runtime Rule Management

​View Active Rules

​Persisted State

​System Prompt Injection

​Creating Rule Collections

​Investigation Pack

​Combat Pack

​Social Pack

​Troubleshooting

​Rule not found

​Rule not affecting behavior

​Rules conflicting

​Command Reference

​Next Steps

Creating Worlds

Creating Characters

Build docs developers (and LLMs) love

Overview

Why Use Rules?

Rule File Structure

Required Fields

Basic Example

Advanced Examples

Tone Modifiers

Content Restrictions

Narrative Style

Mechanical Rules

Content Expansion

Rule Loading Process

Activating Rules

Add a Rule

Clear All Rules

Quick NSFW Toggle

Stacking Multiple Rules

Rule Categories

Tone Rules

Content Rules

Style Rules

Mechanical Rules

Best Practices

Rule Template

Runtime Rule Management

View Active Rules

Persisted State

System Prompt Injection

Creating Rule Collections

Investigation Pack

Combat Pack

Social Pack

Troubleshooting

Rule not found

Rule not affecting behavior

Rules conflicting

Command Reference

Next Steps