Prompt Decorators extend the functionality of large language models by allowing structured, modular control over reasoning, style, and behavior. Each decorator enforces specific response rules, enabling users to declaratively modify how the model thinks and writes — without retraining.
New: Prompt Decorators is now available as an Agent Skill, making it easy to integrate into AI and agentic tools. Jump to the Agent Skill section ↓
- Prompt Decorators Framework
Prompt Decorators are lightweight control markers (for example, +++Reasoning) that alter the response generation process.
They can be combined, scoped, and remembered during conversation, forming a flexible behavioral layer over standard prompting.
The Prompt Decorators framework is organized into two main families, as illustrated in the taxonomy diagram below:
These decorators guide how the model thinks, explores, and develops ideas.
| Subcategory | Purpose | Key Decorators |
|---|---|---|
| Reasoning & Generation | Enforce analytical, logical, or multi-perspective reasoning. | +++Reasoning, +++Debate, +++Socratic |
| Exploration, Interaction & Planning | Encourage open-ended exploration, inquiry, and planning. | +++Interactive, +++Brainstorm, +++Planning, +++Import |
| Inquiry & Clarification | Ensure information gaps are identified and clarified before proceeding. | +++Interactive, +++Rewrite |
| Idea Development & Structuring | Refine, structure, and extend ideas into coherent outputs. | +++Rewrite, +++Planning |
| Evaluation, Feedback & Refinement | Critically assess and improve content through iteration. | +++Critique, +++Refine, +++Candor |
These decorators control how the model expresses results, manages sessions, and ensures consistency.
| Subcategory | Purpose | Key Decorators |
|---|---|---|
| Output Formatting & Expression | Structure or constrain the format and tone of output. | +++OutputFormat, +++Tone |
| Meta-Control & Session Management | Manage decorator state, memory, and operational flow. | +++ChatScope, +++MessageScope, +++Clear, +++ActiveDecs, +++AvailableDecs, +++Validate, +++FactCheck, +++CiteSources, +++Export, +++Dump |
Multiple decorators can be combined to refine responses. For example:
+++Reasoning
+++StepByStep
+++Tone(style=formal)
Explain the economic impact of renewable energy transitions.
This prompt enforces logical reasoning, structured progression, and formal tone simultaneously.
| Prompt Decorator | Parameters | Function |
|---|---|---|
+++Reasoning |
— | Ensures logical explanation before answering |
+++StepByStep |
— | Breaks down complex tasks into sequential reasoning |
+++Socratic |
— | Encourages critical thinking through guided questions |
+++Debate |
— | Generates multiple perspectives before concluding |
+++Critique |
— | Analyzes strengths and weaknesses before improving |
+++Refine(iterations=N) |
iterations – number of refinement cycles |
Iterates through multiple improvements |
+++Interactive(limit=N, style=TYPE) |
limit, style |
Asks clarifying questions when context is incomplete |
+++Planning |
— | Outlines objectives and approach before execution |
+++Brainstorm(limit=N, diversity=LEVEL) |
limit, diversity |
Produces diverse ideas without early evaluation |
+++Rewrite |
— | Reframes the user’s prompt for clarity and precision |
+++Import(topic=STRING) |
topic – conceptual or disciplinary lens |
Applies a named lens to reasoning |
+++Candor(level=LEVEL) |
level – low, medium, or high |
Controls feedback directness while staying professional |
+++OutputFormat(format=FORMAT) |
format – output type |
Structures the response in a specific format |
+++Tone(style=STYLE) |
style – tone or register |
Adjusts the communication tone or style |
+++FactCheck |
— | Verifies factual accuracy of key claims |
+++CiteSources |
— | Ensures all major claims are supported by sources |
+++Validate |
— | Performs self-check on grammar, logic, and consistency |
+++ChatScope |
— | Applies decorators persistently across a conversation |
+++MessageScope |
— | Limits decorator effects to a single message |
+++Clear(targets=LIST) |
targets – decorators to clear (optional) |
Clears all or specified active decorators |
+++ActiveDecs |
— | Lists currently active decorators |
+++AvailableDecs |
— | Lists all available decorators and their status |
+++Export(format=FORMAT) |
format – text, markdown, json, yaml |
Exports or summarizes conversation in a chosen format |
+++Dump |
Alias of +++Export |
Quick or raw output version of +++Export |
Decorator scope controls how long and where a decorator applies:
- Message Scope (
+++MessageScope) → Affects only the current prompt. - Chat Scope (
+++ChatScope) → Persists across conversation turns. - Clear (
+++Clear) → Resets decorators selectively or entirely.
The system retains awareness of active decorators to ensure consistent behavior between exchanges.
Detailed definitions for each decorator are provided in prompt-decorators.txt.
Each definition specifies:
- Behavioral rules
- Response structure
- Parameters (if applicable)
- Compliance and scope requirements
Prompt Decorators rely on the model’s memory of definitions. To enable them in chat-based LLMs (such as ChatGPT, Gemini, Claude, or Grok):
- Provide the definitions file once — paste the contents of
prompt-decorators.txtinto the chat or add it to your custom/system instructions. - The model will then “remember” the rules and behaviors of each decorator for that session.
- You can activate decorators inline by prefixing them in your prompt, for example:
+++Reasoning
+++Tone(style=formal)
Explain why renewable energy transition affects global markets.
- Multiple decorators can be combined to modify reasoning, tone, structure, and verification layers.
- To reset behaviors, use
+++Clearor restart the session.
This approach makes decorators portable and platform-agnostic — no code modification required.
Below are simple examples illustrating how one, two, or three decorators change the model’s behavior. These examples use natural language inputs that can be pasted directly into an LLM chat.
+++Reasoning
Why did the Roman Empire fall?
Expected behavior: The model explains its reasoning process before answering, giving a logical chain of causes.
+++StepByStep
+++Tone(style=formal)
Describe how neural networks learn from data.
Expected behavior: The model breaks down the explanation into sequential steps, maintaining a formal tone.
+++Brainstorm(limit=5, diversity=high)
+++Critique
+++Refine(iterations=3)
Propose innovative applications for large language models in healthcare.
Expected behavior:
- The model first generates five diverse ideas.
- Then it critiques their strengths and weaknesses.
- Finally, it refines and improves the top ideas over two iterations.
These compositional examples show how decorators can stack logically to form complex cognitive workflows without retraining the model.
A dedicated Agent Skill implementation of the Prompt Decorators framework is available, allowing it to be seamlessly integrated into a wide range of AI and agentic tools.
Unlike the chat-based approach (pasting the definitions file), the Agent Skill packages the framework as a reusable skill that can be plugged directly into agentic pipelines, AI assistants, and tool-calling workflows. No manual setup per session required.
The repository also hosts additional skills beyond Prompt Decorators, including:
evo-search(Evolutionary Search) — applies a mechanism inspired by genetic and evolutionary algorithms to iteratively improve the quality of AI-generated responses through selection, mutation, and refinement cycles.tournament-search(Tournament Search) — simulates a competition-style tournament among candidate responses, making it effective for search and discovery tasks where the best answer must be identified from a diverse set of options.
More skills are planned and will be added to the repository over time.
For more information, installation instructions, and examples, visit the dedicated repository:
👉 https://github.com/smkalami/skills
The Prompt Decorators framework is described in detail in the paper:
"Prompt Decorators: A Declarative and Composable Syntax for Reasoning, Formatting, and Control in LLMs"
Mostapha Kalami Heris
arXiv preprint arXiv:2510.19850 (2025)
📄 Read the full paper: https://arxiv.org/abs/2510.19850
@misc{heris2025promptdecorators,
title={Prompt Decorators: A Declarative and Composable Syntax for Reasoning, Formatting, and Control in LLMs},
author={Mostapha Kalami Heris},
year={2025},
eprint={2510.19850},
archivePrefix={arXiv},
primaryClass={cs.PL},
url={https://arxiv.org/abs/2510.19850}
}This framework is open-source under the MIT License. You are free to adapt, extend, or integrate it in your own prompt engineering workflows with attribution.