opty-prompt 0.1.11

Prompt optimization via chain-of-thought

opty

Prompt optimization via chain-of-thought.

opty helps you fine-tune your LLM prompts using a judge LLM and iterative chain-of-thought refinement. Provide a description, give one or two input/output examples, and opty will automatically improve your prompt until the target model's output matches your expectations.

How it works

1. Describe what you want your prompt to do, or load an existing draft from a file
2. opty generates an initial prompt draft (the builder LLM), asking clarifying questions if needed
3. Provide one example input and the ideal output you expect
4. opty runs your draft prompt against the example input using the refiner LLM, captures its chain-of-thought, and analyzes where the thinking went wrong
5. Based on that analysis, the refiner rewrites the prompt to steer the model in the right direction
6. Steps 4–5 repeat until the output matches your target or the step limit is reached
7. The final optimized prompt is saved to a file of your choice

Installation

pip install opty-prompt

Usage

Create an opty.config.yaml in your working directory, then run:

opty-prompt

Configuration

opty has two independent roles:

• Builder — generates the initial prompt draft from your description. This is a straightforward text generation task, so it works well with smaller, faster models.
• Refiner — evaluates the prompt against your example, reads the model's chain-of-thought, and rewrites the prompt accordingly. This requires stronger reasoning, so it benefits significantly from a more capable model.

Each role is configured independently and can use a different provider and model. You can freely mix and match Gemini and Ollama for either role.

---

All Gemini

config:
  builder:
    type: gemini
    model: gemini-2.5-flash
    api-key: <your-api-key>
  refiner:
    type: gemini
    max-steps: 4
    model: gemini-2.5-flash
    api-key: <your-api-key>

---

All Ollama (local)

config:
  builder:
    type: ollama
    model: gemma3:4b-it-qat      # smaller model is fine for building
    ollama-server: http://127.0.0.1:11434
  refiner:
    type: ollama
    max-steps: 4
    model: gemma3:12b-it-qat     # use a stronger model for refinement
    ollama-server: http://127.0.0.1:11434

---

Mixed: Ollama builder + Gemini refiner

Run a lightweight local model for drafting and offload the heavier refinement reasoning to Gemini:

config:
  builder:
    type: ollama
    model: gemma3:4b-it-qat
    ollama-server: http://127.0.0.1:11434
  refiner:
    type: gemini
    max-steps: 4
    model: gemini-2.5-flash
    api-key: <your-api-key>

---

Mixed: Gemini builder + Ollama refiner

config:
  builder:
    type: gemini
    model: gemini-2.5-flash
    api-key: <your-api-key>
  refiner:
    type: ollama
    max-steps: 4
    model: gemma3:12b-it-qat
    ollama-server: http://127.0.0.1:11434

---

Configuration reference

Shared — all types

These keys apply to both builder and refiner regardless of the provider:

Key	Required	Description
type	Yes	Provider to use. One of: gemini, ollama

Shared — refiner only

Key	Required	Default	Description
max-steps	No	4	Maximum number of refinement iterations before stopping

Gemini

Key	Required	Description
model	Yes	Gemini model name (e.g. gemini-2.5-flash, gemini-2.5-pro)
api-key	Yes	Your Google Gemini API key

Ollama

Key	Required	Description
model	Yes	Ollama model tag (e.g. gemma3:4b-it-qat, llama3.1:8b)
ollama-server	Yes	Base URL of your Ollama server (e.g. http://127.0.0.1:11434)

---

Model size guidance

Role	Requirements	Examples
Builder	Light — basic instruction following	gemma3:4b, gemini-2.5-flash
Refiner	Heavy — strong reasoning and chain-of-thought	gemma3:12b+, gemini-2.5-pro