refrase 0.1.0

Model-specific prompt optimization. Your prompts, upgraded.

Refrase

Your prompts, upgraded.

Refrase optimizes your AI prompts for specific models using research-backed, model-specific formatting rules. No LLM required — it's deterministic and instant.

Why?

Most developers write one prompt and hope it works on every model. But models were trained on different data with different RLHF processes. Claude works best with XML tags. Qwen3 needs /think prefixes. Mistral's Magistral emits [TOOL_CALLS] markers that break JSON parsing. Refrase handles all of this automatically.

Quick Start

Python

pip install refrase

import refrase

result = refrase.adapt(
    "You are a helpful data analyst. Extract all key metrics from the report.",
    model="claude-sonnet",
    task="extraction",
)

print(result.system)
# <role>
# You are a helpful data analyst.
# </role>
#
# <instructions>
# Extract all key metrics from the report.
# </instructions>
#
# <output_format>
# Return structured output matching the schema.
# </output_format>

print(result.changes)
# [Change(rule='xml-structure', description='Restructured prompt with XML tags', ...)]

TypeScript

npm install refrase

import { adapt } from "refrase";

const result = adapt({
  prompt: "You are a helpful data analyst. Extract all key metrics.",
  model: "claude-sonnet",
  task: "extraction",
});

console.log(result.system);
console.log(result.changes);

Before / After

Model	Before	After
Claude Sonnet	Plain text prompt	XML-tagged <role>, <instructions>, <output_format>
Qwen3 32B	No thinking directive	/no_think prefix + Tier 3 JSON reinforcement + English enforcement
Magistral	Standard prompt	Suppressed [TOOL_CALLS] and [THINK] markers + ultra-strong JSON-only
GLM 4.7 Flash	Verbose prompt	Simplified + nested object fix + English enforcement
Nemotron 9B	Full prompt	/think prefix + simplified to 3 steps + Tier 3 JSON reinforcement

Supported Models

11 model families, 39+ model variants:

Family	Models	Key Adaptations
Claude	Sonnet, Opus, Haiku	XML tags, role extraction, Haiku simplification
OpenAI	GPT-4o, o1, o3	Reasoning hints, grounding rules
Gemini	Pro, Flash, Ultra	Identity (baseline — no changes)
Qwen	235B, 32B, Coder	Thinking mode, Tier 3 for 32B, English enforcement
DeepSeek	V3, V3.1, V3.2	Self-verification, JSON reinforcement
Mistral	Large, Magistral, Devstral, Ministral	Marker suppression, type fixes, simplification
Llama	405B, 70B, 8B, 3B	Grounding, small model simplification
Kimi	K2, K2.5	Source grounding, English enforcement
GLM	4.7, 4.7 Flash	Nested object fix, English, Flash simplification
Nemotron	30B, 12B, 9B	Thinking mode, Tier 3 for small, simplification
MiniMax	M2	Self-verification, contract-style

Task Types

Task	When to Use
extraction	Pulling structured data from documents
analysis	Evaluating, comparing, or reasoning about data
generation	Creating content, writing, summarizing
code	Writing or reviewing code
general	Everything else (default)

MCP Server

npm install -g @refrase/mcp-server

Add to your MCP config:

{
  "mcpServers": {
    "refrase": {
      "command": "refrase-mcp-server"
    }
  }
}

Adding New Model Adapters

See CONTRIBUTING.md for a 5-step guide to adding support for new models.

Research

Refrase's adaptation rules are derived from empirical testing across 46 model configurations on structured output tasks. Research paper coming soon.

License

MIT