Skip to content

coaxer

Online: https://thekevinscott.github.io/coaxer/

Label examples. Derive the prompt. Consume it as a string.

The prompt is a build artifact; your labeled examples are the source of truth. When the prompt drifts, add more examples and recompile.

Install

uv add coaxer

npm install coaxer

Quick start

Label a handful of examples in a folder (see Getting Started for the shape), then:

coax labels/repo-classification --out prompts/repo-classification

Use the compiled prompt:

from coaxer import CoaxedPrompt

p = CoaxedPrompt("prompts/repo-classification")
filled = p(readme=new_readme, stars=1200)

import { CoaxedPrompt } from "coaxer";

const p = new CoaxedPrompt("prompts/repo-classification");
const filled = p({ readme: newReadme, stars: 1200 });

Core concepts

  • Label folder — one directory per record. record.json holds scalar fields; sibling files (readme.md, logo.png) carry large text or binary inputs.
  • coax — reads the folder, builds a DSPy signature internally, optionally runs GEPA, writes a prompt artifact.
  • CoaxedPrompt(path) — a str subclass in Python (callable instance in TS). The raw template is the string value; calling renders it. Drops in anywhere a string is accepted. Exposes a structured-output schema (p.response_format / p.responseFormat()) for OpenAI's .parse() and Anthropic's tool-use.
  • AgentLM / OpenAILM — LLM backends for the optional compile-time optimizer. See the API reference.