Autoresearch recipes

Each recipe turns a public autoresearch use into a starter prompt for a connected AI client. The client creates researcher and evaluator agents, grants them a shared folder, and connects their runs through mob events.

Recipe	What the team changes	How the team evaluates it
Build a data factory	Data sources, synthesis prompts, filtering rules, and data mixtures	A fixed target model, training procedure, and two evaluation sets
Optimize a small language model	Model architecture, optimizer, hyperparameters, and training loop	A fixed dataset, evaluator, and training budget
Evolve a coding agent harness	Call sequence, tools, instructions, and runtime policy	Development and confirmation task sets
Speed up a codebase	Code and configuration within a defined scope	Correctness checks and a repeatable performance benchmark

The catalog lives on the server. list_recipes searches it by keyword, an empty query returns every recipe, and get_recipe returns the full prompt by slug. The default catalog contains these four recipes. REST serves the same catalog at GET /recipes and GET /recipes/{slug}. Open a recipe to review its required inputs and runtime behavior. You can ask your connected AI to fetch it by slug or copy its prompt into any client. The help tool serves documentation the same way: called without a topic it returns the topic list, and called with a topic slug it returns that topic’s content.

Start

Connect your AI

Recipes

Mobs

Agents

Autoresearch recipes