rl-env-use

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

rl-env-use

Introducing rl-env-use: RL environments for anyone and any domain.

Describe a skill in one line. Get a verified environment that frontier models do not saturate, and a hillclimbed prompt for the model you train. Built on Anthropic’s Automating eval design and hillclimbing workflow, with one change: the tasks are real cases that Browser Use agents mine from the web.

Missions	Claude Code	Browser Use
8 x clinical trial results	0 of 86 facts verified	17 verified, 4 complete cases
6 x clinical trial results, same model both sides	0 of 43	3 verified, 2 complete cases
8 x human genetic variants (UniProt, gnomAD, GTEx)	6 of 82	40 verified

Get a Browser Use API key at Browser Use Cloud.

What it does

• Mines real cases: Browser Use agents find the materials (a trial protocol, a 10-Q) and the outcome published later
• Writes hard tasks: gold answers computed by code, anything a frontier panel solves gets rewritten harder
• Rejects bad tasks: leaked answers, memorized outcomes, ambiguous prompts
• Hillclimbs a model’s system prompt with a train/test split, a noise floor and paired confidence intervals
• Exports the dataset for an RL trainer, or serves it in code with Env

Setup prompt

For Claude Code / Codex / OpenCode:

Set up https://github.com/browser-use/rl-env-use for me.

Read install.md first to install this repo, register the skill, and set up the Browser Use and OpenRouter API keys - ask me to paste them when you need them. Then read SKILL.md for daily usage. After install, don't build anything on your own - just tell me it's ready and wait for me to name a skill.

Then ask:

build me an RL environment for predicting whether a clinical trial met its primary endpoint, from its ClinicalTrials.gov registration

For humans:

rl-env-use create "predict whether a completed trial met its primary endpoint from its registered protocol; ClinicalTrials.gov" --budget 40
rl-env-use climb envs/ --model qwen-35b

Results

The shipped example, envs/trial-forecast: 35 forecasts from 29 ClinicalTrials.gov trials whose results were posted since July 2026, after the models’ training data. Built for $72, $17 of it Browser Use.

A climbed Qwen 3.6 35B-A3B goes from 0.53 to 0.78 on tasks it never trained on (+0.25, 95% CI +0.07 to +0.45), level with Opus 5.5. Always answering 50% scores 0.75, so most of the gain is calibration: the model stopped answering 0 or 1. The climb kept 1 of 12 proposed edits (base-rate priors per question type); every round is in climb-qwen-35b/.

Manual install

Python 3.11+, optionally Node.js 18+ for the HTML reports.

git clone https://github.com/browser-use/rl-env-use ~/Developer/rl-env-use
cd ~/Developer/rl-env-use
uv venv && uv pip install -e '.[test]'
cp .env.example .env && $EDITOR .env    # BROWSER_USE_API_KEY, OPENROUTER_API_KEY
.venv/bin/python -m pytest -q           # offline, ~10 s

See SKILL.md for the full workflow and where every file is saved.

License

MIT, see LICENSE.