Hyperstition

Convert your skills to evals

Give your agents the tools to improve themselves

Then tell your agent: set up hyperstition. How →

How it works

Your skills become evals. Your agent reads the results and improves its work.

  1. 1 · SKILLS

    Start from your skills

    hyp reads the SKILL.md files your agents already follow and turns each rule into a narrow eval.

    5 skills → 186 evals

  2. 2 · RUN

    Run on each change

    hyp splits the diff into small units and a judge answers one question per unit. Tests, types and CI sit beside them.

    108 ran · 78 not applicable

  3. 3 · FINDINGS

    Point at lines

    Each indicator names the eval and the files and lines worth a look.

    20 indicators · 34 findings

  4. 4 · DECIDE

    Decide and improve

    The agent decides each finding, fixes what it chose to fix, and the next run shows what changed.

    6 fix  2 explain
    26 proceed

Numbers from one run of hyp on a pull request to its own repository, 17 October 2026.

Install

One line in your terminal.

$ curl -fsSL https://hyp.dev/install | sh
# then tell your agent: set up hyperstition
  1. Run the line in your terminal. It needs Node.js 22.18 or later and git. It installs the hyp command with npm and puts the skills in ~/.claude/skills for Claude Code and ~/.agents/skills for other agents. A re-run upgrades both. On Windows, run npm install -g hyperstition@latest, then hyp init.
  2. Ask your agent to set up hyperstition. The setup skill takes it through the judge, the evals for your skills, where each one runs, and CI. It asks you each choice that is yours.
  3. A judge that needs an API key? Save it with hyp key set in your own terminal.
Read the setup guide npm