Young Tries AI

Make your coding agent prove "done" with Hill Climb

This guide is for developers comfortable with git and the command line. You install Hill Climb, give your agent a measurable goal, and check progress from the log on disk.

CodingFor developers

Before you start

  • Claude Code or Codex, or another agent that reads SKILL.md skills.
  • Python 3.9+ and git. You need no packages, no server and no account.
  • A goal you can measure with a command, such as a test runner, a benchmark or a bundle size.
  1. This is option A, the plugin install. In Claude Code, add the marketplace first, then install the plugin. With the plugin install, the command is /hill-climb:hill-climb.

    Add the marketplace
    /plugin marketplace add Bil0000/Hill-Climb
  2. Run the second command to install the plugin from the marketplace you just added.

    Install the plugin
    /plugin install hill-climb@hill-climb
  3. This is option B, for any agent that reads SKILL.md. Skip this step if you used the plugin. Clone the repo, then copy the skill folder into your agent's skills directory: ~/.claude/skills for Claude Code or ~/.codex/skills for Codex.

    Clone the repo
    git clone https://github.com/Bil0000/Hill-Climb.git
  4. Do this only for option B. If you use Codex, make the same copy into ~/.codex/skills instead.

    Copy skill (Claude Code)
    mkdir -p ~/.claude/skills && cp -R Hill-Climb/skills/hill-climb ~/.claude/skills/
  5. Give the agent a goal. It builds the check, a command that scores the goal, and then climbs. In Codex, name the skill in your prompt. Other README examples are "/hill-climb fix every finding from the billing audit, backend and frontend" and "/hill-climb keep going until every test passes".

    Paste into your agent
    /hill-climb make the test suite 2x faster
  6. For a big goal, the agent first shows a one-screen plan, called the card, and waits for you to say go. A small goal starts at once. Each step that scores better is committed and every other step is undone. A step that edits the locked check is refused.

  7. The log on disk is the source of truth. The status command shows the score, the trail and notes, which helps if the session was compacted or crashed. The full check output of the last step is in last.txt.

    Run in the terminal
    climb.py status

    It worked if: A finished run prints BAR REACHED, for example: BAR REACHED: 0 meets 0.

Watch out for

  • A step that makes the score worse or equal, or fails a guard such as old tests, typecheck or lint, is reset. Expect some ideas to be undone; that is the design.
  • Locked files, meaning the tests, the benchmark and the fixtures, cannot change during a climb. A step that touches them is refused with "Locked files changed".
  • Noisy benchmarks can produce lucky wins. The README suggests --runs 3 --paired --min-gain 3% to reduce them.
  • After 5 flat steps the run reports PLATEAU and forces a blocker review.

Questions

What does Hill Climb need installed?
Python 3.9+ and git. It needs no packages, no server and no account.
What kinds of goals work?
Any goal you can turn into a number with a command. Examples are failing tests, milliseconds or work counts, bytes, outputs matching a real tool, or a score from a panel of model judges using your rubric.
What happens when a change helps?
It is committed, one commit per win, with the idea as the commit message, in every repo the climb touches. The branch only ever holds proven steps.
Can I use it with several repos or test runners?
Yes. The README shows climb.py init with several --check commands plus the --repo, --guard and --lock options. It points to references/checks.md for pytest, vitest, jest, rspec, unittest, node, cargo and go patterns.

Sources

Based on