使用方法
- 复制安装命令，在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物，并对照来源检查输出。
npx skills add https://github.com/gnibu/bento --skill bento-hillclimb开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况，不等于独立质量或安全测试。
按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Hillclimb
The failure mode: change three things at once, the number moves, you don't know which tweak did it — or whether it was noise. Next week it drops and you can't retrace. One variable at a time, logged, keep-or-revert.
Repo layer first. Read the repo's AGENTS.md/CLAUDE.md and, if it exists, the overlay .agents/bento/bento-hillclimb.md at the repo root; follow it under the overlay rules.
Steps (paste as todos)
- Freeze the harness first. Fix the dataset, the metric, and how it's measured
before touching the thing you're tuning. If the ruler moves, no result means anything.
- Record the baseline. Run the frozen harness once. Write the number down.
- Change ONE variable. One knob per experiment. Never stack untested changes.
- Measure against baseline. Same harness, same input set.
- Keep or revert — and log it either way. Keep only a real gain (significant, not
noise; re-run if the sample is small). Log the attempt, the delta, and the decision. A reverted attempt is still knowledge — write it down so you don't retry it.
- New baseline on a keeper, then back to step 3.
- Don't stop on a lucky first hit. Run enough iterations that one good roll can't
fool you. State the stop condition up front (target metric, or N no-improvement rounds).
Guardrails
- Regression check on every keeper: a gain that breaks something else is not a gain.
- If two knobs seem coupled, that's a finding — still change one, note the interaction.
- The log is the deliverable as much as the final number. No log = you learned nothing
transferable.
