hillclimb
Skill12 starsChanged 7 days agoMITView on GitHub
反复改动、测量,把一个指标(性能、命中率、通过率、评分、体积)推向目标时使用。
12 other projects use a version of this.
Adopted by
Most recently taken up by
- Yeachan-Heo/oh-my-claude-sisyphustheir own version, first seen Nov 8, 2026
- Yeachan-Heo/oh-my-claudecodetheir own version, first seen Nov 8, 2026
- paureis/pau-skillstheir own version, first seen Nov 8, 2026
- Beskars-Code/Skillstheir own version, first seen Nov 1, 2026
- Gabriel-Lafrance/Skillstheir own version, first seen Nov 1, 2026
- zombico/mojulotheir own version, first seen Oct 25, 2026
Used with
Projects with it also use grill-me.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
GoogleSkillgrill-me
A relentless interview to sharpen a plan or design.
Also in 3 projects that use this
Other versions12 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 12
- Yeachan-Heo/oh-my-claude-sisyphus★ 41ksame name
- Yeachan-Heo/oh-my-claudecode★ 41ksame name
- rebase-energy/hillclimb★ 37same name
- michielvandermeer/skills★ 9same name
- zombico/mojulo★ 4same name
- jonmumm/skills★ 2same name
- paureis/pau-skills★ 2same name
- Astatide1337/Skills★ 1same name
- kadeski/skills★ 1same name
- Beskars-Code/Skills★ 0same name
2 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository28 other files it gives its agents.
- AGENTS.mdAGENTS.md
- gpt-imageSkill
- postSkill
- promoSkill
- videoSkill
- cuepadSkill
- architecture-auditSkill
- better-testSkill
- project-setupSkill
- prototypeSkill
- prSkill
- shipSkill
- write-swiftSkill
- ui-craftSkill
- ui-picksSkill
- better-browser-useSkill
- better-computer-useSkill
- herdrSkill
- sshSkill
- researchSkill
- session-dataSkill
- web-searchSkill
- fan-outSkill
- grillingSkill
- retroSkill
- eli5Skill
- writing-for-agentsSkill
- writing-for-humansSkill
The file
Files
爬山
一次改一处,测一次,保留或回退。agent 的典型失败是前几步进步很快,随后在一个好实现、但有天花板的方向里反复调参;这个流程的关键是认出停滞并换方向。
- 定指标与停止条件:一个指标、哪边算好、目标值,再加最少尝试次数,免得早期一次偶然的好结果就收工。用户给了数字就用用户的。完成标准:写下“指标 / 方向 / 目标 / 至少 N 次尝试”。
建评测并冻结:一条命令输出指标,多次采样取中位数并报噪声幅度,同时跑回归检查(必须保持通过的测试与检查)。
- 按生产配置跑(模型、思考档、工具),单次运行设超时上限。
- 用例分优化集和留出集。留出集只在收尾确认时跑,结果不拿来改方案;被拿去找过问题的留出题换新题。
- 评分表和模型裁判先拿几份已知好坏的输出试一遍,确认奖励的是想要的行为(出现过因“缺地址价格”扣分、等于鼓励编造的评分表);人工核对一小批,确认裁判和人一致。整场比较只用这一个裁判。
- 被优化的一方碰不到评测脚本、用例与参考答案;正确性失败记 0 分。
- 评测不动生产账号余额、真实设备和共享额度:用独立的 key 或设备。按量付费的余额每批前查一次;会自动换号的订阅不设额度停止线,以调用真的失败为准(只看一个号的额度,会把还能用的调用拦下)。
- 评分表、阈值、用例或裁判一改,旧结果就不可比:换一个批次名重跑基线。
- 指标或诊断要用真实会话与用量数据时,先用 session-data,不现写脚本扫原始文件。
完成标准:基线、噪声幅度、回归检查结果和每批的预计花费都已记下。
- 开决策记录:在评测目录(不进仓库)放一张表,每次尝试一行:编号、假设、改动、前后指标、回归检查、保留或回退、备注。每次尝试前先读它;中断后据它接着做,已有结果不重跑。完成标准:表已建好,基线是第一行。
- 循环,每轮一个假设:
- 假设写明机制(“X 在启动路径上阻塞首帧,挪到首帧之后”),不写“试试加缓存”。只针对基线多次运行里重复出现的失败;只出现一次的是噪声,专门改它就是过拟合。
- 只改这一处。超过噪声且回归检查通过才保留,否则整个回退;保留时在运行记录里指出改动起作用的那一刻,只有分数涨了不够。不叠加没测过的改动,不凭读代码宣布变好。
- 互不依赖的假设各派一个子代理、各用一个 worktree 并行试;你只出假设、看 diff、裁决。几个候选各自过关后合成一版再测一次,不默认叠加。
- 改动落到最便宜的有效部件:重复的机械活交给脚本,窄流程做成按需技能;写进提示词的规则,模型几次里有一次以上没照做,就不再改措辞,改由脚本或检查保证。
- 完成标准:这次尝试在决策记录里有一行;回退的改动已撤干净,保留的改动叠在一起时回归检查仍通过。
- 停滞就换方向:连续 2–3 次回退,或连续几次提升都落在噪声里,就停止在同一方向上微调。
- 先查测量:各次尝试用的测量和正式评测是不是同一口径、同一条件(出现过团队一直在另一种负载下测速度,校准测量后当轮就破了纪录)。
- 回看决策记录,判断是不是选了一条有天花板的路,然后换一类做法:换算法或架构、换数据或用例的组织、换模型、把几个差一点的改动组合起来,或者试一个更激进的方案;互斥的几个方向并行试。
- 换方向也可以是换分工:加一个专门查测量、专职调试或读运行记录找卡点的角色。执行者说“已经到顶”不算停止理由,那是当前做法的顶。
- 换过方向仍然卡住,带着决策记录交给用户,不空转。
- 完成标准:决策记录里写明这次停滞的判断(测量问题,还是方向的天花板)和换到的新方向。
- 收尾:达到停止条件,或剩下的想法收益抵不上成本时停;不放宽停止条件来凑数。在留出集上确认,再用生产配置在干净检出上复测一次。汇报:指标从基线到最终的变化、尝试次数(保留/回退)、每个保留的改动一行、决策记录的位置,以及下一个最值得试的想法。
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

