North star
Two agent skills to keep a long-running AI agent from getting stuck: north-star (delivery-point self-correction hook) + moa (N+1 multi-model consultation).
npx -y skills add ele-yufo/unstuck --skill north-starAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
长周期迭代项目(数据竞赛/研究/月级重构)的"交付点自我纠偏"机制——PostToolUse hook 每次写交付标记文件时把 vision+结构化 state+失败清单注入 agent 上下文,防 tunnel vision / sunk-cost / context 压缩失忆;stuck 阈值 breach 时联动 moa skill 做多模型会诊。含 install 模板。Triggers:长项目 / 竞赛 / Kaggle / 天池 / leaderboard / 分数卡住 / agent keeps trying same thing / context keeps getting lost / 自我纠偏 / 反思机制 / north star / 防止钻牛角尖。
SKILL.md
2.7 KB, as published. Nobody here has run it
North Star — 长周期项目的交付点自我纠偏
给数据竞赛 / 研究 / 月级重构这类长周期迭代项目装一个机制:PostToolUse hook 每次写交付标记文件(如 docs/CHANGELOG.md)时,把 vision + 结构化 state(consecutive_null_count 等)+ 失败清单注入 agent 下一轮上下文,防 tunnel vision / sunk-cost / context 压缩失忆。不是银弹——只强制在交付点反思。
红线
- 只对长周期项目装:按意图判——用户声明是长项目(竞赛/研究/月级重构),或你从对话看到长周期信号(远截止日 / 迭代指标 / 多阶段)。别对一次性脚本滥装(<1 天、纯探索)。
- hook 只在写触发文件时触发:注入是"交付时刻"事件,不是每轮都灌上下文。
- stuck 阈值 breach → 跑 moa skill:
consecutive_null_count ≥ 阈值时升级到moa做 N+1 多模型会诊。两种模式:框架内卡住喂 STATUS+FAILED_ATTEMPTS;"框架本身错"(上方有断层档 / 多样尝试同收敛 / gate 脱钩)走换框模式,只喂原始题、不喂 STATUS(升级链见references/mechanism.md)。 - merge 进既有 settings.json:项目已有 hook 就合并 PostToolUse 段,别 clobber。
- 机制不让你聪明:只强制反思 + 防 amnesia。别吹成银弹,也别 gaming STATUS。
场景路由表
| 你要做的事 | 做什么 / read |
|---|---|
| 判断是不是长周期项目 / 何时装 | 靠对话意图判断;判据见 references/install.md |
| 用户要装 / 正在装 | read references/install.md(4 定制输入 + 写 5 文件 + merge settings + 测 hook) |
| 理解机制 / MoA 联动 / 排错 | read references/mechanism.md(hook 怎么工作 + stuck→moa 升级链 + 限制) |
| review 安装或用法是否合规 | read references/gotcha-gallery.md |
加载纪律
判断/装项目前按需 read 对应 reference(装 → install.md,排错 → mechanism.md,自审 → gotcha-gallery.md)。不要一次 read 全部。机制本身(hook 注入什么、自检问什么)由装到项目里的 NORTH_STAR.md 驱动,不在本 skill 正文重复。