Skip to content
Back to search

tashan

SkillWorks with Claude CodeGitHub

Use for complex problems, unclear goals or constraints, unknown causes, repeated failures with standard approaches, or explicit requests for deep analysis or first principles. Clarify what matters, derive and verify a solution, and explain it clearly.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/simulationdute--tashan/badge.svg)](/listings/simulationdute--tashan)

HTML

<a href="/listings/simulationdute--tashan"><img src="/api/listings/simulationdute--tashan/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

把整个 `tashan/` 目录放入所用 Agent 的技能目录。以 CodeArts Agent 的项目级技能目录为例:

<project>/.codeartsdoer/skills/tashan/

其他 Agent 的目录约定不同,核心要求只有一条:`SKILL.md` 所在目录名为 `tashan`,且 frontmatter 中的 `name` 与目录名一致。

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 39 lines
  1. ---
  2. name: tashan
  3. description: >-
  4. Use for complex problems, unclear goals or constraints, unknown causes, repeated failures with standard approaches, or explicit requests for deep analysis or first principles. Clarify what matters, derive and verify a solution, and explain it clearly.
  5. ---
  6. # Tashan
  7. ## 1. Match the effort to the problem
  8. Answer straightforward questions directly and stop: if one or two sentences settle it, that is the whole answer, and do not show the steps unless the user asks. Use the steps below only when working through them would change the answer.
  9. ## 2. Clarify what could change the solution
  10. Separate the desired outcome from the proposed approach. Check available evidence first. Ask only about unresolved facts or tradeoffs that would change the decision. Group questions the user can answer together. If a follow-up is easy to answer and applies to only one branch, use "If ..., please also confirm ...". Wait for the first answer if it determines the next question or could spare the user an investigation. Explain your recommendation and what the choice changes. Proceed once enough is known.
  11. ## 3. Reason from first principles
  12. Separate facts from assumptions and identify what causes the outcome. Work backward from success to find what must hold and what can change. If the current approach is stuck, try reallocating resources, changing the sequence, or separating coupled steps within the constraints. For example, condensing steam inside the cylinder also cooled its walls, which had to be reheated on the next cycle. Watt kept the cylinder hot and moved condensation to a connected cold vessel, preserving the pressure drop that drives the piston.
  13. Treat anything the user has not supplied as unknown, not as given. A quantity you did not measure, or a fact you only expect, is an assumption: label it as one and say which branch changes if it is false, or leave it out. Never let an unmeasured quantity drive the conclusion.
  14. Reduce the result to one sentence: what has to be decided, allocated, or kept stable, and what makes it hard. That sentence is the input to the next step.
  15. ## 4. Find the classic problem, borrow its method, then rebuild
  16. Ask which classic problem the stripped-down statement is a variant of, and test the mapping before writing it: name the causal structure, and state in one sentence what causes what here and which of the classic method's assumptions still hold. If you cannot do that, there is no usable mapping: skip this step entirely, say nothing about an analogy, and continue to the next step. Borrow mechanisms, not vocabulary; mark each assumption that fails and what replaces it.
  17. Then set the analogy aside and re-derive from the basic facts. A mature method carries its own field's constraints and a settled solution shape; this problem's constraints differ, so recombine the borrowed parts into what this problem needs — change the sequence, split a coupled step, keep the mechanism and drop the conventional form. If the re-derivation lands on the classic method for a reason you can state, keep it; otherwise follow the derivation.
  18. For example, "peak traffic will overload checkout" reduces to arrival rate against finite service capacity, the classic queueing problem, so borrow its capacity calculation. But queueing theory assumes interchangeable, patient requests, while orders have priority and inventory limits. Keep the calculation, replace the queue with admission control and priority ordering.
  19. ## 5. Check the solution
  20. Test it against evidence or a counterexample. For any required guarantee, identify what can enforce it with the available resources and controls. State unverified prerequisites as conditions. If the constraints rule out a necessary prerequisite, explain why the result is infeasible and what would need to change. For a decision that is costly to get wrong, name one test that could falsify the conclusion and the observation that should stop the plan early; for cheap or reversible decisions, skip this.
  21. ## 6. Give a clear answer
  22. State the conclusion first, then the key reasons, what to do, and how to verify success. Use plain language, with a simple analogy when it helps. Keep the technical detail the task needs.
README.md · README · 78 lines
  1. # Tashan
  2. Tashan 是一套可复用的思考指令(Agent Skill)。当问题复杂、目标不清,或常规方法反复失败时,它把分析过程固定为可执行的步骤:先澄清影响决策的事实,再从第一性原理推导因果,借用机制相同的成熟方法,最后用证据检验结论并给出可验证的答案。
  3. ## 它解决什么问题
  4. 一次性回答擅长处理直截了当的问题,但在以下情况容易失效:约束没弄清就开始给方案、沿用熟悉做法却反复碰壁、把假设当成事实、给出结论却没有验证路径。
  5. Tashan 针对的是这些失效模式。它不提供领域知识,只约束思考顺序。
  6. ## 适用场景
  7. - 问题复杂,多个因素相互制约
  8. - 目标或约束不明确,方案取决于某个关键取舍
  9. - 原因未知,需要从现象反推因果
  10. - 常规做法反复失败,需要跳出既有路径
  11. - 明确要求深度分析或第一性原理思考
  12. 简单直接的问题不需要它。Tashan 的第一步就是让投入与问题匹配,而不是对所有问题都展开长分析。
  13. ## 工作方式
  14. 1. **匹配投入**:能一两句定的事,就一两句答完并停,不展示过程;复杂问题才使用后续步骤。
  15. 2. **澄清**:区分期望结果与当前方案,先查已有证据,只问会改变决策的未决事实或取舍。
  16. 3. **第一性原理**:区分事实与假设,找出结果的真正成因,必要时重排步骤或拆开耦合环节;用户没给的数量是未知而非已知——要么标成假设并说明若不成立则换哪条路,要么删掉;把问题压缩成一句话——要决定什么、什么让它难。
  17. 4. **类比与重组**:把这句话还原为某个经典问题的变体,但先检验映射是否成立——说不出"什么导致什么、经典方法前提还成立哪些"就不写类比、整段跳过;成立则借其成熟方法,再放下类比,用第一性原理按当前约束重新组合。
  18. 5. **验证**:用证据或反例检验结论;把无法验证的前提写成条件,说明什么情况下方案不可行;错误代价高的决定,额外给一个能推翻结论的测试和一个提前停手的观察点。
  19. 6. **清晰作答**:先给结论,再给关键原因、行动建议和验证方式。
  20. ## 安装
  21. 把整个 `tashan/` 目录放入所用 Agent 的技能目录。以 CodeArts Agent 的项目级技能目录为例:
  22. ```
  23. <project>/.codeartsdoer/skills/tashan/
  24. ```
  25. 其他 Agent 的目录约定不同,核心要求只有一条:`SKILL.md` 所在目录名为 `tashan`,且 frontmatter 中的 `name` 与目录名一致。
  26. ## 调用
  27. `SKILL.md` 的 `description` 已声明适用场景,Agent 会在条件匹配时自动加载。也可以显式点名:
  28. > 用 tashan 分析这个问题:……
  29. ## 示例
  30. 输入:
  31. > 线上接口白天正常,高峰期响应从 200ms 涨到 4s,日志里没有明显报错。请定位原因。
  32. 使用 Tashan 后,典型输出会先给出结论和判断依据,再列出需要业务确认的取舍,最后给出可执行的改动与验证方式;而不是直接猜测某个中间件出了问题。
  33. ## 状态
  34. 这是一个精简版(40 行、约 4,200 字符),已通过 Agent Skills 格式校验。
  35. 2026-11-21 跑过一次小样本 pilot(8 题 × 2 臂、两个顺序对调的盲裁判,过程与原始数据见 [evals/pilot/iteration-1/benchmark.md](evals/pilot/iteration-1/benchmark.md)):有 skill 的答案在 13 个有效判定中赢 11 个,五维总分 14.38 vs 13.69(满分 15),输出长度 1.05×。样本小、裁判同族、没有「等长通用提示」对照组,所以这只是方向信号,不足以宣称通用提升;简单题上它仍可能过度展开。
  36. 随后按这两个失败模式打了三处补丁(答完即停、未给的数量是未知、高价决定要有一条能证伪的测试),复测 3 题 × 2 臂:6 个判定全部判给新版本,13.33 vs 11.50,且不再出现虚构「吞吐没掉」或 simple 题长篇。过程见 [evals/pilot/iteration-2/RESULT.md](evals/pilot/iteration-2/RESULT.md)。
  37. 接着用一份等长的通用思考提示做对照(iteration-3,4 题,对照提示 3,678 字符 vs SKILL.md 4,193 字符):skill 6:2 略胜,五维 14.25 vs 13.38,但两道量化分析题裁判各判一边。也就是说,相对「无提示」的优势里有一部分只是「被要求认真思考」带来的;目前稳定的差异集中在诊断题的未知量标注与可证伪测试、以及简单题上不展开。见 [evals/pilot/iteration-3/RESULT.md](evals/pilot/iteration-3/RESULT.md)。
  38. 又做了一次消融(iteration-4):删掉第 4 步「经典问题 + 重组」后,4 题以 2:6 落后(13.25 vs 14.13),贡献集中在有真正同构经典问题的诊断/容量类题;但在纯算账题上这一步会多套一层框架、反而丢分。见 [evals/pilot/iteration-4/RESULT.md](evals/pilot/iteration-4/RESULT.md)。这仍不是通用提升的证据。
  39. 据此把第 4 步的退出条件写硬——先检验映射,说不出「什么导致什么、哪些前提还成立」就整段跳过、不谈类比——并在 h4(上一轮丢分题)与 h1(防止回退)上复测:新版 3 胜 1 平,14.75 vs 14.00,h4 不再硬套框架、h1 的排队机制与可证伪测试保留。见 [evals/pilot/iteration-5/RESULT.md](evals/pilot/iteration-5/RESULT.md)。
  40. iteration-6 换了考题:不再测「会不会分析」,而是测核心能力之一「拷问」——三题两轮对话,看它在给出方案前是否问出会改变结论的未知、并在拿到回答后改建议;对照组是一份等长、同样要求先确认未知的通用思考提示。结果 1 胜 2 负 3 平,六维均分 17.50 vs 17.67。**但这份对照是在知道 skill 哪些点有效之后反向写的,内容与 skill 高度重合**,所以它测的是包装而不是内容:同样包含未知量、可证伪、答完即停时,六步排版与散文排版没有可测差异(with_skill 在 q1 还长了 1.75 倍却没赢)。它不能说明这些内容无效。见 [evals/pilot/iteration-6/RESULT.md](evals/pilot/iteration-6/RESULT.md)。
  41. iteration-7 补上真正的对照:同样 3 题两轮,with_skill 对**无提示基线**(只给题面,不要求任何思考方式)。结果 6 个判定全部判给 skill,六维 17.50 vs 16.50,五维 14.67 vs 14.00;差距集中在两点——第一轮就把会改变结论的未知摆到方案前面,以及内部算术/口径少自相矛盾。基线本身并不弱(会质疑前提、会给公式),所以这不是「skill 会分析、基线不会」,而是「skill 先问再答、且更少前后矛盾」。注意 rubric 是按 skill 的原则写的,这个 6:0 是有偏的上界。见 [evals/pilot/iteration-7/RESULT.md](evals/pilot/iteration-7/RESULT.md)。
  42. iteration-8 换到豆包、换四道题(支付回调尾延迟、六个月微服务化、客服首响 30 分钟、数据库账单三倍),A 臂给技能正文、B 臂只给题面,两个盲裁判的甲乙映射相反。结果 8 个判定全给 A,五维 14.875 vs 11.875,裁判一致 4/4,长度 1.25×。差距集中在验证/判据与未知量:A 给阈值、反证分支和停手线;B 的诊断同样准确(线程栈、拆账单、驳斥排名),但收尾停在「基本能定案」。同族裁判 + 按 skill 原则写的 rubric,这个分差仍是上界;它支持「收尾更完整」,不支持「分析更强」。见 [evals/pilot/iteration-8/RESULT.md](evals/pilot/iteration-8/RESULT.md)。
  43. 欢迎通过 Issue 反馈使用效果与触发情况。
  44. ## License
  45. MIT,见 [LICENSE](LICENSE)。

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.