Claude技能+插件为每一项任务选择正确的模式和努力: Opus计划, Haiku/Sonnet 构建,每个步骤都经过验证,只有在失败时才会升级.
概览
大部分的作品不需要最大的克洛德模型 在最大努力。 这种技巧使克洛德挑选出最便宜的模型和能完成某项任务的努力. 每一个部分都经过核实,克洛德只有在检查失败时才会转向更强大的模型. Claude写了一个很严谨的规格,发给了Haiku或Sonnet,自己运行检查,只有在检查失败时才能向上移动一等. 1. Haiku曾将密码库映射为.router/CONTEXT.md 2. Opus写作.router/PLAN.md 有接受标准(API形状,schema,签名)和小任务. 每项任务都有分级和完成的检查。 你批准计划之前,任何事情建成。 3. 每项任务都由最便宜的配发商负责,只满足其需要。 克劳德核实每一项任务 并把它当作一个检查站 4. 如果Opus任务两次失败,则该部分的计划会得到固定。 5. Opus根据接受标准审查diff,其结论又成为了新的任务。 在.router/PROGRESS中节省了进度.
概述（AI 翻译）
** 强大的模型认为,廉价模型进行打字,真正的检查决定。 **
大部分的作品不需要最大的克洛德模型 在最大努力。 这种技巧使得克洛德选择了最便宜的模型和努力,可以完成一个任务的每一部分**. 每一个部分都经过核实,克洛德只有在检查失败时才会转向更强大的模型.
如何运作 成本 |. 最大一切 最大模型 最大努力 每一项任务 手动转换模型 Escalate最便宜的模型,先核实,再向上移动
它会做什么#
** 小任务(快速路径):** Claude写了一个很严谨的规格,发给了Haiku或Sonnet,自己运行检查,只有在检查失败时才能向上移动一等.
** 大项目(项目路径):**
- 检索: Haiku将密码库一回映入
.router/CONTEXT.md。 - 计划: Opus写作有验收标准、合同第一(API形状、计划、签名)和小任务。 每项任务都有分级和完成的检查。 你批准计划之前,任何事情建成。 3个.
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
Smart Task Router for Claude
The strong model thinks, cheap models do the typing, and real checks decide.
Most work doesn’t need the biggest Claude model on max effort. This skill makes Claude pick the cheapest model and effort that can do each part of a task. Every part is verified, and Claude moves to a stronger model only when a check fails.
| Approach | How it works | Cost |
|---|---|---|
| Max everything | Biggest model, max effort, every task | $$$$$ |
| Manual | You switch models by hand | $$ |
| Escalate | Cheapest model first, verify, then move up | $ |
| Router (this repo) | A plan picks model + effort per subtask, verifies, escalates on fail | ¢ |
What it does
Small tasks (quick path): Claude writes a tight spec, sends it to Haiku or Sonnet, runs the check itself, and moves one tier up only if the check fails.
Big projects (project path):
- Recon: Haiku maps the codebase once into
.router/CONTEXT.md. - Plan: Opus writes
.router/PLAN.mdwith acceptance criteria, contracts first (API shapes, schema, signatures) and small tasks. Each task has a tier and a done-check. You approve the plan before anything is built. - Build: each task goes to the cheapest fitting agent with only the context it needs. Claude verifies every task and commits it as a checkpoint.
- Re-plan, don’t brute-force: if a task fails twice on Opus, the plan for that part gets fixed instead.
- Review: Opus reviews the diff against the acceptance criteria, and its findings come back as new tasks.
Progress is saved in .router/PROGRESS.md, so in a new session just say “continue the project”.
| Tier | Agent | Model / effort | For |
|---|---|---|---|
| E | router-easy |
Haiku · low | Boilerplate, renames, conversions, copying a pattern |
| N | router-normal |
Sonnet · medium | Standard features inside defined contracts |
| H | router-hard |
Opus · high | Planning, core logic, hard bugs, security, review |
| X | (override) | Fable · high | Only after Opus fails twice |
Install
Claude Code (recommended: includes the agents with effort settings)
/plugin marketplace add celiktech24-cell/smart-task-router
/plugin install smart-task-router@smart-task-router
Claude Code, manual install
git clone https://github.com/celiktech24-cell/smart-task-router
cp -r smart-task-router/skills/smart-task-router ~/.claude/skills/
cp smart-task-router/agents/*.md ~/.claude/agents/
Claude app (claude.ai, desktop, Cowork)
Download smart-task-router-skill.zip from the Releases page and upload it in the Skills section of Claude’s settings. The skill runs there too, but only Claude Code supports the agent files, so effort control is Claude Code only.
Use it
Just ask for the task:
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add celiktech24-cell/smart-task-router