routing
Skill1 starsChanged yesterdayMITView on GitHub
Change or explain the Auto Router's settings and routing. Use when the user asks, in words, to switch budget mode (economical, balanced, quality), change which model a tier or kind of work uses, limit parallel agents, turn the dashboard or independent reviews on or off, restrict Fable to certain areas, or asks why a task was routed to a model.
17 other projects use a version of this.
Adopted by
Most recently taken up by
- microsoft/coreai-design-systemtheir own version, first seen Nov 15, 2026
- NVIDIA/nvcftheir own version, first seen Nov 15, 2026
- aazirani/flutter_agentifytheir own version, first seen Nov 15, 2026
- AnyOrgName/php-skillstheir own version, first seen Nov 8, 2026
- github/replicatheir own version, first seen Nov 1, 2026
- darknecrocities/SkillBuildertheir own version, first seen Nov 1, 2026
Used with
Projects with it also use parallelization, prompt-chaining and 1 more.
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.
hajekimSkillparallelization
This skill should be used when the user wants to "run tasks in parallel", "concurrent LLM calls", "fan-out and fan-in", "parallel agents", "simultaneous execution", "reduce latency with concurrency", "parallel sub-tasks", "parallel processing", "concurrent agent execution", "run agents simultaneously", "speed up with parallelism", "async agent execution", "batch LLM calls", "parallel workflow", "parallelize agent tasks", or needs to execute multiple independent tasks simultaneously. Also responds to Korean: "병렬 작업 실행", "동시 LLM 호출", "팬아웃 팬인 패턴", "병렬 에이전트", "동시 실행", "동시성으로 지연 시간 감소", "병렬로 실행", "동시에 여러 작업", "병렬 처리", "동시에 돌려줘", "빠르게 병렬로", "여러 작업 동시 실행해줘". Also responds to Japanese: "並列処理", "同時実行", "並列エージェント", "並行LLM呼び出し", "ファンアウト・ファンイン", "複数タスクを同時に実行したい", "処理速度を上げたい", "非同期エージェント", "バッチ処理", "並列で動かしたい" Also responds to Chinese: "并行处理", "并发执行", "并行智能体", "同时调用多个LLM", "扇出扇入模式", "多任务并行运行", "加快处理速度", "异步智能体", "批量LLM调用", "同时运行多个任务".. Apply this skill to design or implement the Parallelization agentic design pattern.
Also in 3 projects that use this
hajekimSkillprompt-chaining
This skill should be used when the user wants to "chain prompts", "build a pipeline", "break down complex tasks into steps", "sequential LLM calls", "multi-step reasoning", "pipeline pattern", "sequential agent pipeline", "multi-step prompt pipeline", "LLM chain", "step-by-step agent", "prompt pipeline", "decompose task into prompts", "stage-by-stage processing", "chained LLM calls", or when a task is too complex for a single prompt and needs to be decomposed into sequential sub-tasks. Also responds to Korean: "프롬프트 체이닝", "체인 방식 에이전트", "단계별 LLM 호출", "복잡한 작업 분해", "순차적 프롬프트 파이프라인", "단계적 작업 처리", "단계별 처리", "순차 처리 에이전트", "파이프라인 만들어줘", "단계적으로 처리해줘", "연속 프롬프트", "체이닝 해줘". Also responds to Japanese: "プロンプトチェーニング", "ステップごとに処理", "順次LLM呼び出し", "パイプライン構築", "複雑なタスクを分解", "段階的エージェント", "連続プロンプト", "LLMパイプライン", "順番に処理してほしい", "チェーン方式エージェント" Also responds to Chinese: "提示词链", "顺序调用LLM", "构建流水线", "分解复杂任务", "逐步处理", "链式提示", "多步骤流水线", "按顺序处理任务", "提示词管道", "帮我一步一步处理".. Apply this skill to design, implement, or explain the Prompt Chaining agentic design pattern.
Also in 3 projects that use this
NVIDIASkilltesting
Run Python and Rust tests for physicsnemo-curator using uv + pytest with coverage reporting via pytest-cov, and cargo-nextest for Rust tests. Includes benchmark workflows with pytest-benchmark and criterion.
Also in 3 projects that use this
Other versions17 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 17
- github/replica★ 275same name
- NVIDIA/nvcf★ 268same name
- ypollak2/llm-router★ 123same name
- ThomasLangbroek/frugal★ 37same name
- darknecrocities/SkillBuilder★ 13same name
- flpbalada/fb-skills★ 8same name
- AnyOrgName/php-skills★ 4same name
- hajekim/agentic-design-patterns-skills★ 4same name
- jkaninda/okapi-skills★ 3same name
- ffsshhttiikk/opencode-agents-skills★ 2same name
7 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.
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 3 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 16 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
Use the mod's own tool. It validates the change against the schema, backs up the previous file, replaces it in one step, and reloads.
- Very specific: line 17
If that answers "not understood", read the configuration with { "action": "get" } and set exact paths: { "action": "set", "changes": [{ …
- Very specific: line 32
If the tool is not available (the mod is not loaded in this session), say so.
- Distinctive: line 50
Keep simple questions, wording fixes, small edits and anything that leans on this conversation in the main session.
- Distinctive: line 54
Ask each agent for a short report.
- Distinctive: line 61
Never repeat a failing call unchanged.
“Common” is measured: kept by at least 6 in 10 of the 16 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Auto Router
The Auto Router is a mod that runs by itself. It classifies every request locally, picks the cheapest capable model and effort, routes subagents, and shows it all in the Auto Router pane. Nobody has to invoke this skill for routing to happen. Load it only to change a setting or explain a decision.
Changing a setting the user asked for in words
Use the mod's own tool. It validates the change against the schema, backs up the previous file, replaces it in one step, and reloads. Never edit config.json by hand while the tool is available.
- The tool is deferred. Load it with ToolSearch:
select:mcp__auto-router__configure. - For a request in the user's words, pass it through:
{ "action": "request", "request": "switch to economical mode" } - If that answers "not understood", read the configuration with
{ "action": "get" }and set exact paths:{ "action": "set", "changes": [{ "path": "budgetMode", "value": "economical" }] } - Report what the tool answered. If it refused the change, say why in its words. Do not retry with a different value the user did not ask for, and do not change any other setting.
| The user says | Path and value |
| Switch to economical mode | budgetMode: economical |
| Use Opus for complex frontend work | append to routing.overrides: { "domain": "frontend", "complexity": "complex", "model": "opus" } |
| Limit parallel agents to one | guardrails.maxParallelAgents: 1 (and guardrails.defaultParallelAgents: 1) |
| Disable the dashboard | sidePanel.enabled: false, sidePanel.autoOpen: false |
| Use Fable only for architecture and security | routing.headChef: { "domains": ["architecture", "security"], "exceptional": false } |
| Require independent reviews for high-risk changes | guardrails.independentReview: true, guardrails.reviewMode: remind |
A change to a guardrail, a risk floor or the protected sessions is always put to the user for confirmation. That is by design: do not look for a way round it.
If the tool is not available (the mod is not loaded in this session), say so. The user can run /auto-router config <request> themselves, or the file can be edited at ~/.claude/auto-router/config.json against config.schema.json in the same folder, after copying the old file into backups/.
Explaining a decision
/auto-router status prints the current task's route and the reason. The pane's Router card shows the same. A model line reads "assigned" for what the router asked for and "answered (API)" for what the API reported; only the second is evidence of what ran.
The kitchen
| Agent | Type | Default model | For |
| Head Chef | auto-router:head-chef | fable | Complex architecture, security, exceptional reasoning |
| Sous Chef | auto-router:sous-chef | opus | Advanced development, debugging, difficult research |
| Line Cook | auto-router:line-cook | sonnet | Standard development, testing, routine work |
| Prep Cook | auto-router:prep-cook | haiku | Simple searches, summaries, repetitive edits |
| Independent Reviewer | auto-router:independent-reviewer | opus | Fresh-context, read-only review |
Delegation rules
- Keep simple questions, wording fixes, small edits and anything that leans on this conversation in the main session. A subagent starts cold and re-reads what you already know.
- Delegate independent research, an isolated implementation, a complex investigation, or parts that can truly run apart.
- One agent at a time by default. Two at most; a third is refused by the mod.
- For substantial work: inspect, plan, implement, test, review. Not one agent per stage.
- Ask each agent for a short report. Give it the files and the definition of done, not the whole conversation.
- High-risk work gets an independent review before it is called done: give the reviewer the original requirements and the files changed.
Failure rules
- Two substantive failures (a failing test, a type error, a crash) move the task up one tier; the mod does this and says so beside the error.
- An environment failure (a missing program, a permission, a timeout, the network) is not a reasoning problem. Fix the environment or ask the user.
- Never repeat a failing call unchanged.
What the router never does
It never approves a tool call, never sets a permission mode, never changes settings.json, and never reports a model as having run without the API saying so.
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.

