메뉴
Score a coding task by complexity (1-10) and recommend the right model (haiku/sonnet/opus) before invoking a language agent. Holds the per-language rubric for all language/framework experts.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
CI pipeline design, parallel execution, caching strategies, and quality gates. Use when setting up CI/CD pipelines, configuring GitHub Actions, automating deployments, optimizing build times, or implementing quality gates. Triggers on 'CI/CD pipeline', 'GitHub Actions', 'deploy', 'build pipeline', 'continuous integration', 'automated deployment'.
Add comprehensive documentation including JSDoc/TSDoc, inline comments, README updates, and OpenAPI 3.0 API docs. Use when documenting code, generating docs, improving project documentation, adding docstrings, writing README files, or creating API specifications. Triggers on 'add docs', 'document this', 'write README', 'API spec', 'add docstrings', 'generate documentation'.
Greet all installed tools and generate a health report
Switch between model modes or check current mode. Use /model-mode status to see current setting, /model-mode opus-only to always use Opus, /model-mode default for auto-switching.
Generate project scaffolding, boilerplate code, and project structure for React, Next.js, Express, NestJS, FastAPI, Django, and more. Use when starting a new project, generating boilerplate, or setting up project structure. Triggers on 'scaffold', 'create project', 'new app', 'bootstrap', 'boilerplate', 'starter template', 'project structure', 'monorepo'.
Manage the RAG MCP server — index codebases, search semantically, configure backends (ChromaDB/Redis/Qdrant)
| name | route-language-task |
| skill_name | Route Language Task |
| description | Score a coding task by complexity (1-10) and recommend the right model (haiku/sonnet/opus) before invoking a language agent. Holds the per-language rubric for all language/framework experts. |
| argument-hint | <language> <task description> | hello | hello ID |
| user-invocable | true |
| version | 1.0.0 |
| author | Michel Abboud |
| license | Apache-2.0 |
| repository | https://github.com/michelabboud/claude-code-helper |
| issues | https://github.com/michelabboud/claude-code-helper/issues |
Single source of truth for complexity-aware model selection across all language/framework agents in this repo. Implements Pattern B from the routing design — the dispatcher consults this skill before invoking a language agent so the right model is locked in at invocation time.
This skill is also the canonical reference for the rubrics embedded inline in each language agent (Pattern A) and self-referenced by each agent's Complexity Self-Assessment Protocol (Pattern C).
/route-language-task <language> <task description>
/route-language-task rust "rewrite the storage layer to use Pin-projected self-referential streams"
/route-language-task python "add a /healthz endpoint to the FastAPI app"
The skill returns:
Score: 8/10
Model: opus
Drivers: base 1 +2 unsafe (Pin projection) +2 self-referential lifetime puzzle +2 async stream internals (custom Stream/Pin) +1 cross-crate refactor = 8
Recommendation: Invoke rust-expert on model opus.
(Each driver is a distinct rubric concept counted once; cross-crate refactor is +1 per the rubric; the numbers sum to the score.)
| Score | Model | Rationale |
|---|---|---|
| 1–3 | haiku | Boilerplate, formatting, dependency bumps, obvious fixes |
| 4–6 | sonnet | Default — most feature work, refactoring, normal debugging |
| 7–10 | opus | Type-system depth, soundness/concurrency analysis, perf |
Base score: 1. Add modifiers from the rubric. Cap at 10.
As a dispatcher (main Claude or Agent caller):
Pin appears under both the lifetime line and the async line), score it once, at the higher value. Default to base 1 if nothing matches.<lang>-expert agent on the chosen model. Where the sub-agent tool accepts a per-call model (this harness's Task/Agent tool does — pass model: "<chosen>" alongside subagent_type: "<lang>-expert"), the model is locked at invocation. If your harness sets a sub-agent's model only via its frontmatter, surface the recommended model to the user instead of forcing a lower tier.As an agent (Self-Assessment, Pattern C): At the start of every task, run the rubric internally. If your score exceeds the band of the model you're running on, stop and respond:
"Complexity score: X/10 (drivers: ...). I'm on {current_model} but this task scores in the {recommended_model} band. Recommend re-invoking with
model: {recommended_model}. Proceeding now would risk: ..."
Do not proceed without explicit user override. This is a circuit breaker, not a router — the model is locked at invocation, the agent can only halt and request escalation.
rust-expert)transmute, MaybeUninitPin projectionFuture/Stream/Waker, Pin/Unpin, manual poll_*macro_rules!, build.rs codegengo-expert)sync.Cond, lock-free with sync/atomic, race conditions)cgo, unsafe, runtime/scheduler internalsgo generate pipelinespython-backend-expert)__init_subclass__ / class-creation magicProtocol, ParamSpec, TypeVar variance, genericsctypes, cffi, cython)nodejs-typescript-backend-expert)java-spring-boot-expert)Mono/Flux) composition, backpressure, hot/cold flows@Transactional propagation/isolation edge casesruby-rails-expert)define_method, method_missing, prepend ancestry tricksThread, Fiber, Ractorphp-expert)laravel-expert)react-nextjs-expert)as props, ref forwarding)useDeferredValue, useEffectEventvue-nuxt-expert)ref vs reactive, deep tracking, shallowRef traps<script setup>, generic componentssvelte-expert)handle, handleFetch, handleError$bindable / $props varianceangular-expert)ControlValueAccessorandroid-expert, android-dev)StateFlow/SharedFlow buffering, cancellation propagation)SubcomposeLayout, custom modifiers, layout phases)ios-development-expert)@MainActor traps@objc runtimeflutter-react-native-expert)runOnJS boundarieswordpress-expert)+2 modifier contributed, escalate to the next band (3 → sonnet, 6 → opus). Scores are integers, so this is a band-ceiling escalation, not rounding.When invoked with <language> <task>:
<language> (case-insensitive) to a section above. Aliases:
ts, typescript, node, nodejs → Node.js + TypeScriptjs, javascript → Node.js + TypeScript (or React/Vue/Svelte/Angular if framework named)py → Python backendkotlin, android → Androidswift → iOSdart, flutter, rn, react-native → Flutter / React Nativescore = 1 + sum(modifiers), capped at 10.Language: <name>
Score: <n>/10
Drivers: <list of modifiers, with their +N value>
Model: <haiku|sonnet|opus>
Agent: <lang>-expert
Invocation: Agent({subagent_type: "<lang>-expert", model: "<model>", prompt: "..."})
When invoked with hello:
🧭 Hello! I'm Route Language Task v1.0.0. I score coding tasks 1-10 and pick the right model. Use
/route-language-task hello IDfor the full guide.
When invoked with hello ID:
/route-language-task <language> <task description>description: field.Author: Michel Abboud License: Apache-2.0 Repository: claude-code-helper Issues & Discussions: GitHub Issues