← ClaudeAtlas

capability-routinglisted

Use when a task needs a tool or workflow chosen for it - instead of a shell/Python/Node one-off, and for crawlers, parsers, browser automation, API wrappers, feature work, bug fixes, refactors, game/modding changes or releases - and when fetches are blocked, an approach failed twice, or more than one skill could own the next step.
SenjuWoo/Ultimate-AI-Starter-Bundle · ★ 7 · AI & Automation · score 78
Install: claude install-skill SenjuWoo/Ultimate-AI-Starter-Bundle
# Capability routing One question decides this, and it is not "which tool do I have?" ``` WHAT IS THE HARD PART OF THIS TASK? v WHICH AVAILABLE CAPABILITY DIRECTLY OWNS THAT HARD PART? v USE THE CHEAPEST TOOL THAT FULLY COVERS IT v IF IT FAILS BECAUSE THE CAPABILITY IS INSUFFICIENT: ESCALATE THE CLASS ``` The rule is **not** "always use a bundle tool". It is: > When several tools can do the job, take the cheapest reliable one whose > strengths match the hard part. Do not rebuild a weaker version of something > already installed just because shell, Python or Node is familiar. Both halves are load-bearing. A pile of excellent servers the model never reaches for is worse than a small set it routes correctly — and so is a browser stack launched to read a static JSON file. ## Default order, not a law 1. The application's native API, SDK, scripting interface or official CLI. 2. A verified specialized CLI harness, including a CLI-Anything harness when one covers the operation (`cli-hub can "<task>"`; see `tool-discovery`). 3. A specialized MCP for structured access, persistent app state or a live editor. 4. GUI/browser/computer automation when nothing reliable is programmatic. Pick the tool with the best fit, coverage, verification access and overhead — not the first item. A CLI wins for bulk work, an MCP for live state, and one workflow may need both. Never replace a working specialized capability with a fresh ge