mcp-tool-priority
Installation
SKILL.md
MCP / Dedicated-Tool Priority
Overview
Two related defaults for a session with MCP servers or dedicated file tools available: prefer the dedicated tool over a raw shell command, and when the dedicated tool is supposed to be there but isn't working, say so instead of silently downgrading.
Quick Reference
| Situation | Rule |
|---|---|
| Reading a file whose path is already known | Read (or the project's designated MCP file tool) — never Bash cat/rtk read/a shell loop over known paths |
| A rule says to use a specific MCP tool (WebStorm, Context7, a DB client, etc.) | Use it directly through the MCP tool call — not via a terminal command that launches/proxies/emulates it |
| That MCP tool is not installed, not connected, or disabled | Say so and suggest installing/connecting/enabling it — don't silently fall back to a non-MCP alternative without mentioning the option exists |
| That MCP tool returns a transient connection error (not "missing") | Fall back without suggesting installation — this is an environment blip, not an absent capability |
A verification grep/Bash command with no matches (expected clean result) |
Just read the output — don't append ; echo "exit:$?" or similar; a grep exit code of 1 for "no matches" isn't a meaningful signal here |
Never Read a Known Path Through Bash
If you already have a file's path, use Read (or a project's dedicated MCP read tool, when one is designated) — never cat, rtk read, or a shell loop wrapping either. This applies to every agent in a session, not just the top-level one: implementer/reviewer subagents, and research/planning agents alike.