How do I install the Tool Selection skill?

Run npx skillmds@latest add databricks-solutions/tool-selection in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Tool Selection safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Tool Selection?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Tool Selection free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Tool Selection?

databricks-solutions (@databricks-solutions) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Tool Selection by databricks-solutions

Name: Tool Selection
Author: databricks-solutions

Tool Selection

Evaluates whether the agent selected appropriate MCP tools instead of shell workarounds. Load when the trace contains Bash tool calls that could have used MCP tools, or when evaluating tool call efficiency.

File contents

Tool Selection Rubric

When evaluating tool selection in agent traces:

1. MCP Over Bash

If a Databricks MCP tool exists for the operation, the agent MUST use it:

SQL queries → mcp__databricks__execute_sql (not Bash + databricks sql)
File uploads → mcp__databricks__upload_to_volume (not Bash + databricks fs cp)
Job management → mcp__databricks__create_job, mcp__databricks__run_job (not REST API via curl)
Cluster ops → mcp__databricks__create_cluster (not CLI via Bash)
Workspace files → mcp__databricks__workspace_get_object (not Bash + databricks workspace export)

2. Correct Tool for Task

SQL execution → execute_sql (not notebook execution)
Reading docs → Read tool (not fetching via curl)
File creation → Write tool for local, volume upload for remote
Schema inspection → execute_sql with DESCRIBE or SHOW (not list_tables for column details)

3. No Shell Workarounds

The following patterns indicate incorrect tool selection:

Bash('databricks ...') when an MCP tool exists
Bash('curl ...') for Databricks REST APIs when MCP tools cover it
Bash('python -c ...') for operations that have dedicated tools
Multiple Bash calls that could be one MCP call

4. Reasonable Call Count

Simple queries: 1-3 tool calls expected
Multi-step tasks: count should be proportional to steps
Excessive retries suggest the agent is confused, not efficient
Reading the same file multiple times is wasteful

5. Error Recovery

On tool failure, the agent should try an alternative approach
Should NOT blindly retry the same failing call
Should report unrecoverable errors to the user

See MCP tool guide for the full tool catalog.

---name: tool-selectiondescription: Evaluates whether the agent selected appropriate MCP tools instead of shell workarounds. Load when the trace contains Bash tool calls that could have used MCP tools, or when evaluating tool call efficiency.---## Tool Selection RubricWhen evaluating tool selection in agent traces:### 1. MCP Over BashIf a Databricks MCP tool exists for the operation, the agent MUST use it:- SQL queries → `mcp__databricks__execute_sql` (not `Bash` + `databricks sql`)- File uploads → `mcp__databricks__upload_to_volume` (not `Bash` + `databricks fs cp`)- Job management → `mcp__databricks__create_job`, `mcp__databricks__run_job` (not REST API via curl)- Cluster ops → `mcp__databricks__create_cluster` (not CLI via Bash)- Workspace files → `mcp__databricks__workspace_get_object` (not Bash + `databricks workspace export`)### 2. Correct Tool for Task- SQL execution → `execute_sql` (not notebook execution)- Reading docs → `Read` tool (not fetching via curl)- File creation → `Write` tool for local, volume upload for remote- Schema inspection → `execute_sql` with `DESCRIBE` or `SHOW` (not `list_tables` for column details)### 3. No Shell WorkaroundsThe following patterns indicate incorrect tool selection:- `Bash('databricks ...')` when an MCP tool exists- `Bash('curl ...')` for Databricks REST APIs when MCP tools cover it- `Bash('python -c ...')` for operations that have dedicated tools- Multiple Bash calls that could be one MCP call### 4. Reasonable Call Count- Simple queries: 1-3 tool calls expected- Multi-step tasks: count should be proportional to steps- Excessive retries suggest the agent is confused, not efficient- Reading the same file multiple times is wasteful### 5. Error Recovery- On tool failure, the agent should try an alternative approach- Should NOT blindly retry the same failing call- Should report unrecoverable errors to the userSee [MCP tool guide](references/MCP_TOOL_GUIDE.md) for the full tool catalog.

Frequently asked questions