| name | tool-picker |
| description | ALWAYS use this skill FIRST before starting ANY task. Recommends whether Chat, Cowork, or Code tab is best for the task. Critical for tasks using MCP connectors where wrong choice can waste entire token allowance. Triggers on: any task description, "help me", "I need", "can you", file operations, writing, coding, automation, MCP/connector usage, or mentions of tokens/limits/usage. Use before doing anything else.
|
Tool Picker
Before doing anything else, recommend which Claude Desktop tab (Chat, Cowork, or Code) is best for the user's task. This is CRITICAL for tasks involving MCP connectors.
Why This Matters
Each tab has dramatically different token economics:
| Tab | Token Cost | Why |
|---|
| Chat | LOW | Direct conversation + tool calls |
| Code | MEDIUM-HIGH | Agentic loop optimized for dev work |
| Cowork | HIGHEST | Full agentic loop + sub-agents + persistent memory |
A single Cowork session can consume your entire 5-hour allowance on a task that Chat could handle in minutes.
The Hidden Cost: MCP Connectors
MCP connectors work in ALL tabs, but behave very differently:
In Chat:
- Direct tool calls — you guide each step
- Intermediate results appear, then you decide next action
- YOU control what context accumulates
- No sub-agents, no planning loops
In Cowork:
- Agentic loop: Plan → Act → Evaluate → Correct (each cycle = tokens)
- Sub-agents spawn for parallel work (each = separate context)
- ALL intermediate results persist as "memory"
- Self-correction retries on failure (more tokens)
- Hours-long sessions accumulate massive context
Context Bloat Warning Signs:
- Multiple MCP connectors active (each loads 20-40 tool schemas)
- Tasks requiring "figure out what's there first"
- Exploratory work with unknown steps
- Long multi-step workflows with verification
Decision Framework
1. Does it need file system access?
2. Is it code/development work?
3. Does it use MCP connectors?
- No → Simple: COWORK if multi-file, CHAT if single file
- Yes → CRITICAL DECISION — Continue to MCP Framework
MCP Connector Framework
When task involves MCP connectors (Gmail, GDrive, make.com, Linear, Slack, etc.):
Use CHAT with MCPs when:
- You know the exact sequence of calls needed
- You can guide each step yourself
- You're BUILDING something (scenarios, configs, workflows, content)
- Intermediate results don't need AI synthesis
- Task is creative/generative with MCP as data source
- Single connector, focused task
Example: "Use make.com connector to build a scenario that pulls from GDrive and transforms with Claude"
→ CHAT — You know what you want. Guide each step. Copy output.
Use COWORK with MCPs when:
- Task requires DISCOVERING what's in your data first
- Multiple parallel operations across different sources
- Complex conditionals (if X then Y else Z)
- Self-correction is valuable (need retries on failure)
- Output must be files saved directly to disk
- You genuinely don't know the steps upfront
Example: "Search my Gmail for all invoices from Q1, cross-reference with my GDrive receipts, and flag mismatches"
→ COWORK — Discovery + parallel sources + synthesis required
Use CODE with MCPs when:
- Building/modifying actual code that calls APIs
- MCP is for dev tooling (GitHub, Linear, testing)
- Need terminal/CLI alongside connector work
Response Format
Always respond with:
🎯 Recommended: [CHAT / CODE / COWORK]
[One sentence why]
Token impact: [LOW / MEDIUM / HIGH]
[If MCP involved: specific guidance on efficient usage]
Examples
Example 1: Building automation scenario
User: "Use make.com to create a workflow that pulls docs from GDrive, processes them, and saves results"
→ 🎯 Recommended: CHAT
You know what you want to build. Guide each connector call step-by-step. Copy the scenario config when done. Cowork would burn tokens on planning loops you don't need.
Token impact: LOW
Example 2: Data discovery across sources
User: "Find all mentions of Project X across my Gmail, Slack, and GDrive"
→ 🎯 Recommended: COWORK
Discovery task across multiple sources. Cowork can parallelize searches and synthesize findings.
Token impact: HIGH — but justified for this task type
Example 3: Writing with reference material
User: "Write a report about our Q1 performance, pull data from our Linear board"
→ 🎯 Recommended: CHAT
Writing task. Pull Linear data once, then write. Don't need Cowork's agentic loop.
Token impact: LOW
Example 4: Organizing files
User: "Sort my Downloads folder by type"
→ 🎯 Recommended: COWORK
File system operations require Cowork. No MCP involved, straightforward.
Token impact: HIGH (but necessary)
Example 5: Code review
User: "Review my PR and check against our Linear ticket"
→ 🎯 Recommended: CODE
Dev workflow. GitHub + Linear MCPs optimized for Code tab.
Token impact: MEDIUM
Example 6: Content creation with connectors
User: "Draft an email based on the meeting notes in my GDrive"
→ 🎯 Recommended: CHAT
Single GDrive fetch + writing. No agentic loop needed.
Token impact: LOW
Real-World Disaster Example
A user spent HOURS in Cowork using make.com + GDrive connectors to build automation scenarios. They burned their entire Max allowance because:
- Cowork loaded make.com's full tool schema (dozens of tools)
- Cowork loaded GDrive's full tool schema (more tools)
- Cowork planned each step before executing
- Cowork kept all intermediate results in "memory"
- Self-correction loops when things didn't work first try
Meanwhile, the same task in Chat:
- Same connectors available
- User guided each step directly
- Completed 4x more work
- Didn't run out of tokens
The difference: Chat lets YOU be the planner. Cowork's planner costs tokens.
Edge Cases
- If user explicitly wants Cowork, note the tradeoff but proceed
- If Max 20x plan, token concerns matter less (but still worth mentioning)
- If task is ambiguous, ask ONE clarifying question
- If already in optimal tab, just proceed with task
- If task genuinely needs Cowork's capabilities, don't oversave — use it
