Skip to main content

​
What good prompts do

Good MCP prompts are specific about the outcome, the data shape, and whether the client is allowed to write.
  • ask for one task at a time
  • name the target object or workflow
  • say whether you want read-only analysis or a write-capable plan
  • request a confirmation checkpoint before any mutation
  • prefer concise summaries first, then deeper detail on request
If you are unsure which scope or tool family you need, start with mcp_whoami, then ask the client to explain the granted scopes before you do anything else.

​
Read-only examples

These are safe starting points for exploration, triage, and workspace inspection.

​
Write-capable examples

For anything that mutates data, instruct the client to stop before execution and ask for approval.
For write workflows, the best prompt is not “do it all automatically.” It is “prepare the change, show me the diff, and wait.”

​
Prompts by client

​
Claude Code

Use this when you want the assistant to inspect data first and only mutate after you approve the plan:

​
Codex

Codex works well when you separate discovery from execution:

​
Cursor

Cursor prompts should be explicit about the local flow you want:

​
ChatGPT developer mode

Use the same operator pattern in ChatGPT developer mode:

​
Other remote MCP clients

For any other remote MCP client, keep the instruction simple:

​
Prompt patterns that work well

  • First inspect, then summarize, then ask
  • Show the minimum diff
  • Use read-only tools unless I approve a write
  • Explain missing scopes instead of retrying blindly
  • If something fails, tell me the exact MCP/client reason and the next fix

​
Prompt patterns to avoid

  • Do everything automatically
  • Fix the workspace without asking
  • Use whatever scopes you need
  • Keep retrying until it works
  • Write changes and then summarize them

​
Next steps