Start manually before connecting MCP
Use MCP to remove a concrete recurring problem with external data, not as a universal solution. A one-off action that is faster manually does not justify integration complexity.
Before automation, check value, repeatability, and controllability.
| Condition | Question |
|---|---|
| Value | Does manual work consume time or create errors? |
| Repeatability | Does the process happen regularly? |
| Control | Can access be limited and the result checked? |
Validate the process first
- Perform the action manually several times.
- Separate stable data and steps from incidental details.
- Define a limited integration scope.
- Start with read-only MCP.
- Verify returned data by primary identifier or source.
- Record the decision in TASK_SPEC.md or the skill documentation.
- Prepare a manual fallback if MCP is unavailable.
Example: issue tracker
For recurring issue analysis, MCP may list issues, open one, and search its fields. This can reduce copying errors in descriptions, statuses, labels, and comments.
The initial mode should be read-only. Changing status, assignee, or comments is an external side effect and needs separate approval.
list_issues
get_issue
search_issues
Status changes and comments: keep disabled initially.When MCP is unnecessary
- One-off local stack trace.
- Small log or screenshot.
- One-time documentation lookup.
- A case where manual insertion is faster than setup.
- A source that cannot be trusted or verified.
Canonical lesson source · JavaRush
Open JavaRush material