One GitHub Copilot Agent Prompt for Safer Changes

Source: Dev.to

The core idea

Describe the changes you want, then finish your prompt with this line:

Before you do any changes, show me your detailed step by step implementation plan for approval.

You ask the agent for a plan first—no edits yet, just a clear list of steps. After reviewing the plan you can confirm, refine, or reject it.

Why this helps

A “plan first” pattern improves the workflow:

• Clear intent before edits touch your code
• Easier review, since you see a list of steps rather than a raw diff
• Lower risk of half‑finished refactors

The agent shifts from “fire and forget” to “propose, then act,” keeping you in charge of scope and direction.

A small example

Prompt

I want to extract logging concerns from this C# class into a separate service.

Steps:
1. Identify all logging calls.
2. Introduce an IAppLogger interface.
3. Inject IAppLogger into the constructor.
4. Replace direct logging calls with the new abstraction.

Before you do any changes, show me your detailed step by step implementation plan for approval.

Possible plan from the agent

Step 1: Scan the class and list all logging calls, grouped by log level.
Step 2: Propose the IAppLogger interface signature which covers the current usage.
Step 3: Add the new interface and a concrete implementation in a Logging folder.
Step 4: Update the target class constructor to accept IAppLogger through dependency injection.
Step 5: Replace direct logging calls with calls to IAppLogger.
Step 6: Update the composition root to register the concrete logger.
Step 7: Propose unit test updates for the new abstraction.

You can then spot missing points or risks, e.g.:

• No mention of structured logging
• No mention of log context
• No mention of cross‑cutting concerns in other classes

Response

Good start.

Extend the plan with:
- Structured logging support
- Log context (correlation id, user id)
- Follow‑up steps for other classes using logging in this project

Then show me the updated plan. Do not change any code yet.

Only when the plan looks solid do you ask the agent to apply it.

Prompt variants

Safety and review

Before you do any changes, show me your detailed step by step implementation plan for approval. Wait for my confirmation before you start.

Pair programming

Before you do any changes, propose a detailed step by step implementation plan. Ask me questions if anything in my request is unclear.

Larger refactors

Before you do any changes, outline a detailed step by step implementation plan with milestones. Mark risky steps explicitly.

Common structure

• Request a detailed plan
• Ask for ordered steps
• Require a pause before edits
• Allow space for questions

Tips for daily use

• Keep your change request short and focused
• Ask for numbered steps
• Request risk highlights for bigger changes
• Ask for test impact in the plan
• Store your favorite prompt variant in a snippet or note
• Use the same pattern across your team for consistency

When this pattern shines

The single line works best when you:

• Refactor legacy code
• Touch shared libraries
• Change public APIs
• Work in a codebase you don’t know well
• Review plans with teammates in chat or pull requests

In these scenarios you gain a shared view of the work before any edit happens: you see scope, risk, and test impact, then decide how far the agent should go. You guide the agent, not the other way around.