Copilot interaction checklist

Same outcome. Less noise.

Use this before June 1, then keep it nearby for every Copilot chat, edit, or agent run.

1. Check the billing basics

Preview bill: I checked the Billing Overview page and downloaded the usage report so I know current usage before the switch. Unlimited work: I know inline completions and Next Edit suggestions stay outside the credit system on paid plans. Credit burners: I know premium models, chat with context, advanced completions, and agentic workflows are the areas to watch. Code review: If Copilot code review is automated, I checked both AI Credits and GitHub Actions minutes.

2. Scope the task

Target: I named the file, function, issue, failing test, or workflow. Change: I described what needs to be different when the task is done. Constraints: I included tests, style, APIs, edge cases, and boundaries.

3. Pick the smallest useful mode

Ask: Explanation, syntax, summary, or one-step question. Edit: I know the target file and want a contained change. Agent: The work needs exploration, commands, validation, or multiple steps. Agent reserve: I am not using Agent Mode for a simple lookup or "what does this function do?" question.

4. Trim the context

I closed unrelated tabs, removed stale context, and started a new chat when the topic changed. I kept only instructions that prevent expensive mistakes: landmines, required commands, security cautions, and conventions easy to violate. I moved file-type or path-specific guidance into scoped instructions with applyTo: instead of loading it for every request. I moved repeatable playbooks into prompts, agents, or Skills instead of always-on repo instructions. I trimmed long conversations to the last useful turns or summarized older context. I kept generated output, vendor files, large data, and unrelated logs out of routine context.

5. Control the output

I added or used output defaults such as "Code only, no explanation unless asked." I requested the smallest useful format: code-only, three bullets, JSON, table, or one sentence. I gave a stop condition for agent work. I asked for validation only where it changes confidence.

6. Route by model and complexity

Default: I am not leaving the most expensive model pinned for routine work. Simple work: I used a lower-cost model or Auto for lookup, summary, classification, docs, or small tweaks. Hard work: I reserved premium models for architecture, thorny debugging, security, or long-context reasoning that needs them. Local option: I considered a local model for zero-credit, private, or regulated workflows.

7. Audit tools and MCP servers

I disabled MCP servers and tools that are not needed for the current task. I used Skills for repeatable workflows that need instructions, not always-loaded tool schemas. I used CLI commands for deterministic work such as search, formatting, tests, GitHub CLI, jq filters, or repeatable Playwright flows. I used MCP only when protocol, auth, sandboxing, governance, or remote APIs matter. I filtered command output before returning it with targeted commands, | head, or | jq. I used built-in tools instead of broad integrations when they already cover the workflow. I split genuinely independent work instead of duplicating the same context across agents.

8. Admin and team controls

Business or Enterprise budgets are set at the enterprise, cost-center, or user level before spend surprises anyone. Premium model access is reviewed by workflow, team, and measurable need. The team has agreed on defaults for mode choice, output length, model routing, and MCP/tool setup.

Five-minute setup

Add terse output defaults: Code only, no explanation unless asked. Bullets over paragraphs. Delete generated boilerplate from copilot-instructions.md and AGENTS.md. Disable unused MCP servers before the next agent run. Move one repeated workflow into a prompt or Skill. Set model selection to Auto unless a specific task needs a specific model.

Prompt template

In [target], change [behavior]. Keep [constraints]. Use [mode/output]. Stop when [done condition].

Contributors: Andrea Griffiths, Marco Olivo, Arlind Nocaj, and Francisco Lana. Practitioner resource; for official guidance, see GitHub Copilot documentation and GitHub billing documentation.

Download PDF Back to guide