Skip to content

Practical Setup

← Back to Guide

Use this checklist to apply the guide. Follow the linked chapters for explanations, evidence, and tool installation.

Start with the concise response default, then apply only the changes relevant to your workload. Step 4 applies to Coding Agent tasks; skip it for other workflows. Use Step 5 to review results rather than assuming savings.

Step 1: Set a concise response default

Create .github/copilot-instructions.md in your repository:

mkdir -p .github
touch .github/copilot-instructions.md

Start with ordinary concise language:

Lead with the answer. Keep explanations proportional to the task.
Skip greetings, repeated summaries, and unsolicited follow-up offers.
Preserve technical terms, negation, conditions, numbers, units, code, and commands.
Use full sentences when brevity would make an action ambiguous.
For code changes, report what changed and the relevant test result or blocker.

This is a starting point, not a measured savings guarantee. For output formats and optional Caveman style, see Output Control.

Step 2: Add only project-specific guidance

Keep rules the agent needs but cannot reliably discover in the code. Put path-specific rules in scoped instruction files and occasional checklists in on-demand guidance.

See Context Management for placement and Always-On Context Problem for what to keep.

Step 3: Choose mode, model, and context

Before a long run:

  1. Use the mode decision guide to decide whether tools are needed.
  2. Choose a model and supported effort level.
  3. Close unrelated files and attach only relevant context. Convert rich documents using the Markdown workflow.
  4. Inspect loaded tools; disable unused servers and extension-provided tools.
  5. Follow the cache and handoff guidance when session settings must change.

For setup, choose only what the workload needs: RTK or snip for noisy shell results; Graphify for repeated structural navigation.

Step 4: Make Coding Agent tasks reproducible

Prepare the environment

Use copilot-setup-steps.yml to pre-install dependencies deterministically. Keep the setup specific to the repo rather than asking the agent to rediscover it.

Write a precise issue

Name the failure, location, expected behavior, and test:

Bug: login fails when email contains '+' character.
File: src/auth/login.ts, validateEmail() on L42.
Fix: URL-encode the email before passing to the OAuth provider.
Test: add case for "user+tag@example.com" in login.test.ts.

Treat this as a worked example, not a diagnosis to apply without inspecting the code.

Use focused agent profiles when useful

A repeat workflow can use a narrow tool list and task-specific instructions. Do not duplicate the whole project's guidance in each profile. The repo's agents/token-saver.agent.md is an existing terse-style example, not a benchmarked cost guarantee or the required default.

Plan and verify

For multi-step work, use the plan/execute/verify workflow. Define acceptance checks, know the starting test baseline, and keep the final report tied to evidence.

Step 5: Review usage and adjust

Signal Action
Repeated clarification or missing conditions Use clearer wording and inspect task scope
Large loaded-context baseline Review always-on instructions and active tools
Large shell results Try one output filter; check that errors remain useful
Expensive sessions with weak results Review model choice, retries, plan, and acceptance checks
A style rule seems helpful Use the three-arm protocol, not an assumed savings rate

For history-based review, see Copilot CLI /chronicle and VS Code AI Engineering Coach.

Organization rollout

Set budgets, per-user limits, and model policy outside prompt files. Review spend alongside delivered work before expanding access.


Next: Find advice for remaining waste or review the evidence. For model rates and routing, see Model Selection & Pricing.