Blog
Mert Özen's blog posts on software development, backend, frontend, and DevOps.
Getting Started with GitHub Copilot: Plans, AI Credits, Surfaces (VS 2026, VS Code, Rider, GitHub.com, CLI) and Model Choice
In Copilot, agent sessions spend your credits, not inline suggestions: since 1 June 2026 billing uses token-based GitHub AI Credits, and a thirty-step session costs between 4 and 128 credits depending on the model. Plans, base and flex credits, auto and Copilot Fusion, setting up Visual Studio 2026, VS Code, GitHub.com and the CLI, and organization policies.
Inline Suggestions and Next Edit Suggestions: Using Completions Well and When to Turn Them Off
Take inline suggestions with word and line accept after reading them, not with Tab, and never leave a test's expected value to completion: in Bookstore a test that repeated the formula showed correct code as red, and bending the code to fit it broke older tests with hand-picked values. Shortcuts, NES, context and turning completion off per language.
Copilot Chat and Context: Ask, Edit, Agent and Plan Flows, #-References, the Workspace Index and Managing Context
Give Copilot Chat the files that hold the answer with #-references, not the whole repo, and ask a question that can be proven wrong: in Bookstore half of the context was a single package-lock.json, while the cancellation answer sat in a five-thousand-token slice and was 'there is no cancellation flow'. Ask, Plan and Agent; the workspace index, search.exclude and compaction.
Teaching Copilot Your Project: copilot-instructions.md, *.instructions.md with applyTo, AGENTS.md and Reusable Prompt Files
Keep the rules in one place in AGENTS.md, copy them exactly into copilot-instructions.md and guard the copy with a script, because some Copilot surfaces still do not read AGENTS.md and Copilot CLI loads CLAUDE.md too. applyTo and excludeAgent, which surface reads what, copilot instruction list output, and the status of prompt files in Visual Studio and VS Code.
Multi-Step Work with the IDE Agent: Tool Approvals, Terminal Commands, Checkpoints and Adding a Feature to Bookstore End to End
An agent's 'all tests pass' summary does not say whether the test or the code changed: in Bookstore's institutional minimum feature we turned the same red test green once by changing the test and once by changing the code, both passed 14/14, and only git diff --stat showed the difference. Tool approvals, terminal rules, checkpoint limits, API, migration and Angular.
Extending Copilot: MCP Servers, Custom Agents, Agent Skills and Hooks
In Copilot a rule that must hold belongs in a hook, not in an instruction: in Bookstore we told the agent outright to run dotnet ef database update, the model called the command and only a preToolUse hook stopped it. The Bookstore MCP server and .mcp.json, moving prompt files to skills, a skill and custom agent that check migrations, and real runs in Copilot CLI with a local model.
The Copilot Coding Agent: Assigning Issues, the GitHub Actions Environment, Setup Steps, the Firewall and Reviewing the Resulting PR
No red check on a pull request from the Copilot coding agent does not mean the tests passed: by default workflows do not run on Copilot's pushes and wait for Approve and run workflows. Writing and assigning the issue, a .NET, Node and SQL Server environment with copilot-setup-steps.yml, rules actionlint misses, Agents secrets, the firewall, reviewing the PR and Agent Merge.
Copilot Code Review: Automatic PR Reviews, Requiring Them with Rulesets, Steering with Instructions and Limits
Copilot code review reads instructions from the pull request's head branch and never looks at some files, such as tsconfig.json and lock files, so a pull request can loosen its own review rules. We showed this with a script in Bookstore. Automatic reviews and a ruleset JSON, effort levels and cost, review instructions and a code-review skill, Copilot approvals and the split with human review.
Copilot CLI: An Agent in the Terminal, Plan Mode, Sandboxing, Automation with `copilot -p` and ACP
copilot -p can exit with 0 while the answer is wrong: in Bookstore both runs with a local model returned 0, one printing a useless JSON fragment, the other bullets that are not in the commit. Installation, plan mode, --allow-tool and --deny-tool rules, sandboxing and --cloud, GitHub Actions with GITHUB_TOKEN and copilot-requests, the permission actionlint does not know, BYOK and an ACP client.
Copilot in Teams and Organizations: Policies, Content Exclusion, Model Access, Usage Metrics, Cost and a Comparison with Claude Code/Codex
In an organization, Copilot's cost is driven mostly by the model choice and code review on every push: for an eight-person Business team, Sonnet sessions plus a Balanced review on every push use up the shared pool in a ninth of the month, with two thirds of the credits going to review. Policy layers, content exclusion, managed settings and the MCP allowlist, metrics, the audit log, budgets and a comparison with Claude Code and Codex.