Blog
Mert Özen's blog posts on software development, backend, frontend, and DevOps.
How Much Do AI Coding Assistants Really Help? Tool Types, Evidence and the Right Expectations
AI coding tools do not help equally with every task: they speed up isolated, easy-to-verify work, while on work that needs repository knowledge the gain depends on context and review. We cover the tool types from completion to cloud agents, METR's 2025 and 2026 measurements, DORA 2025's amplifier finding and two real experiments with a local model.
Giving Agents the Right Context: AGENTS.md, copilot-instructions.md, CLAUDE.md and Path-Specific Instructions
The tool-neutral way to describe a repository to a coding agent is a short AGENTS.md at the root: build and test commands, folder layout, team rules and topics to ask about first. We cover which tool reads which file, nested and path-specific instructions, a real AGENTS.md for Bookstore, the results of the same task with three different contexts, and a check that catches stale instruction files.
A Feature Workflow with AI: Explore → Plan → Implement → Verify and Spec-Driven Development
The safe way to build a feature with an agent is to split the work into explore, plan, implement and verify, and keep the plan written down and approved, so you read every diff against it. We add a wishlist to Bookstore with real model output, review the plan, go test-first, catch code that deviates from the plan, and cover Spec Kit's flow, context management and parallel agents with git worktree.
Reviewing and Verifying AI-Written Code: Typical Failure Modes, a Checklist and AI Reviewers
The most dangerous AI-written code is code that compiles and passes the tests but is wrong, so read the diff against the plan, read test changes separately and check edge cases with your own tests. With real model output we cover a move-to-cart endpoint that returns 500, a test weakened to turn CI green, a checklist for AI pull requests, how Copilot code review works and the findings of a local AI reviewer.
Using AI Coding Tools Safely: Secrets, Prompt Injection, Package Hallucination and Licensing
The safe way to use an AI coding agent is to assume the model will be fooled and limit its permissions: command approval, a sandbox, network off and no production secrets where the agent runs. With real output we show a hidden issue comment entering the agent's plan in ten out of ten attempts, what gitleaks misses, and five of ten model-suggested packages not existing on nuget.org; we also cover dependency scanning and licensing risk.
AI in the Team: Policy, Measuring Impact, Junior Growth and When Not to Use It
Getting value from AI in a team depends more on the environment than on the tool: a short written policy, rules backed by technical controls, measurement with delivery metrics and habits that protect learning. We cover a real AI policy and pull request template for Bookstore, DORA's seven capabilities, misleading metrics, Anthropic's skill-loss experiment and the work where not using AI is better.