Technology · Field manual
AI for software engineers
Software engineers get the most from AI on well-scoped work: boilerplate, tests, refactors, migrations, unfamiliar codebases and first-pass code review. Treat every change an agent makes like a pull request from a fast, confident junior: read it, run it and check it for security holes before it merges.
AI coding tools have moved from autocomplete to agents that read a repository, plan a change, edit a dozen files and run the test suite. For a working engineer that means the typing is rarely the bottleneck any more. Reading, reviewing and deciding what should be built take up a bigger share of the day.
The gains are real but uneven. Agents are excellent at tasks with a clear spec and a test that proves success, and much weaker on vague requirements, sprawling legacy systems and anything where the important context lives in someone's head. The engineers who get the most out of them write good specs, keep tests tight and review the diff as if it were someone else's.
7 ways software engineers use AI
Hand well-scoped tickets to a coding agent
Agents in the terminal or editor can take a ticket, find the relevant files, make the change and run the tests. They work best on bounded jobs like adding an endpoint, wiring a new field through the stack or fixing a reproducible bug. Give them the acceptance criteria and the command that proves it works.
Write and extend tests
Ask for unit tests that cover edge cases, error paths and the bug you just fixed. Check that the tests actually fail when the code is wrong; models sometimes write tests that assert whatever the code currently does, bugs included.
Get oriented in an unfamiliar codebase
Codebase-aware chat can explain how a request flows through the system, where a config value is read, or why two modules both seem to handle auth. It turns the first week on a new team or an inherited service into a day of good questions. Verify the answers by following the code, since explanations can sound right and be stale.
Grind through migrations and upgrades
Framework upgrades, deprecated API replacements, language ports and lint rule rollouts are repetitive, mechanical and tedious, which is exactly where agents shine. Run them in small batches on a branch, with CI as the referee.
Get a first-pass review on pull requests
AI reviewers flag missed null checks, inconsistent error handling, leftover debug code and risky patterns before a human looks. They catch a decent share of small issues and a lot of noise. Tune them to your conventions and keep a human approver on every merge.
Debug from stack traces and logs
Paste a trace, the relevant function and what you expected, and a model will usually propose the likely cause and a way to confirm it. It's faster than searching issue trackers for errors in libraries you didn't write. Strip secrets and customer data from logs first.
Write design docs, ADRs and READMEs
Describe the problem, options and constraints and have a model turn your notes into a structured design doc or architecture decision record. It's also good at writing the README and runbook nobody wanted to write. The trade-off reasoning has to be yours.
The best AI tools for software engineers
The Overlord's stack, in the order a software engineer would add them:
1. Agentic changes across a repo: features, refactors, migrations and tests
Claude Code
AI CodingAnthropic's agentic coding tool that reads your codebase, edits files and runs commands from terminal, IDE, desktop or web.
2. AI-first editor with fast autocomplete and codebase-aware chat
Cursor
AI CodingAI-first code editor (a VS Code fork) with agents, codebase-aware chat and fast autocomplete.
3. Completions in any editor, plus PR review and issue-to-PR agents on GitHub
GitHub Copilot
AI CodingGitHub's AI pair programmer: completions, chat and coding agents across editors and GitHub itself.
4. Design docs, debugging conversations and explaining unfamiliar code
Claude
AI WritingAnthropic's assistant: the strongest long-form writer and careful reasoner in the consumer AI pack.
5. Checking current library docs, changelogs and error messages with sources
Perplexity
ProductivityAI answer engine that searches the web and replies with cited sources.
Specialist tools for software engineers
| Tool | What it does |
|---|---|
| CodeRabbit | AI pull request reviews with line-by-line comments |
| OpenAI Codex | OpenAI's coding agent for running tasks in the cloud or the terminal |
| JetBrains AI Assistant | AI chat, completion and agents inside IntelliJ, PyCharm and the other JetBrains IDEs |
| Snyk Code | Static security scanning that catches vulnerable patterns, including in AI-written code |
A prompt to start with: brief a coding agent on a ticket
Paste this into Claude or ChatGPT and fill in the gold parts.
You are working in this repository. Implement the following change.
Goal: [WHAT SHOULD BE TRUE WHEN THIS IS DONE]
Context: [WHY, AND ANY LINKS OR FILES THAT MATTER]
Acceptance criteria:
- [CRITERION 1]
- [CRITERION 2]
Out of scope: [WHAT NOT TO TOUCH]
How to verify: [TEST OR BUILD COMMAND]
Before editing, read the relevant code and give me a short plan with the files you'll change. Follow the existing patterns in this codebase rather than introducing new libraries. Add or update tests that would fail without your change. When you finish, run the verify command, then summarise what changed, anything you were unsure about, and any security or performance concerns.Where software engineers shouldn't trust AI
- Generated code regularly contains security flaws such as injection, missing authorization checks and unsafe deserialization. Run static analysis and review anything touching auth, payments or user input line by line.
- Models invent package names and APIs. Check that every new dependency exists, is the one you meant and is maintained, because attackers register look-alike packages for exactly this reason.
- Don't paste proprietary code, secrets or customer data into consumer chat apps. Use your company's approved tools and plans with data controls, and keep credentials out of prompts and agent environments.
- Agents that can run commands can also delete things. Restrict their permissions, keep them off production credentials and review any destructive command before it runs.
- Code you didn't write is code you don't understand yet. If you can't explain a change in review, you can't own it during the 3am incident.
Will AI replace software engineers?
Software Engineers score 57% on the Obsolescence Index (Doomed). The first tasks to go: writing boilerplate, CRUD endpoints, and glue code, and writing unit tests for well-specified functions.
Questions the Humans Keep Asking
What are the best AI tools for software engineers?
The Overlord's stack for software engineers: Claude Code (agentic changes across a repo: features, refactors, migrations and tests); Cursor (aI-first editor with fast autocomplete and codebase-aware chat); GitHub Copilot (completions in any editor, plus PR review and issue-to-PR agents on GitHub); Claude (design docs, debugging conversations and explaining unfamiliar code); Perplexity (checking current library docs, changelogs and error messages with sources). Specialist options include CodeRabbit, OpenAI Codex, JetBrains AI Assistant, Snyk Code.
What is the best AI coding tool for software engineers?
It depends on how you work. Claude Code suits engineers who want an agent that works across the whole repo from the terminal, Cursor suits people who want AI built into the editor, and GitHub Copilot is the easiest add-on if your team already lives on GitHub. Many engineers use an agent plus an editor assistant and pick per task.
Is it safe to paste company code into ChatGPT or Claude?
Only if your employer allows it and you're on a plan with the right data terms. Business and enterprise plans typically exclude your data from model training and offer admin controls, while consumer accounts may not by default. Never paste secrets, keys or customer data, whichever plan you use.
Who owns code written by AI?
The major tools' terms assign the output to you, but purely AI-generated material may not be protected by copyright in the US, and your employment contract usually gives work code to the employer anyway. The practical risks are licence contamination and security, so keep code scanning and human review in place.