Write unit tests: how to do it with AI
Last checked 2026-10-15
IntermediateBot with human review 💻 Coding
Generate tests for existing code and run them in an agent.
1. Bot automation level
Bot with human review The bot does the work, but facts, numbers or code should be checked.
2. Before you hand it to the bot
- Gather inputs: files, data samples, links and any templates.
- Access: make sure you can use the chosen bot/plan (see limits below).
- Context: tell the bot audience, purpose, constraints and examples of good output.
- Privacy check: remove personal, confidential or regulated data unless your plan/workspace is approved for it.
- Success criteria: write down what 'done' looks like (length, format, tests passing, fields filled).
3. Copy-paste prompts
Main prompt
Read <module>. First list the test cases you plan (happy path, boundaries, errors). Wait for my OK, then write <framework> tests and run them. Don't change source code.
Beginner-friendly
Read <module>. First list the test cases you plan (happy path, boundaries, errors). Wait for my OK, then write <framework> tests and run them. Don't change source code. I'm new to this: explain each step in plain words and ask me for anything you need before starting.
Check-your-work
Read <module>. First list the test cases you plan (happy path, boundaries, errors). Wait for my OK, then write <framework> tests and run them. Don't change source code. When you finish, list any assumptions you made, anything you're unsure about, and what I should double-check before using the result.
Free tiers that can handle this (verified): ChatGPT (incl. Work & Codex), GitHub Copilot, Mistral Le Chat. Others: check plan.
4. Step-by-step
- Choose the framework (pytest, Jest, JUnit) and tell the agent which one.
- In an agent (Claude Code, Cursor Agent, Codex, Copilot agent mode) ask it to read the module and list cases before writing tests.
- Approve the case list: happy path, boundaries, error paths, mocks for I/O.
- Let the agent write the tests and run them in the terminal.
- If tests fail, decide whether the test or the code is wrong — don't let the agent 'fix' code to match a wrong test.
- Check coverage locally and commit on a branch.
5. Which bots and plans can do it
| Tool | What's different here | Plan limits | Source |
|---|---|---|---|
| Claude (incl. Claude Code) | Claude Code runs commands in your terminal with permission prompts. | Claude Code and Research need Pro or higher; computer use is Pro/Max only (research preview). | code.claude.com/docs/en/desktop |
| ChatGPT (incl. Work & Codex) | Codex runs tests and commands locally or in cloud environments. | Free tier has limited deep research, Codex and Work; scheduled tasks need Go or higher (per chatgpt.com/pricing). | help.openai.com/en/articles/20001275-chatgpt-work-and-codex |
| Cursor | Agent can run terminal commands to execute the tests. | Pricing page unverified; cloud agents require a connected source-control repo. | cursor.com/docs/agent/overview |
| GitHub Copilot | Agent mode in your IDE edits files locally; the cloud agent works in GitHub Actions. | Free: 2,000 completions + 50 chats; cloud agent listed on paid plans. | docs.github.com/en/copilot/concepts/agents/cloud-agent/about-cloud-agent |
| Mistral Le Chat | — | Free tier has limited messages, searches and coding sessions; Pro raises limits. | mistral.ai/pricing |
| DeepSeek | — | Chat app limits not documented officially; API is pay-per-token. | api-docs.deepseek.com/quick_start/pricing |
Free or paid?
- Claude (incl. Claude Code): Free plan: yes. chat, web search, files & code execution, Artifacts, connectors. Claude Code and Research need Pro. Paid: the lowest price we could verify on the official pricing page is $20/month (other plans may exist). For this task, check the plan.
- ChatGPT (incl. Work & Codex): Free plan: yes. unlimited text chats; limited uploads, images, voice, deep research and Codex. Paid: price not verified, so check the official site. Its free tier can handle this task.
- Cursor: Free: not verified, so check the official site. Paid: price not verified, so check the official site. For this task, check the plan.
- GitHub Copilot: Free plan: yes. 2,000 completions and 50 chat requests a month; agent mode and CLI. Paid: the lowest price we could verify on the official pricing page is $10/month (other plans may exist). Its free tier can handle this task.
- Mistral Le Chat: Free plan: yes. limited messages, searches, coding sessions and basic images. Paid: the lowest price we could verify on the official pricing page is $14.99/month (other plans may exist). Its free tier can handle this task.
- DeepSeek: Free: not verified, so check the official site. Paid: price not verified, so check the official site. For this task, check the plan.
6. Risks and what to double-check
- Agents sometimes write tests that only assert the current (possibly buggy) behaviour.
- Review mocks: over-mocking makes tests pass while proving nothing.
- Run tests and review diffs; never merge unreviewed agent code.
- AI can hallucinate facts — verify anything that matters.
- Don't share secrets or other people's personal data.