AI Unit Tests for Legacy Code: A Safe Workflow
Why AI Helps with Legacy Test Coverage
Generating unit tests for legacy code with AI works best when you treat the model as a fast code-reading and test-drafting assistant, not as an automatic correctness engine. Older services often have weak boundaries, undocumented side effects, database calls mixed with business rules, and edge cases known only through production behavior. An AI coding model can quickly map those behaviors, suggest seams for dependency injection, and produce an initial suite that your team can review and refine.
The practical goal is not “test everything.” Start by locking down high-risk behavior before changing it. Good candidates include pricing calculations, permission checks, data transformations, retry logic, parsers, and bug-prone API adapters. A small set of trustworthy characterization tests can make a legacy refactor dramatically safer.
Step 1: Choose a Small, High-Value Target
Pick one module that is important, reasonably isolated, and currently hard to modify. Avoid beginning with a 2,000-line controller that talks directly to several databases and external APIs. Instead, select one method or class with a clear input and observable output. Gather its source file, related interfaces, existing tests, package manifest, and a few real examples from logs or bug reports.
Before asking AI to write tests, run the existing test command and record the baseline. Note the test framework, command, runtime version, and whether tests require environment variables or containers. This prevents a common failure mode: accepting generated tests that look convincing but do not run in the repository.
Step 2: Ask AI for a Behavior Map First
Do not lead with “write tests for this file.” First ask the model to explain the code and identify testable behavior. Provide the relevant source, dependency interfaces, and any known production issue. Ask it to list inputs, outputs, mutations, exceptions, external calls, hidden assumptions, and branches that deserve coverage.
A useful prompt is: “Analyze this legacy module without changing it. Create a behavior table with normal cases, boundary cases, error paths, side effects, and dependencies to mock. Identify any behavior that is ambiguous and must be confirmed by a developer.” This step exposes uncertainty early and gives you a reviewable test plan instead of a blind code dump.
Step 3: Generate Characterization Tests in Small Batches
Next, request two to five tests for one behavior group at a time. Tell the model to use the project’s existing framework and conventions, reuse existing fixtures where possible, and avoid changing production code unless a minimal test seam is necessary. Include the exact import path and a representative existing test file so the output matches local style.
For a legacy function that calls a payment gateway, begin with tests that preserve current observable behavior: valid input returns the current result, malformed input raises the current error, and a gateway failure produces the current fallback. Mock only boundaries such as network clients, clocks, random generators, filesystems, and databases. Over-mocking internal helper calls makes tests brittle and merely verifies implementation details.
- Keep each prompt narrow: one function or branch family is easier to validate.
- Request explicit assertions: assert returned values, thrown errors, and meaningful external calls.
- Ask for deterministic tests: freeze time, seed randomness, and avoid live network requests.
- Require an explanation: have AI state which behavior each test protects.
Step 4: Run, Diagnose, and Iterate
Paste the generated tests into a branch, run the smallest relevant test command, then run the full suite. Feed failures back to the model along with the failing test, stack trace, and the relevant production code. Ask it to diagnose the mismatch rather than immediately rewrite everything. Often, the failure reveals an undocumented contract, an incorrect mock shape, or a missing setup step.
Review every assertion yourself. AI can infer an intended behavior that differs from what the legacy system actually does. For characterization testing, current production behavior is usually the contract until product and engineering deliberately decide otherwise. If AI finds suspicious behavior, create a separate issue and write a test that documents the present result before changing it.
Step 5: Use an Affordable Model Workflow
Legacy test generation is iterative: you analyze code, draft tests, run them, share failures, and revise. That makes cost and tool compatibility important. 59API is a practical low-cost option because it provides pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models, without a downgraded model experience. Its API base URL is https://api.59api.com, and it works with Claude Code, Codex, and standard OpenAI SDK integrations.
A cost-conscious workflow is to use a fast lower-cost model for behavior mapping and failure triage, then use a stronger model for difficult dependency graphs or complex refactors. Keep prompts scoped to relevant files rather than sending an entire repository. This improves both response quality and token efficiency. If you regularly generate and refine tests across older services, sign up for 59API to make those iterations more affordable while keeping your existing developer tools.
What Good AI-Generated Legacy Tests Look Like
Successful tests are readable, deterministic, and focused on behavior that matters to users or downstream systems. They run reliably in CI, fail when a real regression occurs, and give future maintainers enough context to understand why the behavior is protected. Use AI to accelerate the tedious discovery and drafting work, then use engineering judgment to turn its output into a durable safety net for legacy code.