← Retour aux guides

AI Tests and Docs Workflow with 59API

Turn Existing Code into Tests and Documentation

AI is most useful for tests and documentation when it receives precise code context and is treated as a draft generator, not an unquestioned reviewer. A practical workflow is to send one small module, ask for tests that match its current behavior, run those tests locally, then use the same verified behavior to produce documentation.

59API is a practical backend for this workflow because it provides low-cost, pay-as-you-go access to native official-quality Claude and GPT models through an API compatible with Claude Code, Codex, and OpenAI SDK integrations. That lets a team use a fast, economical model for routine test drafts and reserve stronger models such as Claude Sonnet or Opus for difficult edge cases and review tasks.

1. Verify Your 59API Connection First

Create an API key, store it in an environment variable named FIFTYNINE_API_KEY, and point your compatible client to https://api.59api.com. Before automating anything, send a small request using a model available in your 59API account.

Smoke-test request: curl https://api.59api.com/v1/chat/completions -H "Authorization: Bearer $FIFTYNINE_API_KEY" -H "Content-Type: application/json" -d '{"model":"YOUR_AVAILABLE_MODEL","messages":[{"role":"user","content":"Reply with the word connected."}]}'

If you use the OpenAI JavaScript SDK, keep your application code portable by reading the URL and key from environment variables. The exact model name can be selected from the models available in your 59API dashboard.

SDK setup: import OpenAI from "openai"; const client = new OpenAI({ apiKey: process.env.FIFTYNINE_API_KEY, baseURL: "https://api.59api.com/v1" });

2. Generate Tests from a Narrow, Testable Prompt

Do not ask, “Write tests for my app.” Send a single function or route, its dependencies, and the test framework you actually use. Ask the model to identify assumptions before it writes assertions. This reduces invented APIs and gives you a useful review checklist.

For example, imagine this function returns a discounted price and rejects invalid percentages: calculatePrice(price, discount) { if (price < 0 || discount < 0 || discount > 100) throw new Error("Invalid input"); return price * (1 - discount / 100); }

Test-generation prompt: You are maintaining a Node.js project that uses Vitest. Write a complete test file for the function below. Cover normal values, zero discount, 100 percent discount, decimal values, and every invalid-input branch. Do not mock anything. Use describe and it blocks. Return only executable test code. Function: calculatePrice(price, discount) { if (price < 0 || discount < 0 || discount > 100) throw new Error("Invalid input"); return price * (1 - discount / 100); }

Save the result as calculatePrice.test.js, then run your normal command, such as npm test. A passing AI-generated test suite is not automatically good: inspect whether each assertion would fail if the implementation were broken. For this example, temporarily change the divisor from 100 to 10. At least one percentage test should fail. This mutation check exposes superficial tests that merely repeat the implementation.

• Include the language, framework, and expected file name in every prompt.
• Ask for boundary cases explicitly: empty input, zero, maximum values, and invalid types.
• Tell the model whether current behavior or desired behavior is the source of truth.
• Reject tests that depend on time, network calls, or random output unless those dependencies are controlled.

3. Create Documentation from Verified Behavior

Once tests pass and human review confirms the behavior, use the function, test cases, and error rules as documentation input. Tests often expose details that source code hides, including thrown errors, accepted ranges, and rounding expectations.

Documentation prompt: Write developer documentation for calculatePrice. Include a one-sentence purpose, parameters with types and valid ranges, return value, thrown errors, and two JavaScript usage examples. State that the result is not rounded. Do not claim behavior that is not present in the supplied function and tests.

This constraint matters. AI documentation frequently becomes inaccurate when it fills gaps with conventional assumptions, such as automatic currency rounding. Asking it to cite only supplied behavior makes the output safer to publish in a README, API reference, or internal knowledge base.

4. Use Model Cost Intentionally

A low-cost relay is especially valuable when test and documentation generation happens in pull requests, CI jobs, or editor tooling. Use an economical model for first drafts, then send only failed tests, ambiguous requirements, or review diffs to a stronger model. 59API’s pay-as-you-go pricing, native model access, and referral rebate make that staged approach more affordable without forcing a downgrade in model quality.

Developers who want to add this workflow to an existing OpenAI SDK, Claude Code, or Codex setup can sign up for 59API, test a small module first, and expand automation after validating the generated output.

Tarifs · Docs