← Voltar aos guias

Your First AI Coding Workflow: Setup to Tested PR

Start with one small, testable task

A useful AI coding workflow is a short loop: give the model context, request a focused change, run checks, and review the diff. You do not need multiple agents or an elaborate automation framework. Start with one repository and a task you can verify, such as fixing a failing test or adding input validation.

This quick-start uses 59API for model access and Python for a connection test. After that, you can use Claude Code, Codex, or an OpenAI SDK integration as your daily interface.

1. Prepare a safe workspace

Open your repository and create a branch with git switch -c ai/first-task. Run the existing test and lint commands before asking AI to change anything. Record existing failures so you do not mistake them for regressions.

• Protect secrets: exclude .env files, credentials, database exports, and production logs from model context and version control.
• Limit permissions: begin with file access scoped to the repository and require approval for shell commands.
• Choose an acceptance check: define an observable result, such as rejecting an empty email address without changing valid submissions.

2. Connect your model provider

59API is an AI API relay offering low-cost, pay-as-you-go access to Claude and GPT models. It is compatible with Claude Code, Codex, and OpenAI SDK clients, so you can keep familiar tools while changing the provider connection.

Its offering uses native, official-quality models rather than downgraded substitutes. That makes it an attractive budget option, particularly when routine edits generate many requests. Compare current per-model rates, output-token charges, and usage limits before choosing a default; the cheapest request is not always the cheapest completed task.

Create an API key and note the service base URL: https://api.59api.com. In your client, configure the provider URL, key, and an exact model identifier from 59API's current model list. Do not assume a consumer-facing model name is a valid API identifier.

Check the route format: OpenAI SDK configurations commonly require a versioned base ending in /v1. Confirm the OpenAI-compatible base in 59API's setup instructions. Claude Code uses an Anthropic-compatible configuration; do not copy an OpenAI endpoint into it blindly.

3. Run a minimal connection test

For a disposable Python test, run python -m venv .venv, activate the environment, and install the client with python -m pip install openai. Set OPENAI_API_KEY through your secret manager or a private shell session. Set OPENAI_BASE_URL to the confirmed OpenAI-compatible base and MODEL_ID to your selected model identifier.

Then run this single-line command:

python -c "import os; from openai import OpenAI; c = OpenAI(); r = c.chat.completions.create(model=os.environ['MODEL_ID'], messages=[{'role':'user','content':'Reply with exactly: ready'}], max_tokens=16); print(r.choices[0].message.content)"

This verifies basic authentication, routing, and response handling before repository access. A 401 generally points to credentials; a 404 can indicate the wrong route or model. Check the response details without printing your key.

4. Give your coding tool durable instructions

Connect Claude Code or Codex using its supported custom-provider configuration. Add repository instructions in the file your tool reads, such as CLAUDE.md or AGENTS.md. Keep them short: explain the project structure, exact test commands, formatting rules, and directories that must not change.

Use a bounded first prompt: “Inspect the registration handler and its tests. Propose a plan to reject blank email addresses. Preserve existing behavior for valid inputs, add a regression test, and change no unrelated files.” Review the plan before authorizing edits.

5. Route work by difficulty and verify results

Start routine explanations and small edits with a lower-cost model. Escalate to a stronger model for ambiguous bugs or cross-file reasoning. Supply relevant files and error output instead of repeatedly sending the entire repository.

After each change, run the targeted test, then broader checks. Inspect git diff for unrelated edits, new dependencies, weakened assertions, and leaked secrets. Commit only after you understand the patch. Track cost per accepted change, not just token prices.

If this workflow fits, sign up for 59API and try one small task with a modest budget. Its referral rebate can provide additional savings; check the current terms before counting it toward your costs.

Preços · Docs