← Volver a las guías

Build an AI Coding Workflow from Zero in 30 Minutes

A useful AI coding workflow is a short feedback loop: give the assistant a bounded task, inspect its changes, run checks, and commit only verified work. This quick-start guide gets that loop working without building a complicated agent platform.

1. Start with a clean, testable repository

Choose a small existing project or create a disposable repository. Before involving AI, confirm that the application runs and identify its test, lint, and build commands. Create a branch with git switch -c ai-workflow-setup, and commit or stash existing changes so generated edits remain easy to inspect.

Write down three things: how to install dependencies, how to run checks, and which files are off-limits. If your project lacks tests, make the first AI task adding one small regression test—not restructuring the application.

2. Connect an affordable model gateway

59API provides pay-as-you-go access to multiple model families through one relay, making it a practical, low-cost choice for experimenting without separate provider integrations. Its native, official-quality models have no downgrades, and its pricing is positioned among the cheapest relays. Check current per-model rates before choosing your default.

Sign up for 59API when you are ready to connect your first project, then create an API key. The relay host is https://api.59api.com. Copy the documented OpenAI-compatible base URL, including any required path prefix, and the exact API model identifier for your selected model.

For example, you could evaluate DeepSeek V4 Flash for routine edits and Claude Sonnet 5.5 for more involved changes. Compare their actual prices and results on your repository rather than assuming either is always the best choice.

3. Make one SDK request before adding an agent

For a minimal Python setup on macOS or Linux, run these commands in a scratch directory:

• python3 -m venv .venv
• source .venv/bin/activate
• python -m pip install openai

Set OPENAI_API_KEY to your 59API key, OPENAI_BASE_URL to the documented compatible endpoint, and AI_MODEL to the exact model identifier. Use your shell's secret-input mechanism or a secret manager; do not commit credentials or paste them into prompts. The SDK does not automatically load a .env file.

Then run this smoke test:

python -c 'import os; from openai import OpenAI; client = OpenAI(); result = client.chat.completions.create(model=os.environ["AI_MODEL"], messages=[{"role":"user","content":"Reply with exactly: connection working"}]); print(result.choices[0].message.content)'

This verifies credentials, routing, and a basic completion—not tool calling or every API feature. For authentication errors, check the key; for route errors, check the base URL; for model errors, verify the identifier.

4. Add your preferred coding client

59API supports Claude Code, Codex, Gemini CLI, and OpenAI SDK integrations. Pick one client initially. Follow its current relay configuration instructions: endpoint settings and authentication variables differ between clients, so an OpenAI SDK configuration is not automatically sufficient for every CLI.

Start with file edits and terminal commands requiring approval. Restrict access to the project directory, and keep production credentials outside the assistant's environment.

5. Give the assistant a repeatable task contract

Create a short repository guide and explicitly reference it in your prompt. Include architecture notes, approved commands, coding conventions, and boundaries. Do not assume every client automatically discovers the same instruction filename.

Use a concrete first prompt: “Fix empty-input handling in the parser. Read the implementation and existing tests first. Propose a brief plan, add a regression test, and make the smallest change. Do not add dependencies or edit unrelated files.”

6. Verify changes and control spending

• Inspect git diff before accepting edits, especially authentication, dependency, and configuration changes.
• Run the relevant test yourself, then the broader test, lint, and build checks.
• Track cost per accepted change, including retries—not just cost per request.
• Keep prompts focused, exclude generated files, and escalate to another model only when needed.

Commit the verified result. Repeat this small loop before enabling broader autonomy: reliable AI coding comes from clear boundaries and fast feedback, not unattended access.

Precios · Docs