Abrantes - Use Abrantes with AI

Important disclaimer

This workflow is experimental. Do not rely only on AI to create experiments.

Preferably use a coding-focused AI tool that can read files in a local Abrantes repo clone. Some online tools worked, but many didn't work.
Prompt your target URL and experiment hypothesis.
Use a skill or provide the SKILL.md file with the prompt. (Or reference the Abrantes docs URL if upload is not supported).
Run and validate the output manually before launch.

Best online result in these tests. Worked well with uploaded skill and prompt.

Good online results when prompting with the Abrantes URL.

Desktop workflow worked well. The model scanned repo context and generated correct code.

Good output when permissions were granted for required actions.

Worked reliably in these tests. Allow curl so the model can access docs and pages.

Worked well using the /experiment skill workflow.

Worked after being directed to use the experiment skill. Initial result hallucinated.

It worked well with the free "experiment" plan.

These results were noted on 2026-02-28 and may change as tools improve.

GitHub + Claude Haiku 4.5 (online): could not access the tested page.
Gemini online: hallucinated the implementation.
Mistral Le Chat: hallucinated the implementation.
Grok 4.20 (beta): many selector and implementation mistakes.
Qwen 3.5 plus reasoning: hallucinations and struggled with add2dataLayer.
DeepSeek: generated hallucinations and ignored the real page context.
Google NotebookLM: ignored uploaded SKILL.md guidance.
ChatGPT Mac app (free plan): could not access tested URLs directly in this workflow.

Evaluate a tool + model pair, not only the model alone.
Desktop coding tools can perform better than online tools because they can inspect local repo files.
If your tool cannot access the tested page URL, the tool can't generate a working experiment.
Before launch, validate selectors, assignments, rendering, persistence, and tracking events.