Cross-AI Workflows

The layer between what Claude can decide and what Claude can touch

Claude drafts a flawless image prompt in thirty seconds — lighting, posture, palette, mood. Then you open the generation tool, paste it in, and click Submit yourself. Claude reasoned about the work; reaching the tool was on you. That gap is the subject of this book, and giving Claude a browser login is not the answer. Building the layer between them is.

The bridge is a three-layer pattern: Claude plans, a browser-automation script executes, and Claude verifies. The worked example runs end to end — a 24-page picture book where Claude writes a prompt per page, a Playwright script drives an image generator’s web interface to produce eight variants each, Claude reviews every batch against the page text through the vision API, and a human makes the final pick. 192 images, generated overnight, ranked by Claude by morning.

The three layers are separable. Swap out the generation tool and only the execution script shifts. Change the verification criteria and only the Claude prompt changes. Twelve chapters build the template, then show how the same structure extends to any tool Claude cannot control directly.

One pipeline, all three layers

• The three-layer pattern — Reasoning, execution, verification: where Claude belongs in an automation pipeline, where the script belongs, and why the boundary between them is what makes the whole thing swappable.
• Attaching to a real browser — Connecting Playwright to a live, logged-in browser session so the script drives the same interface you would, without handing credentials to anything.
• The spec file — One file that drives the entire pipeline: page text, scene descriptions, style constraints, and the state every other layer reads from.
• The prompt handoff — Building prompts that Claude and the generation tool both understand, and structuring Claude’s output so the script can consume it without parsing surprises.
• Reliable UI automation — Driving the web interface through submission, detecting when generation is actually complete, and downloading the results — the details that separate a demo from a pipeline that runs 24 pages unattended.
• Progress tracking and resume — A long run will be interrupted. The pipeline records what finished, so a restart picks up at page eleven instead of page one.
• Graceful failure handling — Timeouts, refused generations, and half-loaded pages, handled so one bad page costs one retry instead of the whole overnight run.
• Vision-API verification — Feeding each image batch back to Claude with the page text it should illustrate, and getting a ranked assessment instead of an unsorted folder of 192 files.
• The human selection layer — The final pick stays human. The book covers the selection interface and why the ranked list feeds that step rather than replacing it.
• Generalizing the pattern — Extending the same three-layer structure to any external tool that exposes a web interface but no API.

A preview

Claude did not click Submit. You did. That gap — between what Claude can reason about and what Claude can reach — is the subject of this book. Giving Claude a browser login is not the answer. Building the layer between them is. A script takes Claude’s output, drives the external system, and returns results for Claude to evaluate.

Builders who hit the edge of the API

This book is for readers who already write Python, call Claude from code, and know basic web-automation concepts — waiting for elements, clicking buttons, extracting content. It does not teach Playwright; it documents where Claude fits in the pipeline, how to structure its inputs and outputs, and how to verify that what it produces is actually working. By the end you have a working template for orchestrating any tool Claude cannot control directly.

A longer excerpt is available to newsletter subscribers.

Buy Ebook on Amazon