Overview
What I Built
How It Works
- Claude Code orchestrates the work. It writes and adjusts recipes, runs the engine, reads the result logs, and iterates when a site changes.
- Playwright drives a real Chromium browser through each form. The recipe action types include navigate, fill (with optional sequential typing to look natural), select, click, upload, and a four-strategy click fallback for the awkward sr-only inputs and role-based checkbox buttons that modern forms hide behind.
- Tesseract OCR reads image captchas. A
solve_captchaaction runs multi-PSM OCR to crack the digit captchas on the older directory templates, and asolve_mathaction regex-parses the "X plus Y equals" style questions straight out of the form row. - Reusable JSON recipes are the unit of reuse. Each one is a small script of actions plus a success check, so adding a new directory family usually means copying a similar recipe and adjusting selectors rather than writing anything from scratch.
- Success detection is explicit. A
check_successstep verifies the real confirmation, by text marker, URL change, URL pattern, or in one case a JavaScript alert dialog registered before the click. This matters because an earlier headless version of the engine claimed wins that were actually silent rejections. The recipe-driven version only counts a submission when the site itself confirms it. - The CLAUDE.md instruction hierarchy keeps the whole thing honest. Conventions, structure, and hard rules live in committed instruction files that Claude Code reads first, so the agent stays inside the lines instead of improvising.
Results
- 110+ directories targeted from a single committed source list, spanning a wide range of domain authority.
- 15 reusable recipes that generalize across form patterns, so the engine scales to new directories without new code most of the time.
- Confirmed real submissions, not guesses. In one live session on one product, 73 sites were attempted and 20 came back as confirmed real submissions, including a live public listing that anyone can load in a browser today.
- Proven across more than one product. Running the same 15 recipes unattended against a second product confirmed 11 of the 15 working across repeated runs, again including a live public listing. That second run is also where new capabilities got added: the math-captcha solver, the alert-dialog success check, and the multi-strategy click fallback.
- 60+ chapters audited by parallel agents. On a single long manuscript, parallel AI auditors checked the whole book against an established canon and surfaced consistency fixes a single pass would have missed.



