Updating attorney website profiles with browser automation, then writing the playbook
Updating attorney bio pages one field at a time in a slow WordPress editor was a chore nobody wanted to own, so the goal was to automate the fill-in work without losing accuracy. It nearly did, once. A jurisdictions error the agent reported as done was only caught because someone checked, which is why the playbook now insists on that check every time.
Category
Browser automation with a documented playbook.
What I was trying to accomplish
Update attorney bio pages on the firm's WordPress site accurately from Word documents, and turn a one-off chore into a repeatable, documented process.
The data challenges
Extracting structured content from Word bios, filling a WordPress editor's slow, error-prone nested repeating fields, and mapping free-text practice descriptions onto a 63-term, three-tier taxonomy, flagging ambiguous matches for a person to decide.
The result
Two profile pages staged for review, never published by the agent, a one-page practice-areas reference, and a process document with known pitfalls and mapping rules. A jurisdictions error the agent had reported as done was caught on review, which is why the playbook now includes a field-by-field verification step.
Technical details
Tools: Claude in Chrome, Claude Cowork, python-docx
New and existing attorneys' biography pages on the firm's WordPress site had to be updated from Word documents: bio text, education, bar admissions, jurisdictions, and practice areas. The editor used nested repeating fields that are slow and error-prone to fill by hand.
Claude extracted the structured content from each Word file, then used Claude in Chrome to fill the editor directly, using page scripting where the nested fields would not accept ordinary form input. Free-text practice descriptions were mapped onto the firm's controlled taxonomy, and ambiguous matches were flagged for a person to decide rather than guessed.
The challenges were typical of real websites: a new attorney's page had been created by duplicating an old one and carried leftover content that had to be found and removed, and the editor's repeating fields needed direct manipulation. On the first page I caught a jurisdictions error after the agent reported it done, which is why the playbook includes a field-by-field verification step. Leaving publication to a human was a design decision, not a limitation.
Pull up a chair.
Let's talk through where things stand, and where they could go.