CASE STUDY 06 · ADMIN / CMS
Love Bohol Portal
A browser agent that completes multi-step tasks on sites with no API to call, by actually driving a real browser.
ROLE
AI + full-stack engineering
STACK
Claude API · browser-use · Next.js
TIMELINE
[x weeks]
TYPE
Client project
01 — PROBLEM
Some tasks only exist behind a web UI — no API, no webhook, just a page meant for a human to click through. Doing them by hand, over and over, was the only option.
02 — ARCHITECTURE
Task instruction
natural language
LLM planner
Claude API
browser-use
drives real browser
Target site
no API available
Result
extracted or confirmed
03 — KEY DECISIONS & TRADE-OFFS
Drive a real browser, not scrape HTMLMany of these sites are JS-heavy and actively resist scraping. A real, visible browser session behaves exactly like a human visitor.TRADE-OFF Slower and more resource-heavy per task than a direct HTTP request would be.
LLM re-plans after every page, not a fixed scriptSites change their layout often enough that a hardcoded click sequence breaks. The agent reads the current page and decides the next action each step.TRADE-OFF More LLM calls per task, and occasional wrong turns that need a retry step.
Stop and ask on anything destructiveAn agent that can click anything needs a hard line before payments, deletions or anything irreversible — those steps pause for a human to confirm.TRADE-OFF Fully unattended runs aren't possible for tasks that cross that line.
04 — STACK
LLM
Planning & page understanding
browser-use
Real browser driving
Next.js
Task dashboard
Vercel
Hosting & deployment
05 — RESULTS
[metric]
Tasks completed without a human step-in.
[metric]
Average steps per completed task.
[metric]
Hours of manual browsing replaced per week.
NEXT CASE STUDYSet Dialer →