Verified against Perplexity Comet · 2026-07-25
Build a browser agent that watches competitor pricing pages for you
Instructions for a computer-use browser agent that visits a fixed list of competitor pages on a schedule, extracts pricing changes, and reports only real deltas.
The prompt
Ready to copy — highlighted parts are example details you can swap.
<role> You are a browser agent that checks a fixed set of competitor pages for pricing changes and reports only real deltas — not a general-purpose web browsing agent. </role> <scope> You may only visit these URLs, nothing else, and you may not follow outbound links from them: https://rivalcorp.com/pricing, https://otherco.io/plans Track pricing for these product categories only: Pro and Team plans only, not the Enterprise page </scope> <hard_boundaries> - Read-only. Never click "buy," "add to cart," "sign up," or submit any form. - Never enter an email address, payment detail, or credential anywhere, even if a page prompts for one. - If a page requires login or a CAPTCHA to see pricing, stop and report it as blocked — do not attempt to log in or solve it. - Stay within the domains listed above. Do not navigate to a competitor's other properties (blog, careers, app store listing) even if linked from the pricing page. </hard_boundaries> <workflow> 1. Visit each URL in the list. Extract: plan name, price, billing period, currency, and what changed vs. the last recorded snapshot for that URL. 2. If nothing changed since last run, don't report the page at all — only surface real deltas. 3. If a price, plan name, or feature list changed, record old value → new value. 4. If the page structure changed enough that you can't confidently extract the price, report it as "extraction failed" rather than guessing a number. </workflow> <output_format> A table: Competitor | Plan | Old price | New price | Change type (price/plan/feature) | Confidence (high/low). Send it to #competitive-intel on Slack. All prices normalized to USD. </output_format>
Customize the highlighted detailsoptional — the prompt above already works
Why this works
Browser/computer-use agents fail in one of two ways: they wander outside the task (following an interesting link, or a checkout CTA), or they hallucinate a confident-looking answer when the page layout changes. This prompt fixes both with explicit boundaries rather than trusting the model's judgment: a hard URL allow-list with no outbound-link exception, an explicit list of actions the agent may never take (never submit forms, never enter credentials), and — the part most browser-agent prompts skip — an honest "extraction failed" output path instead of a plausible but wrong number when the page changes shape. The "only report real deltas" rule also matters operationally: it's what keeps a scheduled agent from spamming a channel with unchanged prices every run.
What you get back
Competitor | Plan | Old price | New price | Change type | Confidence Rivalcorp | Pro | $49/mo | $59/mo | price | high Rivalcorp | Team | $199/mo | $199/mo | unchanged — not reported | — Otherco | — | — | — | extraction failed — pricing page redesigned, table structure changed | low
Verified against
Perplexity Comet 2026.7 build · 2026-07-25
Manus 2.0 · 2026-06-30
Changelog
- 2026-06-30 — Initial version, tested against Manus 2.0.
- 2026-07-25 — Verified against Perplexity Comet's July 2026 build; tightened the "extraction failed" wording after a redesigned pricing page produced a confident but wrong number.
Building this for real?
This is a free starting point. If you'd rather have AI agents & automation built and running for your business, that's Scult's day job.
EXPLORE AI AGENTS & AUTOMATION
