Verified against Perplexity Comet · 2026-08-04
Brief the Comet browser agent for a bounded multi-step task instead of one open-ended ask
A task brief for Perplexity Comet that names every site to visit, the exact data to pull from each, the stop condition, and what to do when a step fails — so an agentic browsing run stays inside a defined scope instead of wandering or quietly skipping a hard step.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Run this as a Comet agentic browsing task. Treat the steps below as a fixed checklist to complete, not a loose suggestion of where to start looking. GOAL Confirm the current return policy window and whether it differs for sale items, across our three main furniture retail competitors. SITES TO VISIT, IN ORDER oakwoodfurniture.com/returns, cascadehome.com/policies/returns, birchandbolt.com (search their FAQ or footer for return policy) WHAT TO PULL FROM EACH SITE Return window in days, whether sale/clearance items are excluded, whether return shipping is free or charged to the customer STOP CONDITION Stop once all three sites have been checked, regardless of whether every field was found on every site. RULES FOR NAVIGATION - Stay within the sites listed above unless one of them explicitly links to a page that is clearly necessary to complete the goal (for example, a company site linking to its own SEC filing) — if you go beyond the listed sites, say exactly which page and why. - Do not fill in a missing data point by inferring it from a different, unrelated page just to complete the checklist — mark it as "not found on this site" instead. - If a site requires a login, a paywall, or a CAPTCHA to access the specific page needed, stop at that page and report it as blocked rather than attempting to find a workaround or substituting a different, lower-quality source without saying so. WHEN A STEP FAILS If a site has changed its layout, moved a page, or no longer has the requested information, say specifically which step failed and why, then continue to the remaining steps rather than abandoning the whole task because one step didn't go as planned. OUTPUT FORMAT 1. A table or list, one row per site: Site | Data found | Notes (including "blocked," "not found," or "page structure changed" where relevant). 2. A short summary answering the original goal, built only from what was actually found — if the goal can't be fully answered because of gaps in the table above, say so rather than filling the summary with confident-sounding language that outruns what was actually verified. 3. A one-line note on anything encountered that seemed relevant to the goal but was outside the scope of the sites listed, in case it's worth a follow-up task.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Comet's browsing agent operates by actually navigating pages rather than querying a search index, which means it can reach information a standard web search never indexes well — a specific policy page buried three clicks into a footer, a filing that sits behind a site's own document viewer — but that same navigational freedom is exactly what makes an unbounded brief risky, since an agent told only the end goal has to decide for itself where to look and can drift into pages that seem plausibly related without ever being told to stop. Naming the exact sites and the exact fields to pull from each converts an open-ended browsing task into a checklist the agent can complete or visibly fail to complete, rather than an exploration that could end almost anywhere depending on what it happens to find first. The explicit rule against inferring a missing field from an unrelated page addresses a real risk specific to agentic browsing: an agent under implicit pressure to fill in a complete-looking table can quietly substitute a plausible-sounding number from a different page — a competitor's general FAQ instead of their specific returns page — which produces a table that looks complete but contains a fabricated data point wearing the same formatting as the verified ones. And the per-step failure-handling rule matters because a browsing agent that hits one blocked or restructured page can otherwise treat that as reason to either abandon the whole task or silently skip the failure and report the remaining steps as if nothing went wrong — requiring it to name the specific failure and keep going is what produces a genuinely useful partial result instead of an all-or-nothing one.
Verified against
Perplexity Comet 2026 browser agent · 2026-08-04
Changelog
- 2026-08-04 — Initial publish, verified against Perplexity Comet agentic browsing mode.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
