Your assistant can look things up. Now give it a useful beat.

What actually changed

DeepSeek’s Chinese release notes dated 17 October announce its in-house search API and the V4.1-Flash-1017 update. Its guide distinguishes Search, which returns structured search results, from Fetch, which reads one page into Markdown, a plain-text document format. Those are separate jobs. Finding a promising link is the start of research; reading it supplies the material from which an answer can be checked.

Build a beat, not a firehose

Imagine a student looking for free animation workshops near home. A general answer about learning animation is easy to generate and soon forgotten. A short list of workshops that are actually open this week solves a different problem. Start with a beat: one subject, one place and one audience. The assistant’s job is to find candidates, read organiser pages, and turn the useful facts into a small comparison. A beat is simply a subject you keep following.

For a fictional first edition, make three entries in a document: a weekend workshop, an online talk and a scholarship deadline. Give each a title, organiser, date, eligibility, cost and original link. Mark missing information as unknown. Now ask which two entries your intended reader can actually use. A free workshop two hundred miles away may be less useful than an online session with a small fee. Selection requires knowing the reader’s situation; it is more than summarising whatever the search returned.

Why reading the page changes the answer

Suppose a search preview says applications close on Friday, but the organiser page says applications closed last Friday. A fluent summary of the preview would confidently recommend an expired opportunity. Reading the page lets you distinguish the event date, application deadline and date the article was published. These are three different dates. Keep them in separate fields, then write the short reader-facing explanation from those facts.

This creates an achievable small service: a carefully selected digest for a club or a group of learners. Begin with manual search and an ordinary assistant. If people find the selection useful repeatedly, automation can take over finding and reading candidate pages. Your distinctive contribution is the useful selection and explanation. More links alone do not make the service better.

Make tomorrow’s edition less work

Keep the previous entries and compare what changed: a new opportunity, a moved deadline or an event that has ended. Ask the assistant to draft only the changes, with the original links beside them. This is a simple example of a stateful workflow: today’s work starts from yesterday’s checked state. It can save the reader from rereading the entire list. It also gives you a clear reason to publish again: there is something materially different to know.

Here is a fictional tool budget using that published tariff, not a measured run. Twenty billable searches and sixty billable page reads make eighty requests. At 5 yuan per thousand, eighty divided by one thousand times five is 0.40 yuan, or CNY 0.40. Model generation and your checking time are additional. The guide separately dates charging for server-side tool calls through the named agent clients to 1 November; do not assume all direct API use is free until then.