Five short stories, told from the customer's side: who they are, what was in the way, what they typed, what came back, what it cost. They are composed from real runs — the public demo note and the calls we measured on 2026-10-17 — not from any one customer, and nobody is named. Every number below comes from a logged call.
1. "My agent could finally read the note my client pasted"
Who: an indie developer running Claude Code and an OpenClaw agent for a small e-commerce client in Shenzhen.
Before: the client pasted Xiaohongshu links into chat all day. The agent fetched the URL and got the shell of the page — on the demo note, 860 visible characters, mostly "打开App查看更多". yt-dlp said No video formats found, because it only knows video notes and this one was seventeen pictures. The developer was screenshotting notes into the chat by hand.
What they did: one line in the MCP config, or the xiaohongshu-note-to-text skill from ClawHub. Then pasted the same link.
After: 35,059 characters back — the caption, every one of the seventeen images described and its text read out, the key points with a quote each, the like and save counts, and now the first five comments under the note. The agent answered the client's question without anyone opening the app. Typical Xiaohongshu read: about half a minute.
Cost: one to three credits per note (six images per credit), nothing for a note anyone has already read. No cookie, no login, no account of theirs anywhere near it.
2. "Fifty competitor notes, one page, before the Monday meeting"
Who: a consumer-insight analyst at a skincare brand, preparing a 对标 (benchmark) deck.
Before: the dashboards her team rented listed titles and like counts. To know what buyers actually said she opened notes one at a time and copied comments into a spreadsheet. Fifty notes took an afternoon and the quotes were never checked.
What they did: sent the fifty links to POST /api/v1/digest/batch with insight: true and synthesis: true (from Claude Code it is one compare_links call).
After: each note came back with its caption, image text, the top comments the share page serves, and a consumer-insight block: audience, claims, objections, who is asking where to buy, products and prices named, questions readers ask. Every finding carries a quote, and the engine marks whether that quote was found word for word in the note or its comments. On top, one synthesis over all fifty: the themes and hooks that recur, the claims every post makes, the objections readers raise, the gaps nobody covers — each example citing the note number. A three-note batch we timed finished in 22 seconds; the comparison itself took 13 seconds and billed one credit.
Cost: fifty notes with insight is about 100 credits, plus 5 for the synthesis — roughly ¥15 on the ¥288 pack, or about three percent of a Studio month.
3. "The pitch needed to say why the video worked, not just that it did"
Who: a content strategist at a small agency, pitching a food brand.
Before: she could show the client a viral Douyin or Xiaohongshu video and say "do this". She could not say what the first three seconds did, where the beats fell, or what the title formula was — and the client's own team would ask.
What they did: read the link with breakdown: true (爆款拆解).
After: the hook as spoken and as shown on screen, the structure as timed beats with a quote at each, the title formula, the cover text, the call to action, and a reusable template her writers could fill in. The quotes are checked against the transcript, so when the deck says "at 0:12 she says X", X is really at 0:12.
Cost: the read plus one credit.
4. "I wanted the buying signals, and I wanted to know when there weren't any"
Who: a cross-border seller choosing products for a TikTok Shop, using Xiaohongshu as the research ground.
Before: sorting comments for "求链接", "多少钱", "哪里买" by eye, and never sure whether a quiet comment section meant no demand or just no data.
What they did: read candidate notes with insight: true and looked at the purchase_intent and products sections.
After: where buyers were asking, the quotes are there with the comment they came from. Where they were not, the section says found: false and the engine writes why — on a film-review note we ran, it reported plainly that the material had no purchase link, price or request to buy, instead of inventing one. That is the part that makes the output usable for a decision: silence is reported as silence.
Cost: one credit per note plus one for the insight. A note that cannot be read is not charged.
5. "It lives in our 飞书 table now; nobody calls an API by hand"
Who: an operations lead at an MCN, with a team that works in Feishu (Lark) and Coze, not in a terminal.
Before: an intern pasted links into a form, a developer ran a script at night, and the results arrived as a zip of text files.
What they did: installed the LinkDigest field shortcut in 飞书多维表格, so a column fills itself from the link in the row — transcript, image text, key points, like/comment/save counts, and the top comments. For the workflow builders, the OpenAPI spec went into Coze and Dify as a plugin; batch jobs post back to a signed webhook when they finish.
After: the table updates as links are pasted. The team sees the same fields the API returns, in the tool they already open every morning.
Cost: the same credits as any read; there is no per-seat price.
If one of these is close to your week, the fastest way to find out is to paste a link on the home page — the first look is free — or read the docs. For the 对标 workflow in Chinese, see 对标账号→表格.
Platform notes: Xiaohongshu. Earlier on this blog: Reading links your agent can't open. 中文版：LinkDigest 用户故事。