AI Breakfast Shanghai
← Back to Notes

TL;DR

Demo: A Newsletter That Writes and Sends Itself

This week's centerpiece came from a self-described non-technical member who helps business owners and entrepreneurs adopt AI. His project: a weekly workforce-intelligence newsletter that researches, writes, formats, and emails itself. He hits "go" once a week — everything else is automatic.

It runs as a scheduled Claude task with eight steps. Determine the date — models still get it wrong, so he built a dedicated tool. Read brand and methodology from local seed files. Check a coverage log to skip articles already covered. Web-search the past seven days for workforce restructurings and AI-driven layoffs, weighted to China and APAC. Then write a fixed eight-section report, render a PDF from a reusable Python template, email it via a custom MCP server — a connector that lets the AI use an outside tool — and append the week's articles to the log.

The clever part is memory: the coverage log keeps week two from re-reporting week one's news. A second task fires an hour later, reads the finished report, and turns it into LinkedIn and WeChat drafts. He chose this kind of tool partly for local folders — his brand and context files stay on his machine, editable directly, not trapped in a cloud "project." The hardest part, he said, wasn't code — it was the prompt: turning raw news into "what this means for employees, managers, and executives."

The table's main fix: separate collection from publishing. Search runs only at report time, so a Tuesday story can be stale by Monday — better to scrape daily and write weekly. For deduplication and trend-spotting, one member pointed to Gensim and spaCy, which cluster similar articles and surface "this topic got 100 mentions this week."

Group Discussions

Build an MCP, or Just Write a Script?

To let the agent send email, he built a custom Outlook MCP server — the worst part of the whole project. Register an app in Microsoft Entra ID, grant Mail.Send permission, copy the tenant and client IDs, and grab the client secret immediately — it vanishes when you close the window (he re-registered the app twice after missing it). One stray character in the config file later crashed his desktop app for half a day.

The developers at the table had a blunter view: he overbuilt it. Most services don't need an MCP — write a small script to send the email and tell the agent to run it. The nuance that rescued his approach: Microsoft deliberately cripples plain email-sending (SMTP) on Outlook, so the heavier authenticated path may genuinely have been needed.

The keeper takeaway: default to a script the agent can call, and reach for an MCP only when the service blocks the simple path, as Microsoft does.

Why Your MCP Burns the Whole Token Budget

One member keeps hitting daily limits on trivial work — asking an agent to tidy a small Notion database or delete 60-odd entries — even on Opus. The culprit isn't the task. A connected MCP loads every one of its tool definitions into the model's working memory before the agent does anything.

The fix the table named: dynamic tool sets. Instead of dumping every tool up front, the MCP exposes a lightweight router — just metadata — and the agent pulls only the tools it needs. Not every MCP supports this yet; if yours doesn't, that's where the budget goes.

One more distinction — token cost and token efficiency aren't the same (tokens being the text chunks an AI bills by). A model can look cheap per token yet waste them, burning several times as many for the same answer. The table's read: GPT-5.5 currently leads on efficiency, a $20 plan covers normal use, and heavy MCP workflows are what blow past the limits.

Anthropic Quietly Taxes Third-Party Tools

Why doesn't the OpenCode CLI work on a Claude subscription anymore? The table's read: third-party harnesses — the apps you run an AI model through — reportedly don't play well with prompt caching (a billing discount for reusing the same instructions), so they generate far more compute than the first-party apps.

Anthropic's answer isn't an outright ban. As the table understood it, third-party harnesses get routed to API-rate billing: a $100 plan includes roughly $100 of API credit you can spend this way, but at API rates that runs out fast. Same effect as a block, with a grace period — experiment with outside tools, but not much.

Scraping: When the CAPTCHA Targets Your AI

The hotel-influencer marketplace founder returned to the scraping wall and named what's behind it. Cloudflare Turnstile blocks bots at the door; harder still, Cloudflare now splits login into a separate worker, so the auth flow isn't on the page you're scraping. And some platforms' CAPTCHAs now serve adversarial images — subtly perturbed so a human sees an elephant while a vision model sees a car. Automated solving collapses; he pays human solver services, costly on the worst platforms.

He also sharpened the real problem. Bulk influencer data isn't the bottleneck — he already buys full RedNote and Instagram datasets from a provider for a reasonable fee. What he can't do is search: "which influencers used this hashtag" needs a logged-in session, and platforms ban bot-like behavior. Instagram tolerates roughly 100 posts before an IP block; RedNote is far harsher and returns thin data unless you're logged in.

Practical notes: gallery-dl handles bulk downloads across many sites if you run it locally and don't need it automated, and Apify has a ready-made RedNote scraper several found reliable. For small, specific lookups, paid scraping APIs (about $5 per 1,000 requests, billed per post) are worth it; at scale the economics break.

Quick Hits

The Great Firewall is an AI-adoption tax. A designer admitted she uses AI for almost nothing — "what flower is this" on Doubao — and named the reason plainly: behind the wall, the good tools are enough of a hassle that laziness wins. She's waiting until she leaves the country to bother. Access, not capability, is the real adoption ceiling for many people.

You can't get comfortable in your tooling. A developer who builds hobby apps in languages he doesn't know — Rust, a Swift app for a homemade thermometer — named his real frustration: tools go obsolete monthly. You finish a setup and it's already superseded. He's still hunting for a stable online/offline workflow that survives switching between laptop and phone.

Software that fixes itself. A software architect at a 3D-printing company is wiring monitoring straight to deployment: a scheduled job scans the system's own metrics and error logs, files Linear tickets for the bugs it finds, hands them to AI agents, and lets the resulting fixes merge and ship on their own. Half-built, but closing that full loop is the goal.

Other Resources

  • Claude: Anthropic's AI assistant. Its scheduled tasks and local project folders power the automated newsletter demo.
  • Gensim: Topic-modeling and document-similarity library. Suggested for deduplicating and clustering scraped articles.
  • spaCy: Industrial-strength NLP library. Paired with Gensim for spotting trending topics across a corpus.
  • Microsoft Entra ID: Microsoft's identity platform. Where you register an app and grant Mail.Send to script Outlook email.
  • Notion: Workspace tool with an MCP server — cited as a token hog because it loads every tool into context.
  • OpenCode: Open-source coding CLI. No longer works on Claude subscriptions; routed to API-rate billing.
  • gallery-dl: Command-line bulk media downloader supporting many sites. Good for local, non-automated scraping.
  • Apify: Hosted scraping platform. Its ready-made RedNote scraper was rated reliable for influencer data.
  • Cloudflare Turnstile: Cloudflare's CAPTCHA replacement. One of the main walls blocking the founder's scrapers.
  • Granola: AI meeting-notes app. A member uses it in China by sticking it on the back of a phone.
  • Doubao: ByteDance's consumer AI assistant. One member's only AI tool — for casual questions.
← Back to Notes