Alphabet's net income fell 25% while operating profit rose 30%; Figma's PMs run their jobs on 31 shared AI skills
Stratechery's subscriber-only read on an Alphabet quarter where operating income rose 30% and net income fell 25%, plus the 31-skill AI repository that 60 to 70 of Figma's product managers work out of, shown in full for free.
Two items cleared the twenty-four hours to 08:00 this morning, one from each of the two sources this digest follows.
The Stratechery piece sits behind the paywall. It is his Monday Update on Alphabet's quarter, and all that is public is a two-sentence summary and three section titles: Google Earnings, Facebook Déjà Vu, The Agent Play. 12 What follows is built from Alphabet's own results release and the reporting around the earnings call, so you can see the shape of the argument before deciding whether to pay for it.
The Lenny's item is free in full, and it is the one you can act on this morning. It is a Figma product manager walking through the repository his company's product managers run their daily work through, and its chapter list runs to 33 minutes. 3 A weekly bundle published just after midnight packaged it with the October 28 agent benchmark. 4
AI economics and infrastructure
Google Earnings, Facebook Déjà Vu, The Agent Play — Stratechery
Published November 2, 2026, 19:00 Shanghai. Subscriber-only; the summary line and three section titles are public.
- His public line is that a bad-looking quarter counts in Google's favour. "Google's earnings miss looks a lot like Facebook's, and that's a credit to Google. The company is being more aggressive than they get credit for." 1
- The miss sat in advertising, and the shares fell 10% after hours. Revenue of $127.6 billion came in under the $128.2 billion analysts expected, advertising of $82 billion missed a $83.4 billion consensus, and adjusted earnings per share of $2.67 missed $3.02. 5
- The spending is the part that earns the Facebook comparison. Free cash flow was negative $9.1 billion against expectations near break-even, and the finance chief told analysts that 2027 capital spending will rise again from this year's $195 billion to $205 billion. 5
What Alphabet actually reported
Operating income rose 30% to $40.6 billion, on a 31.8% margin. 6 Net income available to common stockholders fell 25%, from $35.0 billion to $26.2 billion. Almost all of that drop sits below the operating line, where other income swung from a $12.8 billion gain to an $8.2 billion expense. 6
| Q3 2026 | Q3 2025 | Change | |
|---|---|---|---|
| Total revenue | $127,577M 6 | $102,346M | +25% |
| Google Search & other | $63,350M 6 | $56,567M | +12% |
| YouTube ads | $11,207M 6 | $10,261M | +9% |
| Google Network | $7,455M 6 | $7,354M | +1% |
| Subscriptions, platforms and devices | $14,290M 6 | $12,870M | +11% |
| Google Cloud | $30,709M 6 | $15,157M | +103% |
| Operating income | $40,576M 6 | $31,228M | +30% |
| Google Cloud operating income | $11,121M 6 | $3,594M | +210% |
| Net income to common | $26,170M 6 | $34,979M | −25% |
Revenue growth of 25% is a thirteenth consecutive quarter of double-digit growth. 6 Search grew 12%, with year-over-year query growth the highest in five years in Sundar Pichai's account. 6 The advertising miss has a cause attached: chief business officer Philipp Schindler put the larger-than-expected deceleration down to more use of AI Mode, the chatbot interface on top of search, and said ads would begin rolling out there in international markets. 5
Cloud is where the money went. Segment revenue more than doubled to $30.7 billion and segment operating income roughly tripled to $11.1 billion. 6 Sales of TPU Systems, Google's own chips plus the hardware and software around them, were a meaningful contributor, and cloud chief Thomas Kurian appeared on the earnings call for the first time to say so. 5 The cloud backlog stands at $561 billion, more than 55% of which Alphabet expects to recognise within 24 months. 5
Then there is the cash. Alphabet issued $28.7 billion of senior unsecured notes in the quarter and sold 36 million shares through its at-the-market programme for $12.4 billion. 6 The finance chief warned that building TPU inventory will press further on free cash flow. 5
Everything past the section titles sits behind the subscription, The Agent Play included. What is visible from outside is the quarter it argues about. 1
Product and growth
How Figma's PMs actually use AI every day | Ezra Mechaber — How I AI
Published November 2, 2026, 21:04 Shanghai. Free in full: the video, the chapter list and the seven things he covers. Hosted by Claire Vo.
- PM OS is a shared repository of 31 skills, used by 60 to 70 Figma product managers. It began as Mechaber's own Claude Code setup and became the thing every product manager at the company installs and customises. 3
- Most of the skills do the work PMs quietly drop. One works out who owns what and adds the right designer and engineering manager as reviewers on a document; another, which he runs about six times a day, collects the Slack threads, open decisions and people waiting on him. 7
- AutoXP runs growth experiments with a single human checkpoint. Three automations find the drop-off, write the copy and open the pull request; he reviews once before launch, then the system reads the result and starts the next round itself. Five experiments are live that he has not looked at. 7
PM OS is a shared repository
The whole thing is a git repo: an agent instruction file, a folder of skills, and a couple of command hooks. One hook checks on boot whether a newer version exists, which is how sixty-odd people stay roughly in sync. The base template is a hidden folder, so nobody edits the shared context by accident. Each person's copy is a private sub-repository, so their write-backs stay theirs. 7
It began as a terminal interface inside Claude Code, because that was the only place it could live. He added a local web front end once it became clear that reading monospace all day is worse than it sounds. The interface now renders proper formatting and works on a phone. 7 These days he works almost entirely in Codex, partly because browser control got good enough to make it the better tool. 7
The origin was a mismatch. He joined Figma expecting everyone to be using these tools, and found almost nobody outside engineering getting real work done with them. So he started automating his own job and a challenge from the chief product officer — make the whole product org fluent — turned it into something other people could install. 7 At a company offsite this spring every PM installed it, used it and was invited to roast it. Some skills were cut afterwards, and he keeps cutting the ones the models have outgrown. 7
He deliberately kept it uninstrumented, so there is no usage data pointing at it. 7
The skills that carry the load
Figma is around 2,000 people, with several org-chart tools that do not talk to each other, so knowing who owns a thing is a real problem. That is the first skill. It fires inside other skills too: when a product manager writes a document, the right reviewer is added without that person remembering who it is. 7
The second is the one he leans on most. Figma's Slack, in his description, is insane, and he archives roughly ten channels a day. The "what's up" skill reads a defined set of channels, follows whole threads, and returns what needs a decision and who is waiting on him. He runs a variant across his live experiments so he has to open fewer dashboards. 7
Beyond those: a skill that builds a self-contained HTML artifact and publishes it to an internal host, one that produces FigJam workshop templates through Figma's MCP server, and a coaching skill that reads the career ladder, the accumulated feedback and whatever problem he is working through. The coaching one exists so he can do his own first pass before taking a question to a manager. 7
AutoXP, and one checkpoint per experiment
His framing is that a growth PM's worst mistake is wasted traffic — a surface running with no experiment on it. AutoXP is his attempt to take himself out of that loop. 7
It runs as three Cursor automations triggered from a Slack channel the system keeps for itself. Someone can point it at a target, or it can go looking. A data-scientist agent finds the drop-off and looks at the funnel and at mix shift; a writing agent proposes copy; an engineering agent picks up a pull request. Every step reports into the same Slack thread, so the experiment reads as a running log. 7
The control is a single review, before launch. He looks at the data science, the hypotheses, the copy and the code together, and then it runs. When the experiment finishes, a webhook triggers the analysis, and the instructions encode his own rules for exploiting a win or exploring again. 7
The results: a large win on the very first experiment, five concurrent live experiments on surfaces the team would never have spent time or traffic on, and one engineer carrying most of it. He is careful about where it stands — in theory where he wants it, in practice almost there. 7
Two more things he built
Bug bashing. He keeps a standing thread with a browser agent, gives it a test account and a FigJam board of scenarios, and has it work through them and write findings back to the board. It found two things he would have missed. A text field broke when a thousand emojis went into it. And a seat-request screen that displays the team name was putting "a" in front of a name starting with a vowel, a bug that depends on the customer's team name and so never showed up in earlier bashes. 7
A writing assistant on Jev. Jev is a fast, cheap model that answers a narrow question with a label or a number. Mechaber calls it every time typing stops, so it flags placeholders and claims that need support, and pulls matching quotes from his library of research interviews while he writes. Speed is the point: at that latency it feels like an editor sitting beside him. 7 The live demo hit Figma's internal firewall. 7
What it changes for you
- If your team keeps rebuilding the same prompts, the reusable unit is a repository. Shared skills, a version check, and a private sub-repo per person is about a day of work.
- If you run growth, the constraint AutoXP removes is the traffic sitting on surfaces nobody had time to test. The review before launch is what keeps that safe.
- If you are budgeting AI time, the skills that paid off here were the low-status ones: who owns this, what did I miss, what did we decide.
- If you are hiring a product manager who claims AI fluency, ask what they have automated that they used to skip.
Where to start
- 02:04 for the origin of PM OS, and 10:28 for the repository structure if you want to copy it. 3
- 06:30 for the chief-of-staff setup, and 19:30 for the bug-bash routine. 3
- 22:08 for AutoXP, which is the part worth watching twice. 3
- 26:58 for the Jev writing assistant, and 33:15 for the closing round. 3
- For the Alphabet numbers, the company's own release is the shortest route and it is free. 6
References
- 1
- 2
- 3How Figma's PMs actually use AI every day | Ezra Mechaber — Lenny's Newsletter
lennysnewsletter.com
- 4
- 5
- 6
- 7
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.