Steve Newman’s fifteen-app stack exists so he only has to look at one screen

Nathan Labenz introduces Steve Newman as the engineer who created Writely, the startup Google acquired and turned into Google Docs, and who now runs the Golden Gate Institute for AI, hosts the Curve conference, and writes the Second Thoughts Substack. Labenz has spent months on his own Claude Code productivity stack and wanted a show-and-tell from someone who has programmed professionally since 1985. Newman is building on the side — evenings and weekends, about fifteen projects — because his day job is the Institute. The theme, he says, is that he has been drowning.

Until he started building the tools, he would not have had the statistic: something like fifty Substack posts, blogs, newsletters, and other long items hit his inbox every day, on top of everyone he follows on Twitter and a cluster of WhatsApp groups. He was spending hours just trying to read, let alone synthesize. The first thing he built, after muttering about it for a long time, was an RSS reader that pulls those feeds and pre-computes two levels of summary. One is a paragraph he can skim in about ten seconds to decide whether to bother; the other is a page that is supposed to surface novel ideas and notable evidence, his 80/20 stand-in for the original.

The summaries are, by design, dumb. He dumps the full text of each post or podcast transcript into a model with a static prompt and no memory of what he has already read. Last year he imagined a system that would contrast each piece against his history; he never bothered, and he has not been motivated since. He already knows what he has read. A ten-second skim is efficient enough, especially on a day when dozens of pieces recap the same Opus 4.7 launch and he only needs the ones with a new angle.

The second theme took longer to crystallize. A few hundred emails, Slack messages, and WhatsApp pings arrive every day. Only a fraction need prompt attention, but a few really do. His old habit was to rotate through about five apps some thirty times a day, every time his brain came up for air — plenty of chances to get snagged by something that could have waited hours. The project that followed is really about five sub-projects: pull Gmail, Slack, Signal, WhatsApp and the rest, hand each message to a model with a one-page rubric accumulated exception by exception, and pop the timely ones onto a second screen.

He had gone forty years of engineering without a second monitor. He finally bought one so it could hold a rolling three-hour calendar and a list of urgent messages. He named the app Radar, after Radar O’Reilly in the old series MASH, the clerk who handed you what you needed before you asked. The goal is an attention firewall: those items, plus whatever he actually wants to focus on, should be the only things he looks at. He was three-quarters of the way through building it before he understood what he was building. There is no playbook, he says. Everyone is fumbling through a Cambrian explosion and only understanding in hindsight.

When he later shares the live view, the calendar is not his full Google Calendar. Blocked focus time is hidden. His wife’s overlapping events are filtered. Recurring internal meetings get three buttons without opening the event: join the Google Meet, open his private notes, open the shared doc. The default forward button sends to his wife, because they share Amazon and PayPal accounts and many notices are really for her. He has no staging environment and pushes to production constantly. The safety net is that every original still lives in Gmail, Slack, or WhatsApp; Radar mostly changes labels. If Claude deletes a database, he loses the interface, not the mail. In a couple of months of working this way, that has not happened.

Plenty of people are letting models draft replies or act on a digital life. Newman has not gone down that road at all. The project looks daunting, the security concerns are real, and he is conservative by nature: he does not like a tool unless he trusts it and understands what it will do. He expects the pool will eventually be too valuable to stay out of. For now he feels like the man in *Raiders of the Lost Ark* waving at the pit: asps, very dangerous, you go first.

Labenz, who calls his old attitude toward computer security borderline negligent, says this technology changed his mind. He has had the same Gmail account for twenty years. Giving an AI access not just to everything he has written, but to everything anyone ever sent him, feels like a duty of care those people never contemplated. He now takes his time on access questions he would once have clicked through. Newman agrees the data is other people’s data, and that the trade-off between security and utility is getting sharper. He keeps waiting for the shoe to drop on agent setups. A couple of juicy anecdotes circulate. Hardly anything follows.

Every new model report card claims another cut in prompt-injection susceptibility, and yet everyone also agrees the whole system is insecure if you trust it with anything important. He does not see how that tension resolves. Labenz asks why AI phishing and election deepfakes have not exploded either. Newman is mostly confused. He retells the early-1980s Tylenol scare: cyanide in a handful of bottles, a few deaths, and safety seals on products to this day — a harm anyone could have inflicted for decades, which happened once. Set that against the old statistic that an unpatched Windows box on the public internet was owned in minutes. He cannot reconcile those two patterns, and he would like someone who can to say so.

The live demo is a row of tabs. The feed reader is almost only one screen: a list of posts, a one-paragraph summary, a longer overview, and an archive button. He once added an icon that opens a Claude session on the article; he had forgotten it was there, because he never uses it. Cloudflare R2 backups and a pretty-printed backup viewer exist because they were a sentence in a prompt, not because he would have built them by hand. A bookmarklet had to rip the HTML of his Substack subscriptions page to import the list. The feature set is much longer than he would have implemented himself, which is one of the interesting things he has found.

A Hacker News tool reads the linked article and every comment, clusters themes, and lets him open one theme to see that slice of the thread — handy when gold and noise are entangled. A Chrome extension from the current page can drop a URL into his reader, his todo app, or a whitelist of Notion note pages; a Hacker News thread sent to the reader is automatically queued for the theme analyzer. His spam folder is resorted by columns rather than date, so a block of mail to `me@aol.com` — not his address — can be skipped as a slab. Another second-monitor widget is an agent status board: blue means a task finished, red means no agent is running but the todo list still has work for that app. Claude Code hooks POST status to a Cloudflare app; Hammerspoon and AppleScript, which Claude told him to install, jump him to the matching terminal tab.

He asked Gemini, ChatGPT, and Claude where a rusty experienced developer should host a suite of one-user web apps, then made them critique one another. Cloudflare won: enough process hosting, cron, queues, and databases to do the job, far less complexity than AWS, cheap at his scale. He split the work into about fifteen GitHub repos and fifteen Cloudflare databases on purpose, a gut call for agent context, more important in January than now. All Claude sessions run in a Docker container on his Mac with permissions skipped. The project directories sit side by side, so agents can read one another’s code and call one another’s APIs instead of relaying messages through him.

A project called Mirror pulls Google Calendar, Contacts, Gmail, WhatsApp, Slack, Twitter, Google Docs, Signal, and SMS. Getting near-real-time sync was far and away the hardest part of the whole stack, eight different solutions, the biggest time sink. Gmail and Calendar have real APIs, painful for Claude rather than for him. WhatsApp was the clever one: WhatsApp Desktop on the Mac stores messages in an unencrypted SQLite file, so a one-minute cron job piggybacks on Meta’s own sync. He will not talk to WhatsApp’s internal APIs; the community is full of ban stories. Twitter is the worst, a shady third-party that forces him to poll about eighty accounts once an hour before fees become visible. Labenz mentions Twitter’s new usage-based official API. Newman says he will switch the next time the current hack breaks.

The todo app is the one database he would be saddest to lose. It writes a backup every five minutes after a change, live-synced to GitHub, so he could open github.com and read a static list if the app glitched. Everything else dumps into R2. Velocity, though, comes from a logging service he built because commercial options were annoying to configure: a SQLite database on Cloudflare that backends, browser JavaScript, and Android apps all write to. His Claude.md is full of loud rules about logging errors and every database mutation.

When a calendar event fails to show, or the posts since noon have no summaries, he types “debug this.” He uses Jesse Vincent’s Superpowers plugin and its systematic-debugging skill. From Claude.md the model knows the logs exist and has been instructed not to guess. It works close to a hundred percent of the time. If he had one concrete piece of advice for personal agentic coding, this would be it: make every moving part, including the frontend, log to one place, then vigorously remind Claude it can go look.

The coding setup is vanilla: macOS Terminal, one tab per project, all Claude. He has the $200 Claude plan and recently added ChatGPT Pro at the same price, mostly for research, including questions like where to meet a friend for lunch on the peninsula. He tried Codex once, a month ago, with a single prompt; it behaved abysmally, which he marks as a useless data point of 0.002. Coding spend still fits in the Claude subscription. He has it buy extra credits in $30 increments, lately two or three times a week, which he attributes mostly to generating summaries with Opus because they are not that expensive.

During a vigorous month or two he spent maybe fifteen to twenty hours a week, mostly weekends. Now it is more like half an hour a day, bouncing between zero and five parallel agents. He went down the “keep the agents fed” rabbit hole — token-maxxing — and found it stressful. Then he realized the agent is not important; he is. Early on that was not true. With one project he sat around waiting for Claude. Once the work was split into fifteen repos, waiting disappeared: he can prepare the next prompt, or skim the summary queue, and he gives Claude work when it is not an interruption to his own mental workflow.

On walks, which he takes for physical and mental space, he does not try to squeeze in extra prompts. He lets a project percolate, then opens Gmail on his phone, addresses a message to himself, and hits the keyboard’s dictation button. At home he pastes the brain dump into a model with instructions to turn it into a Claude Code prompt. He usually does not even read the cleanup. It is faster to let the agent run and correct a misunderstanding than to proof a three-page rewrite. He is sure it is not the best toolchain. It is always there, and it always works.

He almost never reverts, which still surprises him. He did not start vibe-coding until Opus 4.5, so he has only used recent models, plus Superpowers skills that force big decisions past the user. He never looks at the code. All fifteen projects are TypeScript, and he does not know TypeScript. He does watch the high-level choices and steers Claude off a few false paths. These days the work is incremental features, not architecture, which also helps. The hard part is not Cloudflare or Android; it is unlearning the habit that you adapt your workflow to the tools rather than the other way around, and retiring a career of measure-twice-cut-once. Dive in, do it wrong, throw it away, redo it is the better approach now, and it is not his nature.

Labenz’s own light-bulb is that he has built almost no custom UI, only skills, and still hunts files in a terminal to produce the show. Newman’s claim is that the industry has leaned into verbs: throw a command at Claude Code, Cowork, Codex, or Gemini. That is great for a big task, such as writing an app or reformatting 800 PDFs. It is a poor fit for the tiny actions people do every day. An app is the noun side, and it is less explored. His Google Meet Chrome extension clicks the five buttons he otherwise hits every call — switch accounts, hide his own window, acknowledge that video is still sending — and saves about fifteen seconds, three times a day. It took a couple of minutes to write.

Labenz asks how a 2021 acquisition slog would look now. Newman’s last company had built a better storage and query engine; a larger security firm with an endpoint agent on millions of machines wanted it as the new foundation before an IPO. They were swapping the engine under a system ingesting many billions, probably trillions, of events a day. He would not vibe-code that implementation today. In a year or two he wonders. What he would throw at Claude immediately is the investigation around it. No one had a full picture of query shapes, customer size, or lookback windows, and senior engineers were too busy to explain. He would ask for a hundred tools, a hundred investigations, a hundred reports, and a distillation of which reports a human should actually read — none of it touching production.

Asked how he would hire for an Opus 4.7-era software team, Newman offers two guesses. The first is people who can think outside the box, because the box of established practice is gone. Design Patterns, React style guides, and decades of database lore are not useless, but they are no longer the job if you want the new leverage. Nobody has distilled durable principles for vibe-coding a custom UI onto your own workflow. Posts go stale in a week. What makes sense for him will not make sense for Labenz. Comfort navigating without a map matters more than having memorized the old one.

The second guess is communication. What he hears from people running companies is that what used to be a team is now a person, and what used to be three teams is now three people, each running a suite of agents. That means each person is doing a whole team’s worth of coordination with other people. Skill at talking to humans probably overlaps with skill at talking to agents. On the industry’s size, he does not know whether collapsing engineers-per-line and soaring lines of code add up to more jobs or fewer. Jevons paradox could keep a strong track record. The work may look more like full-stack product management than classical software engineering, and it may not always be the same people.

Labenz notes that the demo contained almost no SaaS UI. Newman is still on Gmail, Slack, and WhatsApp; he mostly uses them as backends and spends less time in their interfaces. That sets up a tug of war. Providers want you inside their app — Slack has talked about tightening APIs, Amazon has restricted shopping agents — while users want the best agent, first-party or not. He has no idea which way customers will roll. What will not be vibe-coded for several revolutions yet is the data backbone: S3, Spanner, and probably the next layer up as a place where data lives, even if not as a UI. That is not the next shoe to drop.

Newman’s picture of change is not a single model score. Thresholds look like ChatGPT’s launch or a December cohort that started Claude-coding because Opus 4.5 had shipped and people had holiday time: capabilities had an overhang, then someone did a visible thing. He uses a coronavirus analogy. Viruses with an R of 0.9 in humans jump, infect two or three people, and die out; one day an R of 1.1 appears and the system tips into a new domain. AI impact, he has written, is the product of eight factors — pre-training, post-training, inference compute, agent scaffolding, app design, user aptitude, workflow refactoring, and adoption — all advancing, some rapidly.

Neither of them has used Mythos. Labenz is struck by Nicholas Carlini saying he found more bugs with it in a few weeks than in the rest of a storied career, including working exploits. Newman’s third position, which he admits is getting harder to hold, is that the public fight is a false binary: either the models are amazing and we are basically at AGI, or they are flawed and AGI is distant. His answer is that the models are amazing and we are still far from AGI. Carlini-style performance is maybe 300 points. Being that smart at all the things, in all the human ways, is more like 50,000. Citing Helen Toner, he notes that long timelines are not what they used to be; quite a ways might now mean five years. It is still some distance.

The argument has a soft spot he names himself: we do not ask models to do what they cannot do, so we do not see the holes. His example is one he calls silly — an AI cannot have dinner with his wife for him — and the difficulty of finding better ones makes him worry he is full of it. The gut remains. On recursive self-improvement he has three unanswered questions. What is the full skill list for making models better, including the taste of someone who has designed RL environments since the project that became o1, and is automating AI R&D a year away or hiding hard bones? When coding and math go superhuman, how smoothly does that generalize to marketing, strategy, teaching, and product? And is robotics another Rodney Brooks-scale decades, or does superintelligence plow through? Waymo is an existence proof he trusts more than himself or nearby drivers, but it took decades, still has weather gaps, and a car has about two degrees of freedom — the budget of one finger.

Before he focused on AI sense-making, Newman wrote about climate, including a claim that AI would not be a big deal for emissions. He now thinks he was a little wrong. He did not anticipate how fast data centers would scale, nor that hyperscalers would, as he understands it, back off climate commitments in the land rush for gigawatts. xAI’s Colossus trucked in the quickest, least efficient gas turbines it could find, and he thinks other hyperscalers are grabbing whatever power they can. Short run, AI is using a lot of power, much of it not clean.

He still does not think electricity is the main climate story. The main story is that AI will reshape materials, batteries, and messy petrochemical processes into cleaner electrochemical ones, and that robots might micro-target fertilizer and kill insects instead of spraying. A world where AI raises global electricity emissions 20 percent could still cut more elsewhere. In the long run he doubts we build a terawatt of gas for a terawatt of data centers; the technologies that make sense at scale still look clean to him. In the big picture, AI is going to roll the dice on the whole world, and climate is along for that ride.

Golden Gate Institute for AI, the nonprofit he co-founded last year, is his attempt to help with the collective sense-making that dice-roll requires. Computer science, economics, cyber, bio, labor, politics, and education all hold pieces, and the pockets do not talk. The Curve, their flagship, puts about 350 people from those communities in a room; people who yell on Twitter, or ignore one another, remember the other person is a human. This year’s dates, he spoils, are October 2 through 4. The team is small and growing. A colleague, Taryn, joked they probably need to double the number of Curves every year. A year is a long time in AI.