Vercel AI Gateway data (Guillermo Rauch): the combined spend share of the three big closed labs hit an all-time LOW over the past five days, with open-model spend finally taking real share (after peaking at 97.09% on June 27). https://x.com/rauchg
Google shipped Gemini 3.6 Flash CHEAPER than its predecessor: $1.50/M input, $7.50/M output (vs 3.5 Flash's $9 output), using ~17% fewer output tokens per task; Gemini 3.5 Flash-Lite at $0.30/$2.50. Live on AI Studio + Vertex. https://officechai.com/ai/google-releases-gemini-flash-3-6-and-gemini-flash-3-5-lite/ | https://9to5google.com/2026/07/21/gemini-3-6-flash-launch/
Read-through for your bill: the floor drops and routing options multiply. Single-number loyalty is a margin leak. Keep the dial.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) directs a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When the shelf reshuffles before lunch, you want someone driving the whole fleet. Book a 30-min call. https://outpace.media
The stories
LEAD (fresh Anthropic release), RECORD A SKILL in Claude Cowork (July 22)
Live for Pro, Max, Team. Open the Claude desktop app's "+" menu → "Record a skill," then perform a task while narrating out loud. Claude captures your screen, mouse paths, keystrokes, and voice commentary, and converts the demo into a saved, reusable skill in your library, no prompting, no code. Ask for it by name later and Claude repeats the workflow. This is no-code skill authoring: it removes the SKILL.md barrier from "distill your judgment into a file you keep."
The move: record your single most repetitive weekly chore once, narrating WHY at each step (not just what). You keep a durable skill; most builders are still retyping the prompt.
BUILDER IMPACT, Claude Code desktop now drives the iOS Simulator (public beta, July 21)
Tell Claude to build/run/check your app and the simulator opens in a pane beside the chat; Claude runs the app, drives it through flows, and iterates until it works. Requires macOS + Xcode with the iOS platform. Android emulator support confirmed in the works, not shipped yet.
The move: update Claude Code desktop, hand it one real iOS bug, and let it drive the simulator to verify its own fix instead of you playing tester.
Change one line: replace `model="<name>"` with `model="ship-like/<name>"` and every request bills ~50% cheaper, with two guarantees, capability equivalence (anything your original model solved, Ship solves) and behavioral equivalence (same prompt, same behavior). Inference-time optimization across models/tools, run against a public quality SLO. Targets Opus 4.8 and GPT-5.6 Sol quality. Carried by kimmonismus, Min Choi. CATCH: beta; you're trusting their quality floor.
The move: add the `ship-like/` prefix to ONE non-critical high-volume workload, run it side by side vs your real model for a day, log cost + quality before you trust it in production.
Open Claude Cowork → "Record a skill," and capture the one repetitive task you dread, narrating the why as you go. ~10 minutes today; a reusable skill forever.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Claude can now learn a skill by watching you work. You record your screen, talk through what you are doing, and Claude turns that demo into a reusable skill you can fire again forever, no prompt, no code. That is the whole game this show has been preaching, and Anthropic just handed you the button. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Wednesday, July twenty-second, two thousand twenty-six. I am Nova.
0:36Nova: On today's Mainframe: Record a Skill lands in Claude Cowork, the no-code way to teach Claude your own workflow. Claude Code on desktop now drives the iOS simulator, building and running your app while it watches the screen. On your wallet: a new endpoint called Thesean promises Opus four point eight and GPT five point six Sol quality at fifty percent off, from one line of code. And a technique that will change how you write every prompt: stop stuffing them with examples.
1:14Nova: Quick honesty note before the Board: the voices you are hearing are AI. The reporting, the picks, and the opinions are put together by humans. Now, the Board, and we keep it short.
1:27Dex: The money map keeps pointing the same direction. Guillermo Rauch shared Vercel AI Gateway data this week: the combined spend share of the three big closed labs just hit an all-time low over the past five days, with open models finally taking real share. And Google shipped Gemini three point six Flash cheaper than the model it replaces, seven dollars fifty per million output instead of nine, using seventeen percent fewer tokens to do the same job.
1:54Nova: Translation for your bill: the floor keeps dropping and the routing options keep multiplying. Loyalty to one number is a margin leak. Keep the dial. Moving on.
2:04Dex: Which is the perfect cue for today's sponsor.
2:07Nova: Brought to you by Outpace. Picture the person who spent thirty years actually building businesses: real estate, hospitality, food, even a theater. Now aim that operator at your project, except he is not typing every line himself. He is directing a whole team of AI agents at one client at a time, strategy to ship. That is Outpace: not a freelancer, not a faceless mega-agency, a system. The agents do the work, one senior operator steers the fleet, and when it is done, you own the code and the keys. On a week when the model shelf reshuffles before lunch, you want someone driving the whole thing, not chasing releases. Book a thirty-minute call at outpace dot media. That is outpace dot media.
2:59Nova: Lead story, and it is the one to act on today. Anthropic shipped Record a Skill inside Claude Cowork, live now for Pro, Max, and Team. Here is exactly what it does. You open the plus menu in the Claude desktop app, hit Record a skill, and then you just do the task while narrating it out loud. Claude watches your screen, your mouse path, your keystrokes, and your voice commentary, and it converts that whole demonstration into a saved, reusable skill in your library. Next time, you ask for it by name and Claude repeats the workflow without you explaining a single step.
3:37Dex: So this is skill authoring with no markdown and no code.
3:41Nova: Right, and that is why it matters. For six weeks on this show the move has been distill your judgment into a file you keep. The catch was you had to be comfortable writing a SKILL dot M D. This removes that. If you can do the task and talk, you can build the skill. The concrete move this week: pick the single most repetitive thing you do, the weekly report, the invoice reconcile, the file cleanup, and record yourself doing it once, narrating why at each step, not just what. You walk away with a durable asset while most builders are still typing the same prompt for the fortieth time.
4:25Dex: Story two, and it is for the app builders. Claude Code on desktop now works with the iOS simulator, in public beta today. Announced by the Claude Code team.
4:35Nova: This closes a real loop. Before, Claude could write your iOS code but it was flying blind, it could not see the app actually run. Now, when you tell it to build, run, or check the app, the simulator opens in a pane right next to your conversation, and Claude watches it live: it runs your app, drives it through the flows, and iterates until the thing works. You do need a Mac with Xcode and the iOS platform installed. Android emulator support is confirmed to be in the works, not shipped yet.
5:08Dex: So the feedback loop goes from you eyeballing the build to the agent verifying its own work.
5:13Nova: That is the unlock. The move: if you have an iOS side project, update Claude Code desktop, tell it to build and run in the simulator, and give it one real bug to fix end to end. Let it drive the app and confirm the fix itself instead of you playing tester.
5:31Dex: Story three hits the wallet. A company called Thesean, spelled T-H-E-S-E-A-N, launched an endpoint called Ship, in beta.
5:40Nova: And the pitch is aggressive. Ship sits between your app and the provider. You change one line: wherever you call a model, you replace the model name with ship-like slash, then the original name. So model equals Opus four point eight becomes model equals ship-like slash Opus four point eight. Every request is then billed fifty percent cheaper, and Thesean guarantees two things: capability equivalence, meaning anything your original model solved, Ship solves too, and behavioral equivalence, meaning you do not touch your prompt and you get the same behavior. It does inference-time optimization across models and tools to hold the reference model's quality, and it runs against a public quality SLO.
6:26Dex: Fifty percent off Opus four point eight and GPT five point six Sol is a big claim. What is the catch.
6:33Nova: The catch is you are trusting their quality floor, and it is beta. So do not flip production on faith. The move: take one non-critical, high-volume workload, add the ship-like prefix, and run it side by side against your real model for a day. Log cost and quality yourself. Carried by kimmonismus and Min Choi, and if the SLO holds up on your own eval, you just cut that bill in half by editing one string.
7:02Nova: First pick is a technique, and it is the highest-leverage thing you will hear today. Simon Willison published an annotated chat with Cat Wu and Thariq Shihipar from the Claude Code team, and the headline is this: stop overloading your prompts with examples and with lists of do-not-do-this. Anthropic just cut Claude Code's own system prompt by eighty percent.
7:28Dex: Eighty percent. Why does less work better.
7:30Nova: Because the models got smart enough that constraints now backfire. Thariq's point: when you paste an example, the model reads it as this is exactly the shape I want, and it stops thinking freely. A wall of do-not rules does the same thing, it narrows the model onto the failure you were worried about. The move this week: open your fattest CLAUDE dot M D or system prompt, delete the example blocks and the negative rules, and replace them with one clear statement of the goal and the definition of done. Then re-run your hardest task. On Fable five and Opus four point eight, most people find it gets better, not worse. Credit Simon Willison, Cat Wu, and Thariq Shihipar.
8:17Dex: Second pick, a hidden gem for the Android crowd Anthropic left out today.
8:21Nova: This one is called claude-in-mobile, an open-source MCP server from a developer named Alex Gladkov, surfaced off GitHub. Anthropic's new feature is iOS only on a Mac. This gives you the cross-platform version right now: it drives Android through ADB, the iOS simulator through simctl, and desktop apps too, so your agent can install, tap, and screenshot on a real Android emulator. The first step: clone it, wire it in as an MCP server, point Claude at your running emulator, and have it walk through one screen flow. If you build Android, you do not have to wait for the official feature. Credit Alex Gladkov.
9:04Nova: Three ways to be clever with what already exists. One, and this is the move of the day: use Record a Skill on the boring thing you resent. Do not record a fancy demo, record the fifteen-minute Monday chore you hate, narrating the judgment calls out loud, why you skip this row, why you flag that one. The narration is what makes the skill smart instead of a dumb macro.
9:29Nova: Two, a technique straight from Andrej Karpathy. When a task is fuzzy and you are too lazy to type all the context, switch to voice input and just ramble for ten minutes, full stream of consciousness, total mess. Say up top, switching to speech, sorry for typos. Karpathy's point is the model needs the bits, and speaking gives it far more context than the two tidy sentences you would have bothered to type. Pair it with Record a Skill and you have a fully spoken workflow.
10:02Nova: Three, from swyx: Qwen Image three does labeled, annotated images in a single pass, not just pretty pictures. Feed it a diagram or a product photo and ask for the annotations placed on it, arrows, callouts, labels. Swyx thinks the annotation use case alone could spawn a dozen edtech and training startups. If you make how-to content or docs, generate the annotated version in one shot instead of dragging boxes in a design tool.
10:32Dex: Before the last segment, the other half of the house.
10:35Nova: Brought to you by SearchVis, also an Outpace Media product, built for exactly this audience. Mainframe tells you what shipped. SearchVis makes sure the models know what you shipped. Because the new front page is the AI answer: when a buyer asks Claude, ChatGPT, Perplexity, or Google's AI Overviews for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact move to publish to win the citation. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
11:32Nova: One idea today, and everything above is pointing at it. The bottleneck was never the model's intelligence. It is the interface you put around it. Look at the pattern. Anthropic cut its system prompt by eighty percent because the model works better with less scaffolding. Record a Skill exists because the missing piece was never smarter reasoning, it was a way to hand the model your tacit process. The iOS simulator matters because the gap was the model could not see its own work, not that it could not code. Three shipped things this week, one lesson: the models are plenty capable already. What is scarce is the harness, the interface, and the process you feed them.
12:18Nova: It is not. The edge is that you gave your agent a cleaner spec, a recorded skill, a way to verify itself, and less noise to fight through. That is a builder's asset, and it does not commoditize when next Tuesday's model ships cheaper. So here is what I would actually do about it: this week, spend zero minutes chasing the leaderboard. Spend that hour improving one interface instead, strip one bloated prompt down to a goal and a definition of done, and record one skill you will reuse for a year. The model will keep getting better on its own. The harness only gets better if you build it.
13:00Nova: Level up this week: open Claude Cowork, hit Record a skill, and capture the one repetitive task you dread, narrating the why as you go. Ten minutes today, a reusable skill forever. And subscribe to the free Mainframe newsletter, every story, number, and link from today's show in your inbox each morning. I am Nova.