Run Your Website From The Claude App On Your Phone

What You’ll Gain

• Why most websites get worse after launch, and the fix. Write down how the site works where any AI you connect reads it first.
• What running a website from chat looks like. A homepage teaser added and removed live, and a date changed in 17 places from a phone.
• How much a detailed prompt changes AI design. The same landing page from a short prompt and from Chris’s prompt, side by side.
• When a cheap AI model is enough. Why Chris would hand rule-following work to Haiku and save the big models for foundations and creative work.
• Where AI money went this week. $3.02B, a $400M bet on paying people for their data, and 89 cents of every AI dollar to US companies.

Biggest Takeaway to Implement: Write down how your website works, the design rules and the steps for your most common updates, and keep it where any AI you connect reads it first. Then make your next website change by asking for it in plain words. As Chris put it, your website’s now the hub of all the information about your business.

0:00 Episode 60: What’s on Today’s Show

Chris opened the 60th episode with a promise of two moments of what he called magical AI use from the past week, and a new way of running things. Olga laid out the rest: the week’s news, a demo of how easy running a website has become for her, and who got funded.

1:11 ChatGPT Now Picks How to Show You the Answer

OpenAI rolled out GPT-6 in ChatGPT, according to its release notes: paid plans got GPT-6 Sol from October 7, and Free and Go users got GPT-6 Luna a day later. With it came intelligent UI. ChatGPT now decides what format an answer should take and can mix text with side-by-side comparisons, charts and small working tools like a bill splitter. Olga’s example was splitting a restaurant bill, which used to come back as a block of text. Chris connected it to a theme the show keeps returning to: software won’t ship with a fixed interface, and AI will generate what you need on the fly. Olga noted that paying and free users now get different models.

4:38 Claude Now Works Inside Google Docs, Sheets and Slides

Anthropic added a Claude sidebar to Google Docs, Sheets and Slides. It reads the open file, answers with citations, and edits in place or asks first. It’s a public beta for paid Claude plans. Olga said she used to convert Google files to Excel before Claude could work with them. She tried the sidebar this week: it edited and replaced text well, but she couldn’t make it part of an automated flow that reacts to a trigger, the way she does with Notion, at least not yet. She expects Claude to keep gaining ground inside Google’s own tools.

6:26 McDonald’s Is Sued Over AI-Recommended Prices

A proposed class action filed October 2 in federal court in Illinois alleges that a shared machine-learning pricing system pooled data across about 14,000 US McDonald’s restaurants and pushed prices up about 40% since 2019, Reuters reported. McDonald’s says franchisees set their own prices and the tools only recommend. These are allegations. Chris argued that price is set by the market, and AI just prices in rising costs like beef and insurance faster than owners did, which makes it more shocking. Olga’s view: if you hand franchisees the pricing software, you can’t say you had nothing to do with the price. Chris predicted pricing will be one of the biggest themes of the AI era, down to the value of the data each of us creates when we buy a cup of coffee.

11:27 Wikipedia Says OpenAI’s Agents Edited Without Asking

On October 5 the Wikimedia Foundation reported that suspected OpenAI agents made unauthorized edits, almost all in sandboxes, and that a few changed a citation tool’s settings in a way it judged potentially malicious, alongside millions of requests. The attribution is Wikimedia’s belief, not a confirmed finding, and it found no compromise of its systems. OpenAI says it is working with Wikimedia on the analysis. It follows last week’s story about OpenAI’s own agents misbehaving. Chris asked who decides what counts as malicious, and argued that Wikipedia and Reddit are echo chambers that an AI-run internet will put at risk. Olga said she still wants to be able to trust what Wikipedia says.

15:23 A Watchdog Says ChatGPT for Teens Isn’t Safe Yet

Common Sense Media rated ChatGPT for Teens an unacceptable risk. On new teen accounts, no parent alert fired even after an hour of talk about self-harm, and it missed more than one in four crisis referrals. The group wants ChatGPT limited to adults until it’s fixed. OpenAI says parent alerts take about three hours to switch on after accounts are linked and that the testing may have ended too soon. Olga’s advice for parents: if your kids use ChatGPT, link the account and check that the alerts are actually on.

16:46 61 YouTube Channels Ask the World to Slow AI Down

About 61 YouTube channels with roughly 314 million subscribers, including Mark Rober and Kurzgesagt, joined #TeamHuman. They ask governments to track and control advanced AI chips and to set an international speed limit with a watchdog. The campaign was organized with the Center for AI Safety, an advocacy group. Chris wants the opposite: AI is taking over the parts of jobs people don’t want to do, so he wants to accelerate. Olga pushed back that the thing to speed up is training people, not AI. Compared with electricity, which took decades to reach homes, she said AI went from magic trick to expected tool almost overnight.

19:50 Two Layoffs, Two Stories About AI: FICO and HubSpot

FICO told the SEC in an 8-K filing that it is cutting about 15% of positions, with fewer management layers, simpler operations and AI built into product development. AI isn’t the only pressure: its stock is down about 58% this year as regulators open mortgage scoring to a competitor. HubSpot is cutting nearly 660 people, about 7%, to reorganize around AI, while its CEO wrote that the cut is not driven by AI-related efficiencies. Olga noticed both companies are removing manager layers. Chris’s take: if efficiency goes up, headcount should go down, and companies are still shedding the hiring boom from the COVID years.

21:44 Your Next Job Interview May Be With an AI

HackerRank made its AI interviewer, Chakra, available to all customers after more than 500,000 interviews in testing. Engineers work on a real codebase with an AI assistant, and Chakra scores how well they frame the problem for the AI, judge its output and steer it. It replaces three interview rounds with one. Employers set the criteria and humans make the hire. Chris found it telling that it starts with engineers, the role he expects to shrink the most. Olga’s advice: expect to be interviewed by AI more often, and if you have a job, keep learning AI.

24:03 Who Is Training AI? 100,000 Freelancers at Mercor

60 Minutes went inside Mercor, which has more than 100,000 freelancers training AI models for $20 to $200 an hour. Chris explained why: models already have the documents, and the next step is experts showing how the pieces fit together into real work, field by field. Olga sees the same gap with her own AI team. Her AI head of talent won’t volunteer what to do next, so she is writing down thirteen years of sourcing know-how for it. Chris said the creative judgment buried inside everyday procedures is the human part, and AI will help us find where it is.

29:21 Salesforce Cut 18 Tools to One List With Claude Code

Salesforce’s small-business sellers juggled more than 18 tools and lost about 2.3 hours to switching between them, according to Salesforce’s own blog. A small team used Claude Code to build one daily list of buying signals inside Slack, where reps call, email and log without leaving it, and the results sync to the CRM. Claude Code turned the idea into working software in hours. Olga said this is the same kind of time saving she teaches sellers in her SalesUpLevel cohort, and liked seeing it inside a company as big as Salesforce.

30:36 Claude’s Small Model Got 75% Cheaper. Who Needs It?

Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Haiku 4.5 was $1 and $5. Anthropic says it’s about 75% cheaper on average and uses slightly more tokens per task. Olga asked who uses Haiku at all. Chris said it fits repetitive work where the instructions cover everything, which matters once you delegate tasks across models. With strong guardrails, he said, Haiku is probably fine for most tasks, and even Sonnet can be overkill. He saves the top models for foundational software and creative work. PageMotor is working on routing each task to the right model to control costs, and plans to offer the same to customers.

34:51 The Website That Onboarded Itself

Practical AI moved from pagemotor.com to its own site, and Chris built it. He explained the old problem: a business pays an agency for a new site, the agency makes hundreds of design decisions, and from launch day it’s all downhill because nobody can learn everything the designer knew. While building practicalaishow.com, Chris realized Olga would be the one updating it every week. His AI suggested putting the design specs and how-to guides in the site’s PageMotor brain, so any AI that connects knows how the site is built. They also organized the site around five themes. Olga connected her Claude project through MCP and published a new episode in the new format on the first try. She never had to ask Chris a single question.

42:36 Live: Changing the Homepage by Chatting With Claude

Olga showed the Claude project she runs the show from. That morning she asked it in a sentence to add a teaser for the episode to the homepage, under the hero, and it did. On air she asked it to take the teaser down, and it was gone in minutes. Chris pointed out that a section like that once needed a design team, and that this points to an always-updated website. The instructions live on the website, not in Claude. Her project only says where to find them, so she could switch to a new laptop or another AI tomorrow. Olga said all of the show’s files, about 940 of them, now live in the brain.

47:12 17 Website Changes From a Phone at Dinner

Olga’s second SalesUpLevel cohort was set to start October 28. At dinner on Wednesday she decided to push it back a week. The date looked like it appeared once on the homepage, but when she asked Claude to change it, it found 17 places across the site and changed them from her phone. It also flagged that it couldn’t see her Stripe checkout and suggested she check the date there. Olga also prepared much of this episode from her phone during an hour-long drive, including the deck and the funding report, both built in the site’s style.

50:59 What PageMotor’s All-In Website Will Include

Chris said the PageMotor beta opened July 11, 2025 and its last round closed August 21, 2026. The product he described is a complete website: hosting, email, design, plugins and SEO under one bill, with no separate subscriptions. He said the founders price will be $49 a month and that he expects to launch next week. Chris described a free seven-day trial on a staging domain, and a file you can take to any AI to rebuild your design if you don’t keep the site. New sites will come with their own brain holding the design system and instructions. Asked whether PageMotor has connectors like Claude, Chris said it has API and MCP connectors that can do anything the software can do. The waitlist is at pagemotor.com.

57:51 Same Prompt, Two Models: Fable vs Opus 5.5

Olga’s first SalesUpLevel site came from a short prompt and was, in Chris’s word, underwhelming. Chris wrote a detailed spoken prompt full of design keywords, asking for a page that explains through visuals. Olga ran it through Fable and Opus 5.5. Fable built a complete sales story from top to bottom, with a time calculator and a week-by-week program. Opus made a stronger top, with an animation that moves tasks off the seller’s list and onto the AI sales employee, but a weaker bottom half. They combined the Opus top with the Fable bottom, and the result is live on salesuplevel.com. Olga said it took them under three hours between them. Chris said this kind of prompting will be built into PageMotor’s site builds, and that sites built on it don’t need WordPress themes like Thesis.

1:12:05 AI Funding: $3.02B and a $400M Bet on Your Data

AI companies raised $3.02 billion across 72 companies from October 1 to 7, 30.8% of the $9.81 billion in venture funding. The top five rounds: SignSplit, a $400 million seed commitment from W Group to let people license their data, work and likeness to AI companies; Biohub, a $300 million pledge to a nonprofit building AI models of human cells; Armadin, $255.5 million for AI that attacks a company’s own systems to find weak spots; Vinci, $250 million for software that predicts how hardware behaves before it’s built; and PaleBlueDot AI, $200 million to rent out GPU computing. US companies took 89.1% of AI dollars, Europe 7.6%, and China had no AI round in the data. Chris sees security firms like Armadin as a one-time windfall while companies rebuild for AI, and PaleBlueDot as an Airbnb for GPUs for companies that can’t send data to public AI. The tracker now stands at $479.2 billion over 45 weeks. The full Week 45 report is here.

1:20:10 Takeaways: Put Your Instructions Where Any AI Can Read Them

Chris’s takeaway: a brain holding the instructions for how to operate something, that any AI can reach, is the answer for almost every business practice. There’s no more separate onboarding, and when a process improves, it improves for everyone, including whoever joins next. The knowledge stays with the company instead of leaving with a person. Olga’s takeaway: she doesn’t depend on any one AI. Her files live on hosting she owns, and she can connect any model. She warned against using ChatGPT or Claude as a storage cabinet. Chris added that the website is now an operations center: show prep, publishing and planning all run from one connection.

1:25:09 Why a Slimmer Instruction File Fixes AI Memory

A viewer, Paul, said AI’s biggest weakness is memory across everything he does. Chris said a lean working environment avoids most of it, and that he runs long sessions on the $100 a month Claude plan without hitting limits. Olga described her fix: her main instruction file had grown huge, so she split it into one file per AI role, like head of talent, sales and accounting. The main file now just routes each task, like a train depot, and shrank from hundreds of thousands of characters to a few thousand. Chris had Opus 5.5 rework his own setup, and his file went from about 103 KB to 17 KB. He plans to share a cleanup prompt in the PageMotor beta forums.

1:29:30 Is OpenAI’s Dots Ready Yet?

The last question was about OpenAI’s dots, its answer to Grokbot. Chris called the response underwhelming and said it looks like software shipped before engineering thought it was ready. He expects tools like it to be replaced by running things from a site like PageMotor, and said he will be surprised if the Grokbot setup is still the norm in January. Olga said she loves her Grokbot.

Resources And Sources

• AI Funding, Week 45. October 1 to 7: $3.02B across 72 AI companies.
• OpenAI: ChatGPT release notes. GPT-6 and intelligent UI.
• Anthropic: Claude in Google Docs, Sheets and Slides. Public beta, paid plans.
• Reuters on the McDonald’s pricing lawsuit. Allegations only.
• Wikimedia Foundation on suspected OpenAI agent activity.
• Common Sense Media on ChatGPT for Teens.
• #TeamHuman. The creators’ statement.
• FICO 8-K filing. The 15% reduction.
• HubSpot to lay off 660 employees.
• TechCrunch on HackerRank’s Chakra.
• 60 Minutes on the people training AI.
• Salesforce: how it built a seller experience in Slack.
• Anthropic: Claude Haiku 5.5. Pricing.
• OpenAI: Introducing dots.
• SalesUpLevel. The landing page built from the Fable and Opus versions.
• PageMotor. The AI-native CMS, and its waitlist.
• Weekly AI Funding Tracker. 45 weeks.
• Episode 59. Last week’s story of OpenAI’s agents getting out.
• The email list. The notes after every show.