What I Watched
October 5 to October 12, 2026
The cheapest software this week was the kind people built around their own business.
Rent Less, Own More
This was a heavy football week in my history, which the filter politely ignored. Of 126 videos, 22 made it through as genuinely professional. What surprised me was how many of those 22 were receipts. Not launches, not predictions. People holding up a bill and showing what they replaced it with.
An invoicing app. A web scraper. Eight websites. An SEO suite. Each story had the same shape: a subscription that stopped fitting, a weekend with Claude Code, and a monthly cost that dropped from three figures to roughly the price of a coffee.
Around that core, the builders were busy figuring out how to make the things they build reliable. How skills actually get read. How a team of agents reviews its own work. Who picks the model when your assistant runs all day. And underneath it all, one uncomfortable question about what happens when everyone knows what AI can do.
Stop Renting What You Could Own
Andy O'Neil's "FreshBooks Wanted $70 More a Month. So I Built My Own." is the cleanest version of this story. FreshBooks moved card payments from Stripe to WePay, and same day payouts suddenly required the $140 plan instead of the $70 one. He tried it for three months. Month one paid for itself. Months two and three did not. His line stuck with me: that is not a feature, that is a toll booth. So he moved invoicing and time tracking into the Chrome extension that already runs his CRM, projects, and calendar. Invoices now draft in his own inbox and go out from him, not from a system.
Creator Magic's "Claude Code Built A Web Scraper: No More $249 SaaS" did the same thing for web data. A polite scraper that reads robots.txt, built step by step from an agents.md file. From a data center server it hit 33 percent success. Routed through residential proxies it climbed past 90 percent, pulled regional pricing for 100 AI tools, and added sentiment scores with Haiku. Total bill for proxies, search, and AI: about two dollars, against $249 a month for the service it replaced.
DAIN's "I Fired GoDaddy. Claude Runs My 8 Sites for $1/Month" brought the receipts too. $1,350 a year in hosting and add ons became 49 cents in July and $1.01 in October across eight sites, on Cloudflare and Firebase. Claude also found a $720 GoDaddy charge he had never noticed and a stale DNS record a stranger could have used to spoof his email. The Next New Thing's "Top 10 Repos explained" rounded it out with OpenSEO, an open source take on the $129 a month SEO suites that runs on pennies of data. None of these people hate software. They just stopped paying for software that no longer fit.
When a tool stops fitting how you work, you pay for it three ways: money, time, and workarounds.
Skills Get Read Like A Contents Page
Jay E's "Anthropic Just Revealed 10 NEW Rules for Claude Skills" walked through Anthropic's updated skills guide, and one detail reframed everything: depending on structure, Claude may only preview the first 100 lines of a reference file. If your key rule sits on line 140, as far as the model is concerned it does not exist. Charlie Automates made the same point in "#1 Way to 20x Your Claude Skills" with the simplest fix I saw all week: put a table of contents at the top of every long reference file so the model knows where to jump.
The rest of the guide reads like good management advice. Keep the main skill under 500 lines and one level deep. Match the degrees of freedom to the risk, so writing the invoice email is wide open while generating the invoice is an exact script. Give multi step jobs a checklist with a go back line. And the one I would underline: if a rule must never break, like never sending an invoice over $10,000 without sign off, move it out of the skill and into a hook. Capital letters are a suggestion. A hook runs every time.
Austin Marchese's "Anthropic Just Released 13 New Skills To Level Up Your Claude Setup" added the vocabulary that ties it together. Utility skills do one job. Orchestration skills route work to the right utility skill, like a contractor sequencing the plumber before the tile guy. His summary of Anthropic's own pattern was the best line of the week: automate the prep, not the decision. The ECC skill pack covered by The Next New Thing applies the same idea to code, with planning and test driven development baked in so the tests exist before the feature does.
Claude may only read the top of a file, so put the map first and move unbreakable rules into hooks.
The Repo Is Free, The Factory Is The Product
Nate Herk's "I Tried 100+ Github Repos for Claude. These 6 Are the Best" was a practical list: Hyperframes for video, the Google Workspace CLI, Apify for data, a humanizer for copy, Addy Osmani's Agent Skills for spec, plan, and build, and Playwright to click through an app like a customer before anyone else does. But the line I wrote down came at the end. No one is going to pay you for a GitHub repo. Those are free. What you sell is the thing you build around it that solves their problem.
ZazenCodes' "Build Your Own AI Software Factory with Claude Code" showed what building around it looks like. A feature request goes to a spec writer on Opus, then a builder, then four separate reviewers for security, UX, UI, and code. Findings loop back to the builder for at most two rounds. An approver either merges to main or escalates to a human, and anything touching billing always escalates. The whole thing runs from markdown files, skills, and subagents inside an existing project, with a dashboard so you can see what each agent is doing.
The platforms are moving the same way. The Claude Code v2.1.290 changelog added a managed agents pattern with templates like a deep researcher. DIY Smart Code's "Is the new Claude app a game changer or a mess?" covered Claude merging chat and Cowork into one app and rebuilding Projects in Claude Code so a coordinator splits a goal into threads, each a full session on its own branch with shared memory. The building blocks are commodities now. The arrangement is the value.
Nobody pays for the free repo; they pay for the reviewed, tested system you assemble around their problem.
Who Picks The Model When It Runs All Day
Always on assistants had a big week. Matthew Berman's "OpenAI COOKED" covered Dots, OpenAI's answer to Grok Bot: an agent that lives in the cloud, connects to your accounts, and tries to get things done before you ask. Nate Herk's "Grok Bot Just Got 2 Massive Upgrades" covered the other side. Grok Bot will now route each request to whatever model it thinks fits, small and fast for simple questions, Opus 5.5 for hard ones, and it gained free native X search. Nate switched his bots off the paid X API on the spot, about five dollars a week saved.
His follow up, "Grok Bot uses my Claude Code subscription," was a two minute trick with a bigger idea behind it. The bots share a cloud computer with a terminal, so you can install Claude Code there, sign in with your own plan, clone your skills repo, and carry your whole operating system in your pocket. It works because his setup lives in files and API keys, not in one vendor's desktop plugins.
The sharpest point was about routing. If the assistant picks one model per prompt, a mega skill that does seven things gets the expensive model for all seven, even when four of them are simple. Breaking big skills into smaller ones lets the cheap model handle the easy parts and saves the heavy model for judgment and taste. Berman's numbers show why it matters: GPT 6.1 Soul at $2 in and $10 out per million tokens performs close to Astra at $10 and $50, while Ultrafast can burn a thousand dollars in ninety minutes.
When the assistant chooses the model, smaller skills are how you keep the bill small.
The Knowledge Gap Is Closing
Nick Saraev's "Here's What I'd Learn Instead of AI Automation in 2027" was the most uncomfortable video of the week, and probably the most useful. His argument has two parts. Automation benchmarks that started near zero are being saturated in months, so agents can increasingly automate a business without a human pointing them at every endpoint. And the knowledge gap that let automation builders charge a premium is closing as more people use capable AI directly. His prediction: 2027 may be the most profitable year AI automation ever has, and also the start of its decline as a pure knowledge play.
He compared it to web design. Anyone can build a website now, yet some designers still close six figure deals, because they are extraordinary at sales and marketing or at the product itself. The models themselves keep getting cheaper to own, too. Better Stack's "AI Models Might Not Need Tokens Anymore!" covered a Meta paper where small byte level models, distilled from a Llama 3 teacher, overtake token models the longer they train and need about a sixth of the data to match them. The Next New Thing featured Mini Mind, a free recipe for training your own tiny language model for about the price of a coffee.
Put those together and the direction is clear. Knowing that AI can automate your email is not worth much when anyone can ask Claude how. What stays valuable is knowing which problem is worth solving in a specific business, building it so it fits, and standing behind the result.
When everyone knows what AI can do, you get paid for fit and outcomes, not for knowing.
Where It All Converges
Every thread this week pointed at the same shift. The parts are free or close to it: open repos, cheap models, hosting at pennies, skills anyone can install. What costs money is software that does not fit, and what earns money is the system that does. Andy's invoicing, the drone scraper, eight sites for a dollar, a factory of reviewing agents, smaller skills for smarter routing.
The knowledge gap Nick Saraev described is closing because the tools are this accessible. That is not bad news for anyone willing to do the fitting. It just moves the value from knowing what AI can do to building the specific thing a specific business needs, with a human still approving anything that touches the money.
Notable Videos This Week
What This Means for My Content
Twenty two useful videos out of 126 is a reminder that my history is mostly football. But the useful ones were unusually concrete. Three of them came with actual invoices on screen, and that is the kind of evidence a business owner can act on.
This is the work Acceleration Works exists to do, so I want more of it in my own feed and in what I publish: fewer tool tours, more before and after numbers from people who replaced something that did not fit.
Rent the commodity. Own the part that makes your business yours.
Own the part that makes your business yours.