What I Watched
September 29 to October 5, 2026
The best AI setups this week were not the most powerful ones. They were the ones built to fit.
A Week About Settings, Not Models
Twenty nine videos went through my history this week, and twenty one of them were about work. The other eight were split between Ole Miss football and stock market forecasts of imminent doom, a combination that does nothing for my blood pressure and nothing for this report.
OpenAI held Dev Day this week, so you would expect a report full of new models. It mostly is not. The videos that stuck with me were about the knobs: effort levels, speed tiers, skills, planning files, and a brand new way to rewrite Claude Code itself.
The through line: defaults are built for the average user. The wins this week came from people who stopped being average users of their own tools.
Claude Code Mods: Rewrite The Tool
Anthropic shipped mods for Claude Code this week, and DIY Smart Code's short "Claude Code just got a massive update" was the cleanest explanation. Every time Claude Code does something (calling a tool, asking permission, drawing part of the screen) it emits an event. A mod is one small function hooked into that event, and it can run before it, after it, or instead of it. The launch examples were token weather, a context forecast above the prompt; blast radius, which holds a delete and lists exactly what it would remove; and replay theater, which steps through every edit in a turn.
Nate Herk's two videos, "Claude Code Mods Are Game Changers. Set Up These 5 NOW." and "This One Saves Me Money," showed what that looks like in daily use. His cache keeper shows how long until the prompt cache expires and pings him five minutes before it goes cold, so he is not paying full price to reread a long session. A record mode swaps emails and financial figures for placeholders while he films. A collision guard checks whether another chat touched a file in the last 30 minutes before a new session edits it. He built most of them by describing what he wanted.
My take: this is the most interesting release of the week because it changes the question from "what can the tool do" to "what do I want the tool to do." The security caveat matters, though. Mods run with the same access as Claude Code itself and are not sandboxed. Build your own readouts and guards first, read the code before installing a stranger's, and keep hard blocks in permission rules. For the record, this report is produced inside Claude Code, so I will be building a few of these myself.
When the tool can be rewritten in plain English, the best setup is the one built around your own work.
The Default Setting Is A Tax
Chase AI's "Your Claude Settings Are Wrong And Its Costing You Money" walked through a long running agentic benchmark across Sonnet 5, Fable 5, and Opus 5. Going from low to high buys real gains. Going from high to extra high bought roughly one point of performance while Fable's cost per task climbed from $9.18 to about $13. Max added nothing except another $9 per task. His rule of thumb: sit on medium, and bump to high only when the job is genuinely complex.
OpenAI's Dev Day put a price tag on speed. Matthew Berman's "OpenAI COOKED" covered Ultrafast, GPT 6 Astra on Cerebras chips at roughly eight times the speed and six times the cost, alongside a new $500 a month plan. Nate Herk ran it in "I Tested Codex's $500/mo Ultrafast. What You Need to Know." One video editing job finished in 19 minutes instead of 50, and ate 17 percent of his weekly usage instead of 4. On a second test the outputs were basically identical. His verdict: keep it as an emergency lever, and build systems that keep you out of emergencies.
The cheap end is moving even faster. The Next New Thing's "9 things you'll actually do with Jev" showed the decision model sorting 500 emails in seconds. Better Stack's "This $40M Startup Was Replaced After Just 3 Days!" covered a free open source alternative that edged past Jev on a typed decisions benchmark, 0.766 to 0.727, three days after launch. And Charlie Automates' "9 Hottest GitHub Repos For Claude Code" included Magpie, a menu for routing jobs to cheaper models. The lesson is boring and valuable: every dial has a cost curve, and the top of the dial is almost never where the value is.
Max effort and maximum speed are emergency levers, not defaults, and the bill shows it.
Skills Are Corrections You Own
Austin Marchese's breakdown, "Matt Pocock's EXACT System To 10x Your Claude Skills," gave me the best definition of a skill I have heard: a correction you got tired of making. If Claude gets a task right the first time, a skill for it is just bloat that costs tokens. So the useful question at the end of the day is not "what did I use AI for" but "what did I have to send back." Pocock's other trick is leading words. Use a term the model already knows deeply, like refactor, steelman, or premortem, instead of paying in tokens to define your own.
Simon Scrapes' "Everything You Know About Skills IS OUTDATED" covered Anthropic's updated best practices, and a few hit home. Claude often reads only the first 100 lines of a long reference file to decide whether it is useful, so long files need a contents list at the top. Detail should match how fragile the step is: plain instructions for reviewing a sales call, a template for a weekly report, an exact script for anything touching money. And skills written for older models can be too prescriptive for newer ones, so test on the model you will actually run.
Then there was InsiderForce's video on Everything Claude Code, a hackathon winner's setup with 64 agents, 262 skills, and more than 214,000 GitHub stars. Impressive, and exactly what Pocock warns about. His point, paraphrased: most people should write their own skills and own their own processes. Acceleration Works runs on a few dozen skills, and I am now auditing every reference file over 100 lines. A borrowed setup is just someone else's corrections.
Build skills from the mistakes you keep correcting, keep them small, and own them instead of copying someone else's.
Plans Live In Files, Agents Live In The Cloud
Brian Casel's "I stopped using plan mode." opened with the news that Anthropic is considering removing plan mode from Claude Code. Casel stopped using it months ago, yet he plans more than ever. Each product gets a separate meta repository with a cycles folder. Every project is a dated folder with a summary file that agents keep updating as they work, so a new session or a new agent reads one file and is up to speed. He turned the whole convention into a single skill so every agent follows it automatically.
His companion video, "Cloud agents fixed multi-tasking," explained why he dropped git worktrees. Every agent now runs on its own computer in the cloud, using a tool called AMP, with an orchestrator agent spawning and talking to the others. One release had separate agents updating the app, writing the docs, and shipping it to customers. The Next New Thing's "Free on GitHub" featured Paperclip, past 95,000 stars, which gives agents an org chart and task assignments because, as its creator put it, everyone has too many agents and too many skills.
Nate Herk's "I Gave GPT 6 Astra $10,000 to Trade Stocks" was the same pattern in a different costume. A string of scheduled wake ups ran through each trading day, and each one left a handoff note for the next so the system behaved like one continuous agent. It lost to the S&P 500 by less than a percent, which tells you very little about trading and quite a lot about handoffs working. Memory in files beats memory in a chat window.
Put the plan in a file every agent can read, and the work survives any session, machine, or handoff.
Always On Staff For A Company Of One
OpenAI's Dev Day keynote, which I caught through PCMag's "Everything Announced at OpenAI Dev Day 2026 in 15 Minutes," was headlined by Dots: always on agents with their own cloud computer and browser that learn how you work, connect to more than 4,000 apps, and will soon text you updates. Alongside them came ChatGPT Spaces, a shared page where people and agents work together, a Decisions API aimed squarely at Jev, and a marketplace for selling to OpenAI's enterprise customers.
The reviews were mixed. Nate Herk's "I Tested OpenAI's Dots vs. Meta's Muse" found Dots buggy enough that he joked they should have been called bugs, and preferred Muse's interface for non technical users. Matthew Berman noticed that every personal assistant now looks the same: a soft geometric shape with eyeballs. It is the ChatGPT clone era all over again, with better posture.
The more useful video was Nate Herk's "Anthropic's CEO: How to Build a 1 Person Business with Claude." He turned Dario Amodei's comments into three filters: skip capital heavy models, build software, and make sure sales and support can be automated without the experience becoming terrible. That led him to a testing product for AI agents, a narrow niche where 168 subscribers clears $1 million a year. The Next New Thing's "9 shocking repos that make money" made the service version of the argument: install a free tool like PostHog for a client and you become the implementation partner. The staff is getting cheaper. Picking a narrow problem is still the hard part.
Always on agents make a company of one possible, but a narrow, repeatable offer is still what makes it work.
Where It All Converges
Every theme this week came back to ownership. Mods let you rewrite the tool. Skills capture your own corrections. Planning files keep your context out of a vendor's chat window. Effort and speed settings decide what you pay. And the one person business stories only work when the owner is clear about the narrow job the system does.
The practical version for any business: set the defaults on purpose, write down what you keep fixing, keep the plan where every agent can read it, and put guardrails where you can see them. None of it needs the newest model. All of it needs someone who knows how the work should go.
Notable Videos This Week
What This Means for My Content
Twenty one useful videos out of twenty nine is the best ratio in months, which means either the algorithm is learning or I am. The most helpful ones were not launches. They were people showing their settings, their folders, and their receipts.
That is the content I want to make more of: fewer announcements, more configuration. The interesting work in AI right now is happening in the settings menu.
Default settings are built for everyone. Your business is not everyone.
Your business is not everyone.