Loops: the work that comes back every day should stop being yours
Ramp runs agents at every step of its engineering process. The lesson I take from it is not about scale, but about which work literally repeats itself.
Claude designed proteins that worked in a real lab, against 14 of 15 targets
Anthropic had Claude run a protein design campaign on its own. Two outside labs built the designs and measured them: 354 of 1,320 bound. The percentage is not the interesting part.
Delete your CLAUDE.md every six months
Boris Cherny cut 80 percent of Claude Code's system prompt and the model got slightly smarter. What that ablation loop is, and what it does to my own 5,100 words of instructions.
Claude Fable 5.1 is out, and the saving lives in the cache
Anthropic shipped Claude Fable 5.1 and Mythos 5.1 today. Token prices hold steady, cache reads get four times cheaper, and your agent will behave differently without a single line of code changing.
Why an app that starts hidden starts slow
Anthropic says Claude Desktop now starts about twice as fast. The cause teaches more than the number: the browser engine had it in power-saving mode while the window stayed hidden.
Let the second agent break what the first one built
Nate Herk built an AI product in a single day and put two coding agents against each other: one builds, the other tries to break it. That habit is worth stealing.
Dynamic workflows: when a fleet of agents earns its keep, and when it just burns money
Anthropic published a cookbook on dynamic workflows in the Agent SDK. What they are, what you gain, where they go wrong, and the rule of thumb I take from it.
Blind tasting against the machine: how to test it fairly
A model does not taste. But how much of tasting actually lives in language? Here is the protocol that measures it, with six glasses and one prompt line that makes cheating impossible.
Artboards in your terminal: what Claude Code's /design command actually does
Claude Code got a /design skill in research preview: editable artboards in the CLI and the desktop app. The more interesting half is not in the announcement but in the plan boundary underneath it.
The AI tools I make content with, and what broke in each one
Nine tools from my own pipeline, with what each costs, what it delivers and where I hit a wall. Including the two I can't make a recommendation about.
Claude Code weekly limits stay 50% higher, now through August 31
Anthropic extended the higher weekly limit for the fourth time, and said for the first time that it would like to make it permanent. The same sentence carries a warning about capacity.
Your Claude Code sessions can now reach each other
Two sessions in the same repo knew nothing about each other. Since version 2.1.224 one can send the other a message. What it fixes, what it doesn't, and where that message actually travels.
The hurdle was git, not the AI
Fifteen non-programmers at ABC Legal built fifty agents in a month. What made that safe was four governance patterns, and they work without the product they came in.
Your wine rack as a data file: what should we open tonight?
Twelve lines in a text file and your AI knows what you have. The question stops being what to buy and becomes what to drink before the window closes.
58,000 stars for a search engine that reads people, not pages
The /last30days skill searches Reddit, X, YouTube and Polymarket instead of indexed web pages. That produces a different kind of answer, and a different kind of bias.
Auto mode becomes the default in Claude Code
From 14 August, new sessions on Pro, Max and Team start in auto mode. What the classifier actually judges, the numbers Anthropic put behind it, and why I would still configure it myself.
Your browser session now follows you. Your recordings stay behind.
The Claude side panel in Chrome now runs a Cowork session, so what you start in a tab you can pick up on your phone. Something also disappears: the recorded workflow.
Test your AI on your own field: twenty questions and an answer key
A model that sounds certain is not the same as a model that is right. Here is the test I built for wine, answer key included, so you can run it on whatever you happen to know best.
Two guardrails in Claude Code quietly changed. One closed, one opened.
Release 2.1.228 holds two lines that touch your habits: synced skills are fenced in, and the read-before-write rule now applies only to older models.
Claude Code, Cowork or Dispatch: which one for what?
Three environments from the same vendor, and they overlap less than you'd think. What I use each for, what they cost, and where I hit a wall with all three.
Claude is going to watermark its text. That proves something else than you think.
Anthropic signed the EU AI Act transparency code and will mark Claude output with a watermark and C2PA metadata. What it measures is which machine touched it, not who thought of it.
Build an AI project that knows your taste in wine
Single prompts forget who you are. Ten questions give you a taste profile you paste into a Claude project once, and never have to explain again.
Blotato now builds your DM automations too
Blotato shipped DM automations for Instagram and Facebook. You create them by talking to Claude or ChatGPT, and they keep running 24/7 on their side.
From 200 issues to almost zero: how Astro put AI agents to work
Astro pointed four isolated AI subagents at its own bug reports. Open issues went from 200+ to about 30. Here is how the pipeline works, and what is worth borrowing.
Claude Wine Mode: not real, but it works
A viral post claims Claude has a hidden Wine Mode. It doesn't. The six prompts do work, and from the wine side I can see where they fall short.
Dictation that never leaves my laptop
After my piece on voice-first I got a question: do you do that outside your AI too? I do. With FluidVoice, a dictation app that runs entirely on my own Mac. Including the AI cleanup.
Accessible design with Claude: why it's there by default
Accessibility has been legally required for many businesses since 2025. What WCAG 2.1 AA means, why Claude includes it by default, and how to get your own site checked.
MCP just got its biggest update yet: what it means for you
MCP is the standard every Claude connector runs on. A new version shipped on 28 July, the biggest since launch. Here is what actually changes in daily use.
Claude Design: from idea to working prototype in minutes
Claude Design turns a plain description into a real page with working code. What it is, who it helps, and how it differs from Figma or Canva.
Four things that break when you move to Opus 5
For anyone building on the API. One change throws a hard error, the others quietly change your bill. Plus the instructions you should be deleting.
The Opus 5 benchmarks, and the four it loses
Anthropic publishes fourteen measurements across twelve benchmarks. Opus 5 wins most of them, but the four it loses say more, and one number puts the whole table in perspective.
Effort is the dial that matters in Opus 5
Five settings, from low to max. The price per token has not changed, so this is where your bill gets decided. And the surprise sits at the bottom end.
I barely type to my AI anymore
I dictate almost everything, this article included. Allen Pike puts a name to my hunch: voice in, visuals out. Plus a few honest limits.
Claude Opus 5 is here: what changes for you today
Anthropic shipped Claude Opus 5 today. Same price as 4.8, close to Fable 5. What that means in practice, whether you just chat, build with it, or pay for it.
Claude Sonnet 5's system card: what Anthropic tests, and admits
Every model ships with a system card: pages on safety, behavior, and welfare. Here's what stands out in Sonnet 5's: a model that criticizes its own constitution.
A Chatbot Answers. An Agent Just Does It.
Agentic AI is the buzzword of 2026. Behind it sits a simple shift: the AI stops giving you answers and starts taking action. And that's exactly where the catch is.
Claude Cowork now runs without my laptop
Cowork leaves the desk. Scheduled tasks now run in the cloud, even when your computer is off. What was actually announced, how it changes my day, and where I stay careful.
Claude Fable 5: you're the bottleneck
At the AI Engineer World's Fair, Anthropic engineer Thariq Shihipar explained why you, not the model, are now the constraint. What Claude Fable 5 changes about how you work.
The plugin that figures out your Claude Code setup for you
Anthropic shipped an official plugin that scans your codebase and recommends which hooks, skills, MCP servers and subagents to turn on. It only reads. It changes nothing.
Dispatch: starting a Claude Code session from your phone
Dispatch can now start Claude Code sessions too, not just Cowork. You assign a task from your phone and pick the session up in the Code tab with Remote Control.
How Spotify lets Claude loose on 20 million lines of code
Spotify VP Engineering Niklas Gustavsson explains Honk, their agent fleet for code migrations, and why 73% of pull requests are now AI-authored.
Effort: the dial that decides how hard Claude thinks
Anthropic added a new dial to the Claude API: effort. You decide how deeply the model thinks, from a quick lookup to the heaviest coding job. Here is what it does and when to pick which level.
Karpathy calls Claude Tag the third redesign of LLM UI/UX
Andrej Karpathy says Claude as a Slack teammate is a new paradigm, after website and app. I read the thread and push back a little.
Claude Sonnet 5 is here: close to Opus, for less money
Anthropic shipped Claude Sonnet 5 today. It works more independently than Sonnet 4.6, lands close to Opus 4.8 on performance, and costs a good deal less. Here is what changes.
I do half my work from my phone
One of the makers of Claude Code starts half his agents from his phone, from the couch, coffee in hand. Not out of laziness, but because the work has changed. On working without sitting at your desk.
The bug was already fixed before I started
A colleague at Claude Code set up a routine that picks up bug reports on its own, posts a fix and pings the pull request. More and more often the work is done before anyone begins.
I dropped Plan mode. Now I do everything in Auto mode.
Boris, one of the makers of Claude Code, no longer uses Plan mode. Only Auto mode. I work the same way now. On a habit that nothing needs anymore, and the one reason I sometimes keep it.
Claude Code gets Artifacts: share a working page from your session
Anthropic brings Artifacts to Claude Code. Build an interactive page from your session and share it with your team through a private link. Now in beta on Team and Enterprise.
Claude designs with me now, and it talks to my code
Claude Design went into beta today on every paid plan. I tried it right away and it works great: more designing, and a smooth handoff to Claude Code.
Builder Day in San Francisco: what seven teams built in a single day
Anthropic ran a Builder Day in San Francisco. Robert Scoble posted the seven best teams. I watched their presentations and pulled out the insights and lessons.
Fable 5 pulled: the government withdrew the strongest model
Three days after launch, the US government suspended Claude Fable 5. What happened, why Anthropic disagrees, and what it means when you build your work on a model like this.
How Hard the Machine Thinks
There is now a dial for how deeply the machine thinks. Since Opus 4.8 it sits on high by default. How I decide when to open it up and when to leave it alone.
Test-driving Claude Fable 5: an afternoon of real work, not a demo
I handed Claude Fable 5 my full content production for one afternoon: two campaigns across eight systems. How the new model feels, what it costs, and what stayed human.
Claude Fable 5: the strongest model you can use today
Anthropic shipped Claude Fable 5 today, the first publicly available Mythos model. Here is what it is, why it feels different from a normal release, and what it changes about how I work.
The prompting playbook: let Claude write your prompts
The best prompt is not one you write yourself. You let the model draft it, you test, and you refine. What that changed about how I build with Claude.
What's New in Claude Opus 4.8
Anthropic shipped Claude Opus 4.8. I use these models every day in sessions and build with them myself. Here is what stands out, and which options I actually turn on.
Claude Cowork: a month of double hours
Through July 5, Anthropic doubles the 5-hour usage limit in Claude Cowork. No button, no request. Here is what it means, and how I use Cowork every day.
One word spun up a fleet of AI agents. Now you have to say it on purpose.
A single word quietly launched a swarm of subagents in Claude Code. Anthropic just made that trigger explicit. Small patch, large principle.
Eight Times the Code, and What It Doesn't Say
This week Anthropic put a number on something I had felt: eight times the output. The number is probably right. The question is what it measures.
What Claude Can Do Now That It Couldn't Last Year
Not a spec sheet, but how I actually use the newest capabilities: connectors, background work, skills, and a homegrown image pipeline. And where I still keep my own hand on it.
The Assistant That Executes
As long as AI lives in a chat window, the advice stays at a distance. Now that remote connections are on by default, the assistant moves into your tools. That is not a feature, it is a shift.
AI Psychosis and the Last Mile
A tech CEO called his peers prone to AI psychosis. He has a point, and it comes down to the distance between the boardroom and the actual work.
The Synthetic Sommelier
What does it mean when AI recommends wines better than a human expert can? An exploration of taste, algorithms, and what we lose along the way.
Why AI Strategy Is Always About People
Organisations pour millions into AI projects that fail. Not because the technology falls short, but because they forget the people.