Posts

The hardest problem in AI writing is not length, speed, or polish. It is the one no model release notes ever mention: making the output sound like you. If you have written long enough to build an audience, you already know what your voice feels like. Sentence length distribution. Where you put a dash. The phrase you reach for every third paragraph. The one word you never use. Strip those out and a technically perfect draft still reads as “someone competent, but not me.” That gap is the entire reason most writers keep a healthy distance from generative AI even as they use it every day. ...

On June 4, 2026, Anthropic shipped three things in 24 hours. None of them is the announcement you’d put on a press release. But read them in sequence, and a roadmap falls out of the noise. The first was an open-source framework called defending-code-reference-harness, dropped on GitHub. It hit Hacker News at #2 with 242 points and 82 comments. [1] The pitch is simple: give Claude a controlled environment to do differential security audits on reference code, and turn “the model finds bugs” from a slogan into a reproducible engineering pipeline. ...

I was up late scrolling X when Guillermo Rauch, the CEO of Vercel, dropped a short thread. No hedging. No build-up. Just a one-word verdict. YES-CODE. An entire category of software, “no-code”, was built under the presumption that code is expensive, difficult, and scarce. Coding agents have forever changed the equation. Code is now cheap, easy, and abundant. [1] Within 24 hours, the same conclusion came back from three people who don’t usually agree on anything. ...

I have been watching GitHub Trending a lot lately, and something is off. By the prevailing narrative, the main battlefield for AI in 2026 is “which model is smarter.” But the repos that keep dominating the trending page are about as sexy as a water meter: Today’s #1 daily is chopratejas/headroom, which claims to compress 60% to 95% of tokens before they enter the LLM. The #1 monthly is colbymchenry/codegraph, a pre-indexed code knowledge graph for Claude Code, Codex, and Hermes Agent, marketed primarily as “fewer tokens.” The #5 monthly is rohitg00/agentmemory, a persistent memory layer for AI coding agents. Over on Product Hunt, the #13 monthly is Tokenwise, billed as a visualization tool for “where your LLM agent bill is going.” And Hacker News has a quiet thread today, “How we index images for RAG”, sitting at 94 points with 14 comments. codegraph alone added 37,000 stars in a single month. That is not a curiosity. That is a collective, very loud signal about where developers are actually spending their money and attention. ...

A lot of AI products look clever but still feel one step away from real work. They can write an email, summarize a meeting, generate a sales script, and produce a polished support reply. But then what? Who actually makes the call? Who follows up with the customer? Who checks whether the email landed in the inbox? Who qualifies the lead? In many workflows, the human still takes over at the important moment. The AI sits beside the work and suggests things. It talks a lot. It does not always touch the business itself. ...

The awkward thing about AI writing is not that it is always bad. Often it is too good at looking like an article. Clean structure. Positive tone. Smooth transitions. A tidy conclusion. The problem is that after three seconds you can smell it. You know the rhythm. Start with a broad statement about a changing world. Add a few balanced sections. End with some version of “the future is full of possibilities.” Every claim is safe. Every sentence is polished. Every paragraph seems desperate to prove that it has a structure. ...

If someone asks me how a normal person should understand Claude Code and other coding agents, I would not start with code. I would start with a much simpler picture: put a pile of PDFs, screenshots, spreadsheets, meeting notes, and transcript files into a folder, then ask the agent to read them, write scripts where needed, and produce an HTML report you can actually open and inspect. That does not sound like “programming.” ...

What caught my eye in Claude Code this time was not the fact that it shipped another command. It was that it finally started treating orchestration as something worth caring about. /workflows looks like a feature toggle on the surface, but it feels more like a boundary line. It moves agents from “helpers that answer questions” toward “execution units that run by rule.” The change is quiet. It is not flashy. But I think it matters. ...

Some tools are not exciting. They are just useful in a way that quietly becomes hard to live without. Zara Zhang recently recommended a Mac utility called Amphetamine. Its job is simple: keep your Mac awake when the lid is closed, while a file is downloading, while an app is running, or during a long task. That sounds almost too small to write about. But I think it points to a very real shift in personal workflows: our laptops are starting to behave less like devices we actively sit in front of, and more like small personal servers. ...

It is tempting to look at today’s coding agents and believe that backend development is about to become a one-shot prompt. I do not buy it. Not because LLM agents cannot write code. They clearly can. They can add endpoints, patch CRUD flows, generate migrations, update tests, and edit multiple files with impressive speed. The real problem is more annoying: the most important parts of a backend system are often not written in the first line of the ticket. ...