David Demaree

bitsandletters.comAgents need a schema, too | Bits&LettersAI agents can now write straight into your CMS. Your content model is their rulebook, and your first line of defense against slop.

AI agents can now write straight into your CMS. 🤯 I wrote for @bitsandletters.com about how content models can help human *and* robot editors on the right track: www.bitsandletters.com/ideas/agent...

Training a bot to take over sorting my email and discovered an unholy eldritch secret lurking in the bowels of the SaaS product I'm trying to replace

I have an app on my phone that emulates the old Game Boy Camera — literally, you load the Camera ROM, and the app hooks it up to the phone camera and Photos library so you can take pictures. Here's my dog in glorious 1-bit:

David Demaree reposted

Andy Bell

@bell.bz

Replying to @bell.bz

People in tech leadership especially, you really need to start measuring the output and determining whether the massive increase in volume is returning anything of actual value. Mix that with your increasingly burnt out workers and there's a real opportunity to get ahead on multiple fronts.

bitsandletters.comNYC-Area Web Design & Development Studio | Bits&LettersBits&Letters is a web and product consultancy in the New York City area. We help startups and businesses get their websites ready for what’s next.

Today's a big day for @bitsandletters.com — I'm pumped to introduce y'all to a new, totally redesigned bitsandletters.com.

But I am delighted with this new site, largely because of how well I think it reflects the original brand work by Trust, my own additions and creative direction, and all those latent decisions that I'd been too timid to let out into the work up to now.

I hope you like the new site and its origin story as much as I do. And if you do, please like/repost this thread if you want, and let me know what you think in the replies:

David Demaree@demar.ee

Today's a big day for @bitsandletters.com — I'm pumped to introduce y'all to a new, totally redesigned bitsandletters.com.

Today's a big day for @bitsandletters.com — I'm pumped to introduce y'all to a new, totally redesigned bitsandletters.com.

bitsandletters.comHow Claude and I rebuilt B&L's website | Bits&LettersI spent most of this summer stuck on a redesign of bitsandletters.com. AI agents got me moving again by sorting out which of our past design decisions still mattered.

As part of the relaunch, I wrote a new essay about the process of building the new site with AI, including some emerging principles like "latent decisions" and treating agentic creation like Layer Tennis: www.bitsandletters.com/ideas/how-c...

My AI journey has been long, and I still would say I contradict myself + contain multitudes on the subject. My essay about using 2.2 billion tokens to make a website also references the "capitalist dream of AI" (derogatory).

Today's a big day for @bitsandletters.com — I'm pumped to introduce y'all to a new, totally redesigned bitsandletters.com.

claude.aiBits&Letters Brand SystemTry out Artifacts created by Claude users

While an LLM was integral to bringing all the pieces together for the new site, I had Claude go back to first principles and the original brand concepts by Trust Design Shop to build and document a new design system that was truer to me, Hannah, and Jesse's original ideas: claude.ai/artifact/8a...

Using that brand system, Claude took Figma wireframes, a fresh content doc, and hours of back-and-forth with me in a single epic megathread to shape a new Astro & Sanity codebase that'd reflect the new, more expressive visuals while keeping/improving UX and SEO.

Today's a big day for @bitsandletters.com — I'm pumped to introduce y'all to a new, totally redesigned bitsandletters.com.

AI — specifically Claude Opus 5.5 — was a big part of designing and building the new site. It's the first personal or work site I've ever made where I wrote _none_ of the code, except maybe bits from prior site builds that Opus reused. (Are there any? I don't even know!)

The year's hottest video game: a management simulator where your junior employee tries to "contribute" by pointing out problems rather than fixing them, and you have to offer real-time feedback.

. @aworkinglibrary.com, @ethanmarcotte.com, and other friends who used to work on Editorially (😭) may be pleased to hear I just used the word 'stet' in a sentence, but less pleased at who the sentence was for (it was an instruction to AI, which to my surprise and delight knew what 'stet' meant)

BTW, here's a glimpse of my Claude Doc-powered writing process. While sometimes I'll `@claude` in the body to ask it to generate some slop for me to rewrite, my instructions are usually "link this mention of _Bird by Bird_ to the page on bookshop.org" or "remind me to change this number at pub time"

emails.bitsandletters.comWe're not going to talk about AI

This week's newsletter explores the confusing space between "SEO slop" — human written but meant for machines — and AI slop, where it's machines all the way down. Does a human touch matter if the goal is to get cited by ChatGPT??? emails.bitsandletters.com/posts/we-re...

youtube.comDefinitely Not Talking AI with David Demaree (ep320)I talk with David Demaree, a software product leader and the person...

In my favorite movie, Nick Cave repeats to himself, “I’m NOT gonna tell you about a girl…” Then he says to the crowd “I’m gonna tell you about a girl” and starts playing. This podcast I was on with @ichris.bsky.social is like that, except replace “a girl” with “AI” youtu.be/Df5iMA9SBao...

bitsandletters.comNo-code, vibe-code, same problems | Bits&LettersDitching Webflow for a vibe-coded site feels like an upgrade, but the hard parts of running a website come along for the ride.

On day one of vibe-coding your company site, you feel like Superman. By day ten, it feels like your keyboard is made of Kryptonite. No-code or vibe-code, you still have to deal with the same problems: www.bitsandletters.com/ideas/no-co...

bitsandletters.comIs it an agent, or just a cron job? | Bits&LettersMany AI "agents" are just scripted workflows wrapped in prompts, burning tokens for no good reason. We explain the difference, and show the cost based on real-world testing.

The other day an AI consulting client asked one of my favorite questions: "is that really an agent, or just a cron job?" I blogged about the difference between the two, and did some actual testing to find out how much the difference costs in time and tokens: www.bitsandletters.com/ideas/is-it...

As I talk to business owners and fellow agency folk about how they're adopting/using AI, a few themes keep coming up: 1) Lots of people are diving in to experiment, relatively few feel like they know what they're doing, but that isn't stopping them from trying.

6) Most of the above applies to a tiny fraction of users/businesses. The vast majority are using ChatGPT on the free tier, or Claude on the cheapest pro plan, mainly just getting answers or typing things out rather than giving AI agents real tasks yet.

An image that comes to mind is, if AI is a pool at a resort, most people are hanging with their kids in the shallow end or lounging poolside, with a few crazies like me crowded together at the deepest end — almost nothing in between.

5.2) Speed = at DevDay, OpenAI put more emphasis on workflows where people are actively using AI to shape work in real time, in which case speed becomes super important (hence Ultrafast mode). Quality is important, but the feedback loops are tighter/faster

5.3) Value = Still seems like some heavy AI users see it as a badge of honor to spend thousands on AI. But more and more people want to get the most out of _one_ subscription, but value is a distant third in the discourse even though it's likely the most important thing to businesses as time goes on

5) Most businesses are chasing (in this order) quality, speed, and value. I can't honestly say I see huge differences in results from Opus 5.5, Fable 5.1, 6.1 Sol, or 6 Astra, but some folks are convinced [newest Anthropic model] is the only good model for what they do. (contd)

5.1) Quality = a lot of interest in whether models can go off and work autonomously for hours at a time with highest-quality output, with no human in the loop, or humans just involved at the end to review work and give notes

3) Almost no one had or has a good baseline for what "good" or "faster" looks like, so AI effectiveness is being measured on vibes, not hard data. More stuff is being produced, so it feels more productive in the short term, but outcomes are basically the same.

4) Likewise, everyone has some back-pocket tests for AI result quality, but almost nobody outside tech is doing benchmarking or structured evals to say with any confidence whether Opus 5.99 is actually better at writing than GPT-67 Andromeda, or whatever. Again, vibes.

2) The % of people who are truly "AI-pilled" (following all the trends, experience with multiple models/versions) is incredibly small. Most AI adoption is, like, buying Claude for Teams and having everyone use the Claude desktop app to make work happen faster.

bitsandletters.comHow I Think About AI Agents | Bits&LettersHow AI agents work, why I give them focused jobs, and how I use them at Bits&Letters to find leads, gather newsletter ideas, and keep track of tasks.

I've got AI agents looking for newsletter links, finding potential clients, and turning emails into tasks. One robot even acts as HR and a performance coach for the other bots. Here's how I think about putting them to work: www.bitsandletters.com/ideas/how-i...

emails.bitsandletters.comTo code and to not codeField notes from OpenAI DevDay 2026, on what it means to "build" if one no longer has to code.

In this week's newsletter, I report back from OpenAI's DevDay in San Francisco, and ponder what it means to "build" but not necessarily to "code." emails.bitsandletters.com/posts/to-co...

bitsandletters.kit.comPeople don't want to be productiveApart from ChatGPT, no AI products have seen meaningful consumer adoption. Will "agents" be different?

This week's newsletter is about AI agents and why most people don't see "productivity" as all that exciting. Also, this week's dog photo is *chef's kiss* 🥐 bitsandletters.kit.com/posts/peopl...

A few thoughts on yesterday’s twin AI model releases: The pitch from the big labs is changing from "this thing is smarter than humans" to "it can do the same work faster and less annoying for less money," implying that there are, in fact, limits to what AI is capable of.

More on compute — Google has the biggest compute advantage (they've been building/rolling out efficient AI processors for years, TPUs) but their models are mid. OpenAI and Anthropic have the leading models, but have struggled to find enough compute at a non-bonkers price.

If/when open-weight models fully pass GPT/Claude in capability, AI compute and user-facing services may start to be better businesses than frontier AI labs. Like, AWS/Azure selling high-performance instances for running models, and products using those models.

The fact that these models (Opus 5.5, GPT-6 Sol) perform about as well as the super-flagships is the clearest sign yet that model capabilities are plateauing. And if you can do a bunch of work on Opus without paying for special "Fable limits", why use Fable?

Compute is the main factor, and the moat. Opus 5.5 can be better and faster because Anthropic has made a lot of progress toward delivering inference-as-a-service more cheaply. It's not like their unit costs make sense, but they aren't having to do so much to ration access to this model.

David Demaree reposted

Dave Willner

@dwillner.bsky.social

17 Sept

Replying to @masnick.com

Certain kinds of smart people tend to have a "sexy apocalypse" problem when it comes to thinking about risk. E.g. skynet turning everyone in paperclips, while bad, is also *kind of cool and fun* at the same time. Where the reality is that most actual disasters are both stupid and disappointing.

reddit.comFrom the ClaudeAI community on RedditExplore this post and more from the ClaudeAI community

Let me tell you the agency/freelancer version of this: AI has caused every non-technical founder/owner/manager to believe that coding is now solved. Speed is the only factor. Even if they’ve never gotten anything usable out of Claude, they believe it’s possible if they just had the right prompt.

Andy Bell@bell.bz

If you think you’re alone in being miserable about being Claude‘s condom, you’re not www.reddit.com/r/ClaudeAI/c... Once you get past the loud minority — those who are dead behind the eyes, but smiling for the cameras — no one who is forced to use LLMs all day is happy in this industry

19 Sept

If your agent says: - Nothing is redundant for the current shape. The shape carries more than the product needs. - What reaches the model, and what bounds it. - The product, and where the code diverges from it that's not your agent, that's J.R.R. Tolkien

19 Sept

This morning I asked Fable to go over my website/ops monorepo with a vengeance, looking for stuff other agents (including previous Claude runs) had over engineered, overexplained, or generally left a mess. Four hours later, this was the result. This is an AI-generated diff I can get behind.