AI Policy

Last updated: Nov 6, 2026

AI is Pandora's (tool)box. It's been opened, turned upside down, dumped all over the world, and we can never put it back.

In this world, where VC-backed con artists are desperate to stop losing money and continually overstate the capabilities of what they're selling, we (ahem, Sarah) tend to talk more about the negatives to balance things out. However, many of the TTS team use AI on a daily basis and are having a lot of fun with it, while others hate it and refuse to touch it with a 10 foot pole.

The differentiator tends to be developers: we love it, because it genuinely makes tedious things much quicker and easier, although we're all using it a little bit differently. Sarah started out mostly using it to write code snippets and continues to (slowly, carefully), expand the use cases.

Here's the key point to keep in mind: AI is a tool, like a hammer, and you shouldn't try to use a hammer to bake a cake.

OK maybe that's a weird analogy, but the idea here is that AI use makes sense in certain scenarios, and that's where we use it. The line is a different for everyone, but broadly speaking:

AI should handle grunt work, (guided) coding, data analysis, and automation-type tasks to free up time for human brains to work on real problems, like high-level strategy, writing, communication, and user-friendly design.

AI can assist with some of these things, but like a junior intern, it needs a lot of guardrails. Here's a very broad overview of how Sarah is using AI as of this writing (OK let's just switch to first person now).

Models:

  • Claude Pro (99%) – all the things
  • Gemini (1%) – Google and SEO-related questions
  • ChatGPT free (.0000001%) – occasional testing to see search results
  • OpenAI API – a couple defunct test/beta projects

Writing

  • If I send you an email, I actually wrote it. I did not have AI write it (unless you're being fired, in which case I might have asked Claude to say it nicer).
  • If I write a blog post, I actually wrote it, from scratch, by typing on a keyboard. Novel, I know. Occasionally I'll use AI to generate an outline (mostly for my newslettters) because you're supposed to write outlines and I never do because they're annoying even though they save me editing time and I totally should oh well.
  • I have no problem using AI to write website blurbs in low-stakes scenarios. For example, a client had team bios that were bullet point style factoids, and we had a summary section on another page where I wanted a single sentence mini-bio. Asking AI to turn a fact list into a single sentence (that I usually tweak myself) is a great use case.
  • I will use AI to clean up client-provided copy that's messy or hard to follow.
  • I did not use AI to write this page. I just started writing, at like 3 AM because my founder brain won't shut off, and that's why this page sounds the way it does. You're welcome.
  • For any AI-generated or edited copy, I usually always tweak it myself. I've been a writer my entire life, after all, and I have OPINIONS.

Strategy

  • I will probably ask Claude what should be included on an AI policy page and use those recommendations to fill in any gaps or things I missed.
  • For technical integration type things, such as connecting two pieces of software to each other, I might talk through the options I'm considering with Claude to see if I'm missing pros/cons or surface helpful documentation.
  • I will occasionally ask AI to parse out tasks from clients who write overly long emails that don't clearly explain what they even want to do. This makes it easier to create a task list.

Design

  • You know what, I actually have zero problem with clients using AI to play with logo concepts. It's not great for brand designers, but it does make our job easier, because while it might give you an easy way to see and communicate what you like, it will still need manual cleanup. So instead of recreating grainy logos from your next-door neighbor's cousin with a bootleg copy of Photoshop, I'm now cleaning up logos from a robot that outputs cleaner files to start with. I'm ok with this because we're no longer a branding agency; we're just trying to get your dang website online sometime this year and I need a logo to do that.
  • Brand guides: can AI create these? I sure hope so. Forever testing this kind of thing, but what a great example of something annoying, tedious, and a great job for the unpaid intern.
  • I have experimented with AI mockups for quick turnaround projects. I generally try to avoid mockups anyway as I prefer to design during development. I feel mostly the same way about this as I do with AI logos; it's actually more helpful if a client spends their own time creating a website mockup to show me that they like, vs. trying to describe with words. I can then review, improve, make suggestions and build the professional version with styling they've already approved.

Websites/Hosting

  • I use AI with Github, and some static sites, with extremely limited write access and only for certain projects.
  • I'm testing AI build features for WordPress, mostly in hopes I might be able to reboot our abandoned DIY WordPress package now that new technology is here.*
  • I do not use AI for any DNS management. Domain setup, connections, updates, etc., are all handled manually by mostly me. Humans break DNS all the time (forget that, DNS just breaks itself all the time). DNS is not something I want AI to put it's grubby 1s and 0s on.

MCPs and Integrations

  • Very limited testing, and with mostly read access. For example, I'm using the MCP for my project management system, mostly for data analysis (a task AI excels at).
  • I do not just connect every new MCP the moment the email announces it. It must serve a purpose, and be tested carefully. The lower the risk of damage, the higher the chance I'll test it. The MCP for marketing reports that are a total pain to set up? Great use case. The email signature software that now, unbeknownst to me, has linked my headshot to a manifesto in praise of Darius Amodei... ehhh let's maybe not connect that MCP.

Things change quickly, and if anything major changes, I'll update here, but this gives you a good starting point to understand my overall philosophy.

If you're interested in a deeper dive, in 2025 I wrote a 3-part series on AI in my newsletter: OK Fine, I'll Talk About AI [Part 1].

*DIY WordPress packages are actually a great educational jumping-off point. I'm here to solve problems. Years ago, I built that package to solve the problem of a slightly more tech savvy individual who wanted to use WordPress but didn't have the tech chops to figure out the server setup part of things.

They never worked. I provided the keys to the kingdom, and there were great demo templates to install in just a couple clicks, which would allow the user to click to update certain bio sections, whatever. But that was still too overwhelming for people.

What AI does is provide a text, chat-based interface that can interpret intent. So instead of having to learn how to use a specific CMS (open an editing window, click to edit a module, find the Advanced tab to adjust the padding, no wait it's coming from the column...), you could ask the AI, in plain language, and it could do that for you. I believe this is the future of AI. HOWEVER, AI can suck at updates, which is why I think still having a CMS for manual edits is the best of both worlds. I'm excited about the possibilities here.

In this case, AI is a support layer between a client sending an email asking me, say, how to update the contact form to the new one they just saved, and instead of me manually logging in, doing it, taking a screenshot and explaining it, the AI can do all of that instantly, meaning my time with the client is spent discussing their goals, their current site's capabilities, and what I recommend they do to build this new process (strategy, i.e. human brain work).