Describing outcomes, not pressing buttons

Thanks for being here for the first issue of this newsletter., If you’re reading this, you’re a legend. This is my first attempt at something like this, and I’m going to work it out as I go.

Quick introduction for anyone who doesn’t know me: I’m one of the co-founders of Friyay. I’ve been working in AI for almost three years now, and I still get excited, every single day, about what these tools can do. I think they can help a lot of people, and my goal with this newsletter is simple: help as many people as possible get real value out of AI.

Here’s where it all comes from: everything in this newsletter is drawn from the last two weeks of my actual work. That’s one-on-one coaching with some genuinely brilliant people, AI training programs across a real mix of Australian businesses, building AI tools and assistants for companies, and running and updating the tools we use inside our own business. There’s usually a few good nuggets in there, and that’s the idea: pull them out every fortnight and share them.

I’m excited and a bit nervous putting this out. If anything lands, or anything doesn’t, reply and tell me. I’d genuinely value the feedback.

Ok, enough jibber jabbing:

Something showed up across my sessions these past two weeks that I keep thinking about. Not a feature. More of a shift.

I ran a check-in this fortnight with a team that’s been using Claude for a couple of months. We were going through what had stuck and what hadn’t. One person said almost in passing: “I keep catching myself about to do something in the spreadsheet, and then I stop. Because I realise I shouldn’t be touching it at all.”

That’s it exactly. The shift from doing the task to describing the outcome. (Microsoft’s 2026 Work Trend Index names this same shift, from doing tasks to delegating outcomes, as one of the defining changes in how people work with AI this year.)

A different thing showed up with a client I coach one-on-one, an executive who’d been using Claude for complex research: feeding in transcripts, reports, and correspondence. He kept losing track of where things were saved, and was re-explaining his folder structure to Claude at the start of every conversation. So we sat down with Claude and talked through what he actually needed. Claude proposed a structure, what he now calls his vault, he pushed back on a couple of things, and once he’d approved it we set it up in OneDrive.

The bit that stuck with me: it was never really about Claude being able to organise folders. It was that they’d co-designed not just the structure but the purpose of each folder, what belongs where and why. Because of that shared understanding, he can now ask Claude to find something or file something, generically (”this relates to X, put it wherever it belongs”) or specifically (”save this in the Y folder”), and Claude gets it right either way, because Claude has access to his OneDrive and SharePoint. The first time it happened without him spelling it out, he looked at the screen and said: “Does that mean I can do that every time?”

The third thing that surprised me this fortnight wasn’t voice, it was how few people are actually using it. More people I work with have started talking to their AI tools instead of typing, using tools like Wispr Flow or Superwhisper. Not because I’m forcing anyone into it, I’m just strongly encouraging it. Every single person who tries it doesn’t go back. Not dictating: actually talking back, asking follow-up questions out loud, giving feedback by speaking. One person said: “I speak so much faster than I can type. And it kind of picks up on how I talk.” (Stanford research has found speech input can be up to three times faster than typing.) If you don’t want to subscribe to a new tool just yet, ChatGPT and Claude both have this built in already: there’s a small microphone icon on the right-hand side of the prompt box. Tap it and you can talk through your prompt instead of typing it. If there’s one thing in this issue I’d want you to actually go try, it’s this. Not for my referral link (message me if you want it, you get a stack of free minutes), I just don’t want anyone still typing when talking is this much faster.

Three things to try

• Before you do something manually in Excel or Google Sheets, try describing the outcome to ChatGPT or Claude and ask it to use excel or google sheets to achieve the outcome.

• Ask Claude to design your folder structure before you do any reorganising. Open a chat and say: “I want to restructure my Drive so you (ChatGPT or Claude) can find things easily. Here’s what I work on: [describe it]. How would you suggest I organise it?”

• Try talking instead of typing. The tool I use is Wispr Flow; my co-founder uses Superwhisper. Both work across Claude, ChatGPT, and anything else you type into. Try one for two sessions before you decide. (Reply to this email for my Wispr Flow affiliate link — free minutes for you, free minutes for me.)

What’s new in AI (worth knowing)

Fable

Thursday was officially a Fable-less day, until Anthropic brought it back online and made this a Fable-ulous edition to write about it in. Honestly, it barely changed anything for me: I run Sonnet 5 for 99% of what I do, and Sonnet 5 only shipped a couple of weeks ago itself. I reckon almost nobody needs Fable day to day. It’s a genuinely good marketing moment for Anthropic, but the real use case is narrow: complex, multi-step planning. That’s the only place I reach for it. I’ll run the planning conversation in Fable, get it to drop the finished plan into a downloadable document, then start a fresh chat (or a new task in Cowork) and switch models before doing the actual work. The model selector sits bottom right of the prompt box: click the arrow and pick Sonnet. I skip Opus entirely. Sonnet gives me results just as good, without burning through the tokens Opus does.

Sol, Terra and Luna

OpenAI announced three new models this week: Sol, Terra, and Luna. I went straight to my ChatGPT model selector to try them and found nothing. They’re limited preview only right now, available through the API and Codex to select partners. If you looked and came up empty like I did, that’s normal, it won’t be long until they roll out more broadly. When they do, I’ll be testing all three against Sonnet, Opus, and Fable to see how they compare. One bit of trivia that stuck with me: this is the first time OpenAI has named model tiers instead of just version numbers (3.5, 4, 5.5). That naming pattern, Sonnet, Opus, Fable, Mythos, has been Anthropic’s for a while now. Feels like a sign of who’s setting the pace at the moment.

Claude Design as an MCP

I recently added Claude Design to my Claude Code setup, and it’s changed how I review things. When I ask Claude to work on a task or show me options, it sometimes creates a web page for me to look at. What I find really cool about it is that Claude decides when to do this — I don’t have to ask. It just pops up. And I find I’m much better able to review and give feedback when I can actually see it visually. To set it up, add Claude Design through the MCP settings in Claude Code.

If anything from the above resonates with where you are, reply to this email. I read every one.

Gareth

P.S. The insights in this issue came from workshops and one-on-one coaching sessions I run on AI adoption. If any of it sounds useful for your team, or you’d like to work through your own AI adoption, reply to this email and we’ll work out what makes sense.