We sit down and have a chat — over a coffee, on the phone, or on a video call. That's genuinely most of it. Some conversations take a few minutes; if there's real complexity to a project, it can run closer to an hour. Either way, that conversation is where the actual information comes from.
Everything else — the website, the flyer, the accounts, whatever needs doing — comes out of what we talk about there.
Everything below this gets a bit more technical — how those conversations actually turn into finished work. Worth a read if you're curious, easy to skip if you're not.
I record our conversations — usually just on my phone, using an app called Wispr Flow — and each transcript, along with the evolving project work itself, goes into a Github repository dedicated to your company.
Transcripts are treated as inputs that get versioned alongside everything else: code that builds your website, design files or PDFs for marketing materials, photos and assets you've sent me, notes and decisions from our conversations, deliverables ready to use. When new conversations happen, they sit alongside earlier ones, with timestamps and full Git history. If something's said differently in a later conversation, we've got all versions and can track when the understanding shifted. It's all in one place, version-controlled and accessible.
Depending on what suits you, you can have read access to the repository itself — so you can see the full history and how things have evolved — or I can pull together summary documents or exported versions. Whichever's more useful to you.
Most of the actual work — turning a conversation into a working site, a flyer, updated copy — I do with AI. I run a Claude Pro subscription (£20 a month), usually on Sonnet 5 at medium effort, and OpenCode running Qwen 3.8 Max for extra capacity when Claude Pro's quiet. All of it runs off a Mac Mini M4 and an iPhone.
One thing worth being upfront about: people naturally aren't perfectly consistent between conversations — a detail gets remembered differently, something genuinely changes, or two things you've told me just don't quite line up. The Github repository surfaces those inconsistencies rather than quietly picking one version and moving on, so we can sort out what's actually true together, in a friendly way.
It's close to what Andrej Karpathy has described as a living, constantly-updating document, rather than a one-off report that goes stale the moment it's written.