Previously at the Volcano Base I’d been working with AI inside a “second brain”. Since then, I’ve been prepping an automation system for testing by a business. I also used the recall.ai software development kit to create an Obsidian plugin for meeting transcription and analysis. Scratching my own itch and replacing the paid subscription tools I’d used previously.

Oh, yeah, and I’m still “open to work” due to pipeline blockages. It’s about as uncomfortable as you imagine it is.

At some point during every project, there arrives a ceremony of sorts. Documents are signed. Credentials are handed over. Someone says something like "from here, it's over to you." Everyone nods gravely, as though a baton has been passed in an important race, and then goes home for the weekend.

This is the handover. It’s a very important ritual. It’s also, to a significant degree, fiction.

Not dishonest fiction. Useful fiction. The kind that helps everyone feel something has concluded and allows the consultant to invoice without guilt. But a fiction nonetheless, because the thing you've built didn't stop when you handed over the keys. It's still running. It's still touching real data. It's still about to encounter a Tuesday that nobody planned for.

Here’s something that happens on every project of this kind: the client is slow to test.

Not through negligence. Not because they don't care. But because they have, as it turns out, an entire job. A billion meetings a week. A Teams channel they're trying to mute. A pile of things that were urgent before your project arrived and will remain urgent regardless.

So they test when they can. But then, inevitably, wonderfully, they start using the thing. And they start saying "wouldn't it be good if it also did X." And X is always a reasonable thing. Something that would genuinely help. Something that was, quite reasonably, outside the scope of the original brief.

And here’s where some consultants feel a small, quiet panic.

No garden, once planted, contains only what you planted in it. This is one of the more reliable facts about gardens. Seeds arrive on the wind. Creatures take up residence in corners you didn't designate as residence. Something with too many legs discovers that your inbox is structurally similar to the habitat it prefers and raises a family there.

In a technology garden, the uninvited guests tend to arrive as: an API that quietly changed its date format over a bank holiday weekend; a user who discovered, entirely by accident, that if you submit the form at 11:59pm on a Sunday it creates a duplicate record with the subject line "undefined"; a European privacy regulation that drifted in as a seed on the wind and is now rather difficult to remove without disturbing the whole bed.

None of this was in scope. None of it was anyone's fault. It just moved in.

"Wouldn't it be good if..." isn’t scope creep. It's something much better: it's evidence that the system is already working in the client's imagination. They're not complaining about what you built. They're dreaming about what comes next. That's a success, not a problem.

The panic happens because we've framed the whole engagement as a race to a finish line. If there's a finish line, everything after it is either overrun or failure. But gardens don't have finish lines. Nobody reaches the end of a garden.

You plant a garden. You establish it. Then you step back and let it grow at its own pace, in its own season. The client tends it slowly, between meetings, with one eye on their inbox. New things come up. Some are weeds. Some are exactly what you hoped for. Some arrived from next door without asking and are technically someone else's problem but are very definitely now in your garden.

And then, when the season turns, you come back.

Even if your client tested everything thoroughly and signed off with great conviction and the system was truly, definitively done, the underlying technology would keep moving.

The AI model that powers part of the workflow gets a new version that is demonstrably better at everything except the one specific thing you were using it for. The API that fetches the data changes its authentication in a way that is thoroughly documented in a changelog that nobody reads because it is a changelog. The platform you integrated with releases an update that works slightly differently in a way that's technically not a breaking change but is absolutely a breaking change.

The line in the sand doesn't just get moved by client wish lists. The tide moves it too.

This isn’t a flaw in the project. It's just the nature of building things with living components. Pretending otherwise is the kind of optimism that looks brave right up until the moment it looks naive.

Here's what I’m telling myself: stop selling projects. Sell growing seasons.

Season one has a defined scope, a timeline, and an end date. You plant what was agreed. You hand over the keys (still) but with a different story attached. This isn't the finish. This is the end of the first growing season. The system will evolve. Things will move in. The wish list will grow. That's expected. That's the point.

Then there's a natural pause. The client lives with what you built. They find the wish list. They discover what the second season should contain. And when the time comes, they hire you again with a new scope, and a clear sense of what they're buying.

The client's day-job tempo is no longer a problem. It's the growing season doing what growing seasons do.

You're not the gardener working for free forever. You're the one who comes back when it's time to plant again, deals professionally with whatever moved in, and charges accordingly for both.

Some interesting stuff I discovered on my adventures.

Figma gets into the AI workflow business
Figma has quietly rebranded its AI design tool as Weave and folded it into its ecosystem. Node-based workflows, multi-model access (OpenAI, Google, Runway), professional compositing tools, and an "App Mode" that auto-generates simplified UIs from complex workflows. Worth watching if you do anything visual and want AI in the loop without losing control of the output.
weave.figma.com

Teaching an AI to use Obsidian
Kepano (Obsidian's CEO) has released a set of agent skills that let AI assistants work directly inside Obsidian vaults: editing notes, building Bases, generating Canvas maps, running CLI commands. 24,000 GitHub stars in short order. The line between "note-taking app" and "AI workspace" is getting harder to find.
github.com/kepano/obsidian-skills

The case for leaving Substack
A cleanly put-together resource for newsletter writers who've been on the fence about Substack's content moderation record. Not a rant, just evidence, migration guides for seven alternative platforms, and politely worded templates for nudging your favourite writers to follow.
leavesubstack.com