The Ecosystem Problem

• 1 Junedate the side-project gate fell due
• 30 Septemberwhere I moved it, and it passed too

On Monday 6 April I asked my AI co-pilot to audit what I was actually working on. The answer came back: seven active workstreams. Seven. On a house.

It began with a simple idea: electrify the house, put up panels, fit a battery, write some automations, save some money. One project. Clean edges. Then the tools I was using (a good home automation platform, an AI co-pilot, a laptop, a bit of time in the evenings) made it so easy to start the next thing that I started it. Then the one after that.

At my own kitchen table, about my own battery, my own hot water tank and my own weekends, the list looked like this. There was the core energy control work, the automations that run the battery. There was a data platform built around a big longitudinal store of everything the house measures. There was the writing-up, which needed a hosting and publishing set-up I had not finished. There were the dashboards, four of them queued for a rebuild. There was a small cluster of mini PCs running local language models. There was an AI research folder full of notes on models and agents I thought I might one day stitch together. And there was a side project idea built on that cluster, with a lot of planning and nothing else.

When I wrote them down and asked, in so many words, “is this sensible?”, the answer was a very polite, very well-evidenced “no”.

Handing my scope to a critic

The technique is not complicated. I pointed a structured critique workflow at my own project: the same pattern I use to review blog drafts and automation plans before I ship them. I gave it the list, the hours I realistically had each week, and the dependencies between items, and asked which ones were delivering value to me, the homeowner, and which were exploratory sprawl I had drifted into because the next idea was always shinier than the one in front of me. I gave it the usual permission: don’t be nice. Be useful.

The verdict sorted the seven into three bins. Three got the time: the energy control work, the data platform underneath it, and the writing-up. The battery automations were saving real money, and the data platform was the reason I could prove it. My April working figure was about $1,500 a year of grid import avoided, before any export revenue, annualised from a four-month sample: a number to think with, not to bank. The billing had been checked against the invoices down to the cent. Two were kept running but not built on: the dashboards, once the rebuild was done, and the cluster. And two were put behind gates: the AI research folder, parked until the optimisation work needed something from it, and the side project idea, which got a date.

None of this was a surprise; I half-knew all of it. But half-knowing is the problem. Half-knowing lets you keep every plate spinning because you never actually decide to drop one. Writing it down and asking for a structured read turned half-knowing into a decision.

Run it again

I ran the audit twice, two days apart; the second run was two rounds of critique. The second, on 8 April, is the one I learned most from, because a lot had moved in forty-eight hours. The patches that stopped the battery automations misbehaving after a restart were deployed. The feed-in pricing question that had stalled the writing was settled. The dashboards were rebuilt and the cluster was finished and benchmarked. The second pass retired “seven is too many” as the headline worry and put something sharper in its place: winter was eight weeks away, and the data I needed to tune for it did not exist yet.

It also wrote down a caveat I have since come back to. Part of the relief was temporal, not structural. If several of those workstreams woke up at once, the sprawl would be back.

The co-pilot could read what I had logged and give it back in a shape I could act on.

The gate

The most useful single thing the audit produced was a kill-or-commit gate on the side project idea. A date: 1 June. To commit, it needed the writing live and some outside interest in the idea. If not, I would stop working on it and not let it become a zombie. The default was to stop, not to continue.

It felt lighter at once. Not because I had decided to kill it. I had not. But because I had decided to decide. The open loop was the weight, not the work.

What happened on 1 June

I did not make the call.

The test could not run: the writing was not live, so there was no outside interest to measure. The gate fell due and passed without a decision, and the day after I moved its due date to 30 September. The 30th of September came and went the same way. As I write this, on 26 October, the item is still open and marked overdue.

In April I expected the idea not to survive the gate. It neither survived nor died, because I never closed it. A gate whose default is to stop only works if you let it close, and I let mine slide twice.

In late October the co-pilot drafted a tidy-up plan for my home projects, and it recommends marking the idea formally dormant. That is a proposal on the page, not yet my answer.

The cluster, now

In April the audit kept the cluster running for the house but gave it nothing new to do; the language models on it had no real job. Now it has one. On 18 October I set it up to run the heavy optimisation runs from my retailer comparison work, which take twenty minutes or more on the desktop, spread across the nodes. Keeping it ticking over without building more on it turned out to be the right call. The AI research folder, parked in the same breath, is still notes.

The pattern

If there is a general lesson here, and I am cautious about claiming one, because this is still just me and my house, it is something like this. My projects grow without anyone pruning them, because the next idea is always cheap to start and no idea is ever quite dead enough to throw away. An AI co-pilot is good at pruning for the same reason a friend with no stake in your garden is good at telling you which shrubs to dig up. It is not smarter than you. It is just not attached.

The specific trick is to treat your own project the way you would treat a draft. Hand it to the critic. Ask for the structured read. Run it again when things have moved. Write the decisions down with dates on them: “decide by 1 June” is a decision, “let’s see how it goes” is not.

And then the bit I got wrong: the co-pilot can help you draft the gates, but you have to be the one who lets them close.