I read Austin Henley’s post Automating my job away on a Friday night and did the laziest possible thing with it.
Henley’s whole piece builds to one prompt. A friend who runs a startup tells his team don’t do anything three times - if a task comes round more than twice, automate it. Henley took that to its logical end and pointed his coding agent at its own history.
So I copied his idea, swapped “Copilot” for “Cowork”, and pasted this in:
Cowork, read my Cowork session logs from the last 7 days
and propose opportunities for automation.
Draft a reusable skill for each.
That’s the whole thing. Copy it, point it at your own logs, change the 7 if a week is too short. I didn’t write a spec. I didn’t give examples. I gave it one sentence and went to make a cup of tea.
It came back with five new skills.
I gave it good material
This worked because I’m a heavy user and I leave a trail. Cowork logs every working session, and the last seven days were not quiet. Roughly thirty of them. Super and Age Pension modelling, monthly finances for the Board, fact-checking my mate Russell’s economic forwards, benchmarking local AI models for my agents, filing Medicare PDFs into Obsidian. The usual sprawl.
That sprawl is the point. An agent can’t spot a pattern in work you only did once. The reason the prompt landed five skills and not zero is that I’d already done each of these jobs three, four, six times - by hand, slightly differently each time, never quite the same shape twice.
Here’s the part that genuinely surprised me. It didn’t just list tasks. It clustered them, and it showed its working by citing the session numbers that justified each one.
What it handed back
It noticed I’d modelled my retirement finances six separate times - SMSF drawdowns, super versus the Age Pension taper, a CGT position on a holding - so it wrote au-retirement-modeller, with the Australian super and pension rules baked in so I stop re-explaining them every time.
It noticed four sessions where I’d turned a public data source - an IMF forecast, an ASX sector breakdown, an S&P chart - into an Excel workbook with a chart and some Australian context. That became markets-data-workbook.
It saw that three times I’d taken an informal bit of economic commentary, fact-checked it claim by claim and drafted a measured reply in my voice. That is now commentary-factcheck-reply, and yes, it specifically named the Russell forwards.
It caught the three sessions where I’d benchmarked a local AI model against a cloud one for my agents and wrote autonomi-model-benchmark so the test runs the same way every time and the results are actually comparable.
And it spotted four sessions of hunting down an email or a PDF and filing it into my Obsidian vault under my own storage rules, so it built email-pdf-to-vault to do exactly that, cross-links and all.
The interesting part isn’t the tools
The five skills are handy. They are not the story.
The move that matters is the one Henley is really pointing at, via Karpathy’s autoresearch and the current rush toward self-improving agents. It isn’t an agent that runs your tools. It’s an agent that watches itself run your tools, notices what it keeps doing, and builds the missing tool without being told the tool was missing.
You stop writing scripts and start writing the thing that decides which scripts are worth writing. The prompt above is one sentence long because the agent supplies the judgement. All I supplied was the history.
The catch nobody mentions in the demo
Henley is honest about the sting in the tail and I will be too. Automating the obvious work doesn’t hand you a clear afternoon. It surfaces the glue work that was hiding underneath, the bits too small to ever bother with. The five-minute job you did twice a day becomes a thirty-second job you now do twenty times, and the context-switching between half a dozen agents is its own tax. He found himself busier, not freer, and making silly mistakes he never used to make.
I believe him, because I can feel where this goes. Five skills is five new things that can run slightly wrong while I’m looking the other way. A skill that bakes in the Age Pension taper is a gift right up until the taper rules change and I’ve forgotten the assumption is buried in there.
So these five are on probation. A drafted skill is a hypothesis, not a promise. The test isn’t whether Cowork could write it - clearly it can, in the time it took my tea to brew. The test is whether the third version of each is still right six months from now.
The rule was never automate everything. It was don’t do anything three times. The quiet upgrade is that the machine now does the counting for me, and it counts better than I do.
Go and paste that prompt into your own agent. The worst case is it finds nothing, and you’re out thirty seconds.
Sources:
- Automating my job away - Austin Z. Henley
- autoresearch - Andrej Karpathy