When I'm stuck on a decision, I ask: can I walk this back next week?
If yes, I decide today.
If no, I slow down and write it out.
Most of what I stall on is the first kind.
Idea via Jeff Bezos (two-way doors), in my own words.
What are you stalling on that you could undo?
4
If I've searched for the same kind of thing three times this month, that's an agent.
Not a big idea. Just a repeat search I hand off with clear rules.
What I look for. Where. What counts. When to stop.
What do you keep searching for over and over?
3
If the first version doesn't embarrass me a little, I waited too long.
So I ship the rough one and watch where people get stuck.
That's my fix list for the next version.
Idea via Reid Hoffman, in my own words.
What's the roughest first version you've shipped?
Every agent I build gets an off switch before it gets a schedule.
A deadline. A result. Or a week with nothing found.
An agent with no stop rule doesn't keep helping. It turns into noise.
How do you decide when a tool has done its job?
1
Before I trust a new agent, I test it on a search where I already know the answer.
If it finds it, it gets real work.
If it misses, I fix the instructions and run it again.
A known answer is the cheapest test I have.
How do you check an AI's work?
2
New agents only get the jobs I'd do with a checklist.
If one misses, I can see exactly which step was vague.
After a few clean runs, it gets something bigger.
What job would you hand an AI on day one?
6
I reread my own instructions before I blame the agent.
Usually my spec was the fuzzy part.
Fuzzy: "Find good deals."
Clear: "Within 50 miles, listed this week, with photos."
The agent did what I asked. I just asked badly.
What's the vaguest instruction you've handed off?
4
My first search agent tried to do everything. It got slow and confused.
What worked was one agent with one job.
So now I add the second piece only after the first runs clean.
Idea via John Gall, in my own words.
What's the smallest version of your next system?
13
My first version of anything is a plain list in a file.
No dashboard. No app. No logo.
If I keep opening the list, it earns a nicer home. If I don't, I just saved a weekend.
What's the boring version of the thing you want to build?
23
I show the work before it's finished.
Half-built tools. Rough drafts. The thing that broke last night.
Waiting for done means feedback comes too late.
Showing it unfinished means it comes while I can still use it.
What are you holding back until it's "ready"?
1
15
My agents tell me when they're not sure.
Every result gets marked sure or not sure. The sure ones I trust and move on. The not-sure ones get a second look from me.
My time goes to the few results that actually need me.
What would you want your agent to double-check with you?
1
22
I write the job description before I build the agent.
What it watches. What counts as a hit. What it has to bring back. When it stops.
If I can't write those four lines, I'm not ready to build it.
What would your agent's job description say?
1
11
My first search agent tried to do everything. It got slow and confused.
So I split it up: one finds, one checks against the rules, one writes the report. And instead of pinging me all day, they send one digest.
Would you rather get pinged all day or get one digest?
7
I ship things before they feel done.
Not sloppy, just early enough that real use tells me what to fix.
The version in my head is always better. It's also always useless, because nobody can use it.
What's something you shipped before it felt ready?
25
Every AI agent I build has one rule: no receipts, no result.
If it finds something, it brings back the link and when it checked.
If it can't show where it found it, I throw it out.
It's the fastest way I've found to catch made-up answers.
What's the first rule you'd give your agent?
1
11
I'm building AI agents that go find things for me.
Lesson one: give them a spec, not a vibe. Then one agent per source, and every result has to come back with receipts. 36 seconds on how I set them up.
What would you send an agent to go find for you?
1
7
Small annoyances are a tax I pay every day.
So on Saturday morning I look back at the week and ask one thing: what bugged me more than twice?
That one gets fixed today. A 20-minute fix on something that annoys me daily pays for itself fast.
What's the small thing that annoyed you most this week?
5
Before I build a dashboard, I track the number by hand for a week.
If I stop checking it, it didn't matter.
If I keep checking it, now it earns a dashboard.
A week of tally marks costs nothing. A dashboard nobody opens costs a weekend.
What's the one number you check every day?
5
Rent vs buy isn't rent vs the mortgage payment.
Owning adds property tax, insurance, HOA and repairs, and buying and selling both cost money. How long you stay changes the math. 42 seconds on the full picture.
How long would you stay in your next place?
8
Anything I do more than once gets the third-time rule.
First time, I do it by hand.
Second time, I write the steps down.
Third time, I automate it.
Automate too early and you lock in the wrong steps. Wait too long and you keep paying in hours.
What's the task you've done by hand way more than three times?
3
















