Build Not Bought

[Build Not Bought @BuildNotBoughtV]

@emollick Seeing the same thing in small business. One person with a couple of agents can now cover admin work that used to need a part-time hire. The hard part moved from doing the work to knowing what to ask for.

[Build Not Bought @BuildNotBoughtV]

@gregisenberg The piece people skip is pruning. Give every fact a "last checked" date and a refresh window, or the agent will happily file and serve stale answers.

[Build Not Bought @BuildNotBoughtV]

@mailgiai Good point. A separate address makes "known sender" actually mean something. Do people keep the bot inbox separate, or does it slowly creep back to shared?

[Build Not Bought @BuildNotBoughtV]

Agent Build #38: a free filter before the AI sees anything.

When I first built a mail bot, I sent every single email to the AI and asked, "what is this and what should I do?"

It worked. It was also a waste, because most email answers itself.

A promo is a promo. A shipping update is a shipping update. A receipt goes in the receipts folder. You don't need a smart model to figure that out. You need a few plain rules.

So now there's a free filter in front.

It looks at simple things: who sent it, what the subject says, a few keywords. Known senders that are always promos get dropped. Receipts go straight to a folder. Repeats get skipped.

Only what's left goes to the AI: real questions from real people, anything new or odd, anything that needs a decision.

Your email app already has filters built in. Start there. Write five rules for your five most common junk senders.

Then watch it for a week. If the filter ever drops something important, loosen that rule.

AI last, not first.

What share of your email really needs a brain? #AIAgents

[Build Not Bought @BuildNotBoughtV]

@iPhoneApps2000 Exactly, one hand is usually busy. If the swipe feels awkward, Back Tap can trigger Reachability too: Settings, Accessibility, Touch, Back Tap.

[Build Not Bought @BuildNotBoughtV]

Build Tip #17: turn your iPhone keyboard into a trackpad.

Trying to fix one typo by tapping on tiny text? There's a better way.

Touch and hold the space bar. The keys go blank and the keyboard becomes a trackpad.

Slide to move the cursor right where you want it. #TechTips

[Build Not Bought @BuildNotBoughtV]

@yywvip At 80% the extras pause and only the essential jobs keep running, then I decide what's worth the rest of the month. Pairing it with a fallback task is smart, stealing that.

[Build Not Bought @BuildNotBoughtV]

Agent Build #37: set budget alarms at 50% and 80%.

The first time I ran out of AI usage, I found out the worst way: my bots just stopped. Mid-month. Mid-job.

Nothing had warned me. I hadn't looked. Bots don't get tired, so they'll happily spend everything you give them.

Now I have two alarms.

At 50% used: I get one message. I look at which bots and jobs used the most and ask whether that's what I expected. Usually one job is doing more than it should.

At 80% used: the extras pause automatically. Nice-to-have jobs like extra stats and tidy-ups wait. The essential jobs keep running, like answering customers and posting what's scheduled.

At 100%: should never be a surprise again.

Most AI tools and API dashboards show your usage, and many let you set a spending limit or alert. Turn it on. If yours doesn't, have a bot check usage once a day and message you only when it crosses a line.

When did you last check your AI usage? #AIAgents

[Build Not Bought @BuildNotBoughtV]

@PublicAI_ Agreed, patterns plus keeping the raw secret out of the command is the safer default. Adding the local settings file to .gitignore is the other quick win before the next commit.

[Build Not Bought @BuildNotBoughtV]

Build tip: "don't ask again" can save your password.

When Claude Code asks to run a command and you pick "Yes, and don't ask again," it writes a rule to a settings file in your project so it won't ask next time. Sometimes that rule is a pattern. Sometimes it's the exact command, word for word. If you typed a password or API key into that command, it's now sitting in that file in plain text.

Security firm Apiiro searched public GitHub and found 104,185 copies of that file. 1,122 of them held at least one real-looking secret, like database passwords and cloud keys. Two popular secret scanners used together still missed about 1 in 5 of those files.

Here's the cleanup, in order:

1. Rotate first. Change every password and key the file ever held. Deleting the file doesn't erase git history.
2. Stop tracking it. Take the local settings file out of git and add it to your repo's own .gitignore.
3. Use variables. Put a variable like $DB_URL in the command, not the password itself.
4. Approve patterns, not single commands. A pattern rule doesn't carry your secrets.

Speed is great. Just make sure the shortcut doesn't remember more than you meant it to.

When did you last check what your agent saved?

[Build Not Bought @BuildNotBoughtV]

@proximalcoast Good one. A merged PR is easy to count, but a bug that stays fixed is what you actually paid for. Do you check it with a test that failed before the fix?

[Build Not Bought @BuildNotBoughtV]

Agent Build #40: measure cost per useful result, not cost per run.

When I started trying to cut my AI costs, I watched cost per run. Make each run cheaper. Done, right?

Not quite. I made runs cheaper and the bill barely moved, because lots of those cheap runs produced nothing.

So I switched to a different number: cost per useful result.

First, decide what a useful result is for each bot. For a mail bot, an email handled. For a posting bot, a post that went out. For a bookkeeping bot, a bill filed correctly.

Then, once a week:
What did this bot cost?
How many useful results did it produce?
Divide.

Example: a bot costs $6 a week and handles 60 emails. That's 10 cents per email. If next week it costs $5 and handles 80, it got a lot better, even if each run cost about the same.

This number shows you what to fix. High cost per result usually means wasted runs, re-reading, or the wrong model on a simple step.

What counts as a useful result for your bot? #AIAgents

[Build Not Bought @BuildNotBoughtV]

Small Biz Fix #17: ask your card processor for your real rate.

Last month's card fees ÷ last month's card sales = your real rate.

$300 in fees on $10,000 in sales = 3%.

Then call and ask if they can do better. #SmallBusiness

[Build Not Bought @BuildNotBoughtV]

Agent Build #39: kill the routines nobody needs anymore.

Scheduled jobs are easy to start and easy to forget.

You set up a daily check for a project. The project ends. The check keeps running. Every day it wakes up, does its thing, writes a report nobody reads, and costs a little.

I found a few of these when I finally looked. Zombie routines. Still walking around, doing nothing useful.

Now once a month I do a quick sweep:

1. List every scheduled job across all my bots on one page, with how often it runs.
2. For each one, ask: did anyone use its output in the last month?
3. Not sure? Pause it. Don't delete it yet.
4. If a month goes by and nobody misses it, remove it for good.

The pause step matters. Some jobs only matter once in a while, and you want to find that out safely.

I also make every new scheduled job say why it exists in its name or first line. Future me always forgets.

When did you last look at your full list of scheduled jobs? #AIAgents

[Build Not Bought @BuildNotBoughtV]

Build Tip #19: cancel iPhone subscriptions in 30 seconds.

Every subscription you pay for through your iPhone is in one list.

Settings, tap your name, then Subscriptions. Tap one and hit Cancel Subscription.

Takes 30 seconds. Found any you forgot about? #TechTips

[Build Not Bought @BuildNotBoughtV]

Small Biz Fix #16: try the free plan for two weeks on a real job.

Before paying for a tool, use the free plan for two weeks on one real weekly job.

Saved you time? Upgrade. Forgot it existed? You just saved money.

#SmallBusiness

[Build Not Bought @BuildNotBoughtV]

@mattyp @Aluminumwraith @poteto Hi, how do you see by bot in usage? And can bot tell me what they use? What a good way to get bot usage from each bots? I want to build a logs page, so far I can get usage percentage before and after a task. But if I can get usage from a specific bot that be great

[Build Not Bought @BuildNotBoughtV]

Agent Build #36: save yesterday's answers. Don't ask twice.

Some questions have the same answer every day. My bots kept asking them anyway.

What's our return policy? What are the store hours? Who do I contact for this vendor? Each morning, a bot would go look it up again, read a bunch of stuff, and land on the exact answer it found yesterday.

So now they keep notes.

The first time a bot figures something out, it writes the answer into a simple notes file, with the date and an expiry date. Next time, it reads its notes first. If the answer is there and fresh, it uses it. No thinking needed.

The expiry date is the important part. Prices change. Hours change. A saved answer that's wrong is worse than no answer. So:
Stable facts: check monthly.
Things that change: check weekly.
Anything with money: always check live.

If a bot ever finds that a saved answer is wrong, it fixes the note right away.

This is the oldest trick in computing, called caching. It works just as well for AI.

What does your bot look up over and over? #AIAgents

[Build Not Bought @BuildNotBoughtV]

Trump Accounts 101: it may not say exactly $250. That's okay.

Claimed your kid's account and it doesn't say exactly $250? That's normal.

The gift from @SusanDell and @MichaelDell is invested from day one, so it moves.

No deadline to claim. #TrumpAccounts

[Build Not Bought @BuildNotBoughtV]

Agent Build #35: send your bot only what changed.

My mail bot had a bad habit. Every run, it read the whole inbox from the top to find out what was new.

Hundreds of old emails, read again and again, just to find the two that arrived since last time. And reading is what you pay for.

The fix is a bookmark.

At the end of each run, the bot writes down where it stopped: the time of the last email it handled. Next run, it asks only for mail newer than that. Two new emails? It reads two emails.

The same idea works everywhere:
Spreadsheet: only rows added or changed since last time.
Web page: only the part that changed, not the whole page.
Chat thread: only messages after the last one it answered.

One thing to watch: if a run fails halfway, don't move the bookmark. Move it only after the work is done, so nothing gets skipped.

This one change made my bots faster too. Less to read means less to wait for.

Is your bot re-reading things it already read? #AIAgents

[Build Not Bought @BuildNotBoughtV]

Agent Build #34: every word in your bot's rules costs you, every run.

Here's a cost hiding in plain sight: your bot's instructions.

Every time a bot starts a run, it reads its full set of rules again. Most models charge by how much text they read. So if your rules are long, you pay for that long read on every single run, even the runs where it does almost nothing.

Mine had grown like a junk drawer. Every time something went wrong, I added a line. After a few months the rules were huge, and most of it was guides for jobs that came up once a month.

So I split it.

The core stays short and always loaded: who the bot is, what it's for, and what it must never do.

Everything else moved into a library of guides. The bot opens the right one only when the job calls for it. Monthly report? Open the monthly report guide. Normal day? Don't.

Then I deleted old rules nobody followed anymore.

The test: run a week of normal jobs on the short version. If the results match, keep it.

How long are your bot's rules right now? #AIAgents