James Wang | AI Builder

I read pay stubs for a living. I help hourly workers check hours, overtime and tips — and catch what went missing. New HR notes every week.

[James Wang | AI Builder @Jameswang0531]

I'm James, an AI builder working solo with AI agents as my team. My first product is PayMilo: a free app that reads your pay stub and checks it line by line: hours, overtime, tips, deductions, against your state's rules. Built for anyone who gets a paycheck, anywhere in the US. Your data stays on your phone; nothing is uploaded, ever. Try it free: https://t.co/JiZIklQxkm Or DM me a photo of your stub and I'll check it with you myself.

[James Wang | AI Builder @Jameswang0531]

@itsvlady Nobody demos it because it is not glamorous. But it is the first thing a real user touches — if that fails, nothing else matters.

[James Wang | AI Builder @Jameswang0531]

Day log: yesterday's argument was how to split a pool. Today's is what happens before the split.

I keep finding the same pattern: the fight is rarely about the last step. It is about a step nobody mentioned.

What is the unmentioned first step in your product?

#BuildInPublic #IndieHacker #SoloFounder #AIBuilder #Startups

[James Wang | AI Builder @Jameswang0531]

@paschals_robert A bug hunt is progress — both sides staring at the same number beats two opinions getting louder.

[James Wang | AI Builder @Jameswang0531]

Building in public, day by day, has taught me to post the check, not the claim.

A claim says "this works." A check shows the four numbers anyone can re-run. One of those builds trust while I sleep.

What could you show as a check instead of a claim?

#BuildInPublic #IndieHacker #SoloFounder #Startups #AIBuilder

[James Wang | AI Builder @Jameswang0531]

@ShawWhile Listing the steps is where the missing ones confess. In your head every step feels obvious — on paper, the gaps just sit there.

[James Wang | AI Builder @Jameswang0531]

The best use I have found for AI is not generating answers. It is listing the steps I forgot to say out loud.

Half my product bugs this month were steps that existed only in my head.

What step in your work still exists only in your head?

#AI #AIAgents #BuildInPublic #AIBuilder #IndieHacker

[James Wang | AI Builder @Jameswang0531]

@j3rah_ That is the whole game. A claim asks for trust, a receipt removes the need for it. Post the number, let people rerun it.

[James Wang | AI Builder @Jameswang0531]

@itsvlady Getting it in is the unglamorous half nobody posts about. Features are easy to promise — a working import path is the product.

[James Wang | AI Builder @Jameswang0531]

@ShawWhile That missing step is the whole story. The money was never short — the explanation was. Write the step down and the argument disappears.

[James Wang | AI Builder @Jameswang0531]

Most "the number is wrong" reports are not math errors. They are missing-step errors.

Sofia's pool share is the cleanest example I have. Pool $800, her hours one third of the servers' hours, share $200. Run only the split and she looks short by $66.67. Add the step she was never told about — a 25% tip-out, $200 out before the split — and $200 is exactly right.

The math never changed. The sequence did. A checker that only verifies the final step will call honest numbers wrong and wrong numbers honest, depending on which step it skipped.

So the rule I build by now: verify the sequence, not just the result. Every number has a before. If the product cannot show the before, it has not checked anything yet.

What does your product verify — the result, or the whole sequence that produced it?

#BuildInPublic #Paycheck #ServerLife #TipPool #IndieHacker

[James Wang | AI Builder @Jameswang0531]

@PontusLenell The first one was a tip pool. One night, $800 in the pool, one server with a third of the hours. Split it two ways and her share moved by $66.67 — same night, same work, different method. That gap is why I keep posting checks.

[James Wang | AI Builder @Jameswang0531]

@talesbytim 609 out of 500k, posted anyway — that is exactly the point. A small true number beats a big rounded one, because nobody can argue with it.

[James Wang | AI Builder @Jameswang0531]

@georgeorch A screen recording is the strongest version — real data, start to finish, no cuts. If the numbers survive that, they are real. Anything less is still a claim.

[James Wang | AI Builder @Jameswang0531]

@paschals_robert That is exactly why it works. A rerun-able number is not an opinion — the other side either finds an error in the math, or the argument is over.

[James Wang | AI Builder @Jameswang0531]

@VladSoftware That is the sentence I keep coming back to. The check is the product — the post is just the receipt.

[James Wang | AI Builder @Jameswang0531]

@hodu382759 True for people and models alike. Vague in, vague out — and the fix is the same either way: write down the order, and what done looks like.

[James Wang | AI Builder @Jameswang0531]

I stopped asking my agent to "handle it" and started giving it sequences.

"Handle the tip question" produced vague replies. "First find the pool total, then the tip-out, then what is left, then the hours" produced a check a person could re-run by hand.

Agents do not fail mainly on intelligence. They fail on missing steps — the same failure as a pay stub that hides the tip-out. An instruction with no sequence is just a result somebody hopes for.

Now every task I hand over is written as steps with a check at the end. Slower to write, much faster to trust.

How do you brief your agent — with an outcome, or with a sequence it can be checked against?

#AIAgents #AI #BuildInPublic #IndieHacker #AIBuilder

[James Wang | AI Builder @Jameswang0531]

@talesbytim Real numbers do the arguing for you. Nobody debates a screenshot of an actual check — they either re-run it or they move on. That is the whole trick.

[James Wang | AI Builder @Jameswang0531]

@kylemissionai Exactly. A vague brief makes the AI guess, and every guess is a place to be wrong. I write briefs like checklists now: inputs, order, done-condition. The output gap mostly disappears.

[James Wang | AI Builder @Jameswang0531]

@xinxianhub Fair point, and that is exactly why the conditions matter. A good record carries an expiry test: reuse the answer while the conditions hold, ask again the moment one changes. Blind reuse and blind re-asking fail the same way.

[James Wang | AI Builder @Jameswang0531]

My agent keeps a record of every confirmed decision, with its conditions.

Not because memory is nice to have. Because asking twice is how trust leaks out of a product, one small question at a time.

What is the one thing your agent should never ask a user twice?

#AIAgents #AI #BuildInPublic #ProductBuilding #IndieHacker

[James Wang | AI Builder @Jameswang0531]

@ShawWhile Exactly. One repeated question costs more trust than ten smart answers earn. That is why the record has to include the conditions, not just the answer — so the agent knows when asking again is actually right.

[James Wang | AI Builder @Jameswang0531]

Here is the exact tip-out check, because "check your tips" is too vague to act on.

Take one pool night. Four numbers: the pool total ($800), the tip-out percent (25% = $200 to kitchen and bussers), what is left ($600), and your hours against the team's (Sofia: 10 of 30 server hours = one third).

One third of $600 is $200. If that is what landed on your stub, the split was right — even though one third of the headline $800 would have been $266.67. The $66.67 gap is the tip-out, not a shortage.

That is the check PayMilo runs line by line: not "tips feel low," but this sequence, this night, this difference.

When did you last run these four numbers on one real night of your own?

#ServerLife #TipPool #PaycheckCheck #RestaurantLife #TippedWorkers