Last month I was about to spend money making new ads. I didn't need them. I built an agent to check first, and it caught what I missed. I already had a winning ad. I was just starving it.

For two years, the creative calls were mine. Look at the ads, decide what to scale, what to kill, what to test next. Last month I gave that job to an agent and pointed it at a live ad account. It read the account better than I would have, and faster.

Here is the whole run, with the real screenshots. One thing before you scroll. I have blacked out the client's brand name and their ad account ID. That is the only thing removed. Everything else is exactly what the agent did.

What I actually built

This is not a clever prompt. It is a full system. I took the way I diagnose ad creative, every rule I use, and packaged it into a skill that Claude loads and runs. One instruction file and six reference files. Competitor research, creative taxonomy, the diagnosis rules, the brief format, a learning loop, and how to run the whole thing.

I handed it over. It loaded clean.

One command, one account

Then one line from me. Run the engine on this account, last 30 days, give me the brief.

The account is a global jewellery brand. Thirty ads, eleven campaigns, about ₹13.67 lakh spent in the window, roughly $16K. 579 sales. On paper it looks healthy. Blended return on ad spend of 2.27. Click-through of 4.3 percent. Nothing screams problem.

That is exactly when you get lazy. Good top-line numbers hide the thing that actually matters.

The winner I was starving

The agent read every ad, looked at the actual creative, and found what I had missed.

Sit with that. I was about to go make new creative while my proven winner sat starving for budget. The fix was not a new ad. The fix was giving the winner more room to run.

It caught a second problem too. Almost every ad ran the same offer. Buy 2 Get 3. Ninety-three percent of them. The account had one idea and kept running it. That is a risk, not a plan.

The plan it handed back

Here is the part that made me trust it. It did not say make ten new ads. It split the budget into three lanes and told me where each rupee should go. Scale what is working. Fix what is close. Protect a slice for genuinely new angles.

Forty-five percent of budget to scaling, thirty to fixing, twenty-five protected for new ideas. The new-idea lane was not guesswork either. It pulled live competitor ads and proposed angles this account had never tried. Trust and craftsmanship proof. An occasion story. Real customers on camera.

It grades its own work

Most AI tools stop at the answer. This one wrote down every call it made, each with a date to come back and check it.

Next cycle it reopens the file and grades itself. Confirmed, partial, or wrong. The predictions start earning their confidence over time. An agent that marks its own homework is worth far more than one that only generates.

What this actually means

The lesson is not that AI makes ads. It is the opposite. The agent stopped me from making ads I did not need.

Most accounts do not have a creative problem. They have a seeing problem. They cannot tell what is already working, so they keep making more of everything. The lever here was reallocation, not creative. An agent that looks before it spends beats one that generates all day.

I have run this read by hand for years. Watching it happen in one window, on a live account, with a tracker and a follow-up date waiting at the end, was the first time the job actually felt like it had moved.

I am packaging this engine so you can point it at your own account. If you want it, start here.

And if you want to build agents like this with me, live, the next cohort is open.

Until next issue,

Apurv