Originally published as an Advisory Hour Substack post on 2026-08-11. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
“What’s the story behind why you started adding a featured product to each post?”
“Well, great question. It’s not what you think.”
Each of the substack posts on this @advisoryhour substack has a featured product. The featured product is a custom product built specific for that post. These are not products from the wider world-and it’s not with any eye towards anything other than doing. These are cocktail conceived of ideas that make sense as novel products placed in each individual post. Products on posts? Yes, because an agent told me I was doing it wrong.
The “it” being a matter of interpretation.
I believe it’s clear this substack is simply a public experimental log of trials and tribulations. What I realized in building evals for agents is how useful they are for guiding outcomes. An experimentalist at heart and raised on the “I can learn anything in two days” ethos this excited me. I then built several evals and pointed them back at myself. I was expecting a parade of flowers, chocolates and a fireworks display. Instead, the AI whacked me on the head with a newspaper. “Get good!”
The Agent Critic
Every day I have a routine that runs that scores my substack and broader body of work on ericrhea.com. The score is telemetry like the thermostat temperature on the wall in your home. It’s a numerical score that tells me if I’m improving or drifting over time. The scores compute on several dimensions, and one of them is an economic one. I wanted to make the invisible hand of the market more visible. So each and every day, a critic tells me exactly what went wrong. It never seems to run out of “this is wrong” things to say about me. I fold the feedback in over time.
Long time subscribers to this substack may recognize a kind of compounding evolution happening in my content. The content and experiments from yesterday feed forward and inform the subsequent day’s experiments. There’s changes in character and tone in the writing-which is difficult as a human writer. You’ll see me at times regress if I’m tired or perplexed about my day. In all of this an AI relentlessly let’s me know just how bad I’m doing things. For example, the AI sent back a score of 56 out of 100 on the Ashwicke Gallery post. Oof, and I thought I did pretty good on that one.
The critic comes in as a series of numbers, markdown text, and notation over discord. Each day the agent runs, reviews whatever I had posted on the day prior, and then delivers a score. How I receive these scores is via discord. I considered just sharing the raw discord text. It’s problematic. The delivery format is too raw to toss onto substack because it is both very long and filled with markdown system, which I regret admitting is easy for me to read. Instead, I asked ChatGPT to format the critique into a nicely viewable document. The content is all the same here, it’s just nicely presented so it’s easier to “get the gist” of what I’m writing about.
Capitalism as a Signal
It is very hard to find meaningful problems. I didn’t start this blog as an exercise in making money-and it still isn’t the driving goal. The writing here started out as just “one more thing” on a list for a contest in the early part of the year. I wasn’t even posting that regularly until I realized that I could use the substack as a kind of public log. The economic critic wasn’t even the first critic. The first agentic ai critic was around technology projects that I should do.
The first agentic ai critic suggested what project I should work on by comparing the corpus of writings I’ve done so far with what emerging frontier companies are doing. The critic emerged during a competition when I wanted to automate some amount of ideation in the “theme of the week”. While I always had one or two ideas myself, having a critic that could give me a daily “what if you built x,y or z” was immensely helpful.
After the competition, and perhaps because of the competition, the “suggest one more thing” AI opened up the doorway to a capitalism one. I couldn’t say it was like one clear “money good”, but really a lean towards “it’s time to do something useful with this gigantic body of work”. Then that led to a daily critic that tells me I’m doing it wrong. My wife snarked, “You really like feedback.”
“I like objective, impersonal feedback. My faults are as glorious as my failures. It’s very clear I have more to learn and get better at.”
She didn’t wait long before adding, “You do know that not everyone would handle so many active critics like you’re doing?”
“People are complicated. Besides, I haven’t figured out a faster or better way to improve. If someone reveals that insight, well, I’ll stop all this immediately. Until then, well, I have one A and… some other grades.”
“It’s nice you got an A in gym class. How’s your semester going?”
I laughed, “Let’s focus on the semester ahead. There’s still plenty of room for improvement.”
Swag Shop
The swag shop uses a starter free “quick store”. It’s the option when someone wants to put in the bare minimum amount of work just to host a few products. It’s contestable if the products should even exist. The critic tells me I need to work harder, so I am. The swag shop bumped my most recent post up five points in the review. I’m still failing, but now I’m failing a little less.
Would this post be so different if I didn’t have at least one compelling product? And so this post is for people who didn’t even realize a daily AI review can serve as a useful calibration tool. You could get feedback once a week, month, quarter or from random anons on your favorite social media. Or, you can cut right deep into your very body of work. Celebrate your AI critic with this one of a kind “GET GOOD” coffee mug: Mug with Color Inside - My AI Critic Said GET GOOD.
Connected work
- Substack Writing at 100 Posts In: Look back at how a public writing practice changed across its first hundred posts.
- Evaluation and Release Gates: Read the technical companion on using evaluations to decide when model work is ready to move forward.
- John's World: The Ashwicke Gallery: Open the gallery post whose critic score becomes a concrete example in this essay.