A picture tells an AI what something looks like. It does not tell it why the thing looks that way, and the why is the part a builder usually forgets to hand over. Nico Lafakis had a story about that from his 3D work, and Chris Carolan had the same lesson from the other side of the table, on Value-First AI Daily, Friday, October 30, 2026 (Episode 46).
This piece is written for someone who did not watch. The show’s sealed board of three news stories was not read on air, so nothing here covers it. The hosts used the time for their own builds and stories instead. The hosts’ words are cited by their time on the recording, and where we join things up, we say so.
Episode — Value-First AI Daily
Market Simulator Max, Live with Nico
Episode 46, October 30: Chris Carolan and Nico Lafakis, no guest. The recording runs just under 54 minutes. Nico’s beta tool is walked through between about the 13-minute and 28-minute marks, and his 3D builds run from about 30:00 to 38:30.
Open the episodeA first pass that did not work
Nico said his initial background is in 3D animation (32:43 to 32:53). When he saw what Claude could draw, he wanted to push it as far as it would go, and his first try was a cinematic scene in the style of Pacific Rim, with the pieces of a machine coming together. By his own account it fell short, and he said so plainly: “very clearly this does not work” (33:25 to 33:46), because the look did not suit the data the scene was meant to represent. He kept something from it anyway, a view of what 3D modeling and rendering can do.
What worked was a longer route (36:18 to 37:09). He moved between Claude and GPT to settle a visual style, made concept art, built storyboards from the art, built a motion reel from the storyboards, and only then gave Claude the reel with instructions about the goal. Even then he started small. The first build ran on a small model (he named Haiku), to see whether the rough was, in his words, “in the ballpark”, and he stepped up to a larger one each time the rough held (37:09 to 37:51). That is his account of his own process.
The link without the route
Chris heard his own gap in it. He said Nico’s storyboard route was “what I was missing” (38:52 to 39:24). He had told Nectar, his AI assistant, that he wanted one of these, and had handed her the link to Nico’s finished piece. Our reading: a link shows a result and none of the steps that produced it.
Nico put it in terms of context (40:59 to 41:17). Hand a model only the images, he said, and it lacks “what came before those images happened.” His own fix was to have the assistant take screenshots, research where the idea in them came from, and work backward, which he said would have led it almost at once to the Tron: Ares scene the look came from (41:29 to 42:12). Then comes the check (42:12 to 42:30). Before anything is built, ask the model to describe what it believes it is about to build. If the description matches what you meant, go ahead. If it does not, correct it then, before it has built the wrong thing.
That check costs one reply and lands before the work, which is where a mismatch is cheapest to find. The first rough on the lightest model is the same idea at the scale of the picture. Both are our reading, not the hosts’.
Chris on the paperwork
Chris took the same point to the value. To get the outcome you want, he said, there is no shortcut past “describing, documenting, telling it what the value is” (43:36 to 44:07). The sentence is about building with AI, and it is also close to the Value-First position that value has to be named before anyone can build for it. That second link is ours; he did not make it on air.
The same thing, in a simulator
Nico’s main demo that day was Market Simulator Max, a tool he built for filling a HubSpot account with practice records. He called it a beta, “not live” and “not for purchase” (19:06 to 19:20), and invited anyone willing to test it to tell him.
What aired was a walk-through of the setup, not a run. From about 12:49 to 27:53 on the recording, Nico stepped through each setup section on a draft built from valuefirstteam.com. No simulation was started on air and no record was shown arriving in HubSpot. The screen at the final step carried a banner saying new work could not start for a few minutes while the app was being updated, which neither host mentioned. Everything below about what the tool writes into an account is Nico describing it, and we have not seen it do so.
His description is where it connects to the rest of the day. The tool is meant to fill an account with records “in a realistic fashion” (13:25 to 14:04), so a client demo has something better than “dummy data” entered once and never touched again (14:11 to 14:36). Setting one up, he said, comes down to “describing what it is that your client does” (15:04 to 15:19). Pointed at a business’s website, it infers the kind of business and its products from what it finds, and Nico said those products are then added to the deals once a simulation starts (20:05 to 20:16). The input is a description, and the output can only be as good as the description. Chris closed the stretch with a line of his own: “context is literally everything” when you work with AI (28:47 to 28:57).
Ask once more
Chris told one more story, about the other end of the same problem. The day before, in a HubSpot user-group session on building an agent, he was testing one on a company record when Breeze, HubSpot’s AI, told him it had no access to that record and asked him to go and find the ID. He retold the exchange rather than reading it. He typed “no, period, you,” and by his account Breeze then fetched the record (06:24 to 06:43).
What he drew from it concerned the human side (06:49 to 07:22). Much of the human contribution right now, he said, is a small course correction, and the cause is usually not an AI making things up. It is context that is “a little bit off”: something left out, or an assistant believing it lacks access it has. He credited that framing to Nectar. His rule of thumb followed (08:03 to 08:31): if you have seen the system do a thing before, do not overthink it, and ask it once whether it is sure. Nico added that how you word the first request makes a large difference to how well the model understands what you want (09:04 to 09:24).
What to carry out of it
The habit we take from the day is small. Before the next build, write the paragraph that came before the picture: what it is for, who it is for, and what it has to hold. Hand that over with the picture. Then ask the tool to say back what it thinks it is making, and read the answer before anything exists. When it tells you it cannot do something it did last week, ask once more with the missing piece in hand. What is the paragraph for the thing you are about to hand over?
Worth passing on?
