At the end of a long design session — an isolated USB-to-RS-485 dongle, maybe forty decisions deep: transceiver swapped for a 5 V part to survive 1 Mbps over 200 m of cable, termination and bias on DIP switches for multipoint buses, isolation split across an ADuM1201 and an isolated DC/DC — I typed the laziest possible review request:
"Looks fine. What did we forget?"
Not "check the termination network." Not "verify the enable pins." Just — what did we forget. The kind of question you'd ask a colleague leaning over your shoulder, and exactly the kind of question tools don't usually answer.
What came back
A checklist, and not a generic one — each item checked against the actual datasheets of the actual parts on the board:
The CH340C has a pin for exactly what we were doing the hard way. Pin 14 on the SOP-16 package is TNOW — it goes high during transmission, and the datasheet explicitly recommends it for RS-485 direction control. We'd been planning to derive the TX-activity LED and direction switching from TXD. The chip's designers had already solved that problem; the solution just lives on a page of a Chinese USB-bridge datasheet that nobody reads past the pinout table.
The isolator's DISABLE pin was floating in the schematic. ISO-family isolators have an active-low DISABLE (pin 7) that must be tied — to which rail depends on the fail-safe behavior you want. We had specified neither. That's the classic end-of-session gap: not a wrong decision, a missing one, invisible because nothing in the schematic looks wrong.
And one "we're fine" that saved a mistake in the other direction: I'd have reflexively added series resistors on USB D+/D−, because most USB chips want them. The CH340C datasheet says the opposite — "directly connect to USB bus, do not use series resistors." A review that only finds problems makes you paranoid; one that also confirms what's already right lets you stop second-guessing.
Plus the unglamorous tail: a PTC fuse on USB power, an RC network for the cable shield. Nothing exotic. Completeness rarely is.
And then I caught it making a mistake
Fairness requires the other half of this story. A few turns earlier, while reworking the power chain for the 5 V transceiver, the assistant proposed feeding a 5 V-output LDO... from the isolated 5 V rail. A 5 V-to-5 V linear regulator. My entire review comment:
"LDO 5V→5.0V Are you serious?"
It owned the error immediately — "you're right, that was a silly mistake" — deleted the part, and produced the corrected power tree: the isolated DC/DC already outputs the 5 V rail; the transceiver hangs off it directly.
I'm including this because it's the honest shape of working with an AI design companion. It reads every datasheet and forgets nothing on the page; it can still assemble parts into a momentarily absurd topology. You bring the topological common sense; it brings total recall with citations. The review runs in both directions — and both directions caught something real that afternoon.
The takeaway
"What did we forget?" might be the highest-value question in engineering, and it's the one that requires a counterpart who has actually read your parts' documentation — all of it, including page 14 of the USB bridge and pin 7 of the isolator. That counterpart used to be the senior engineer two desks over. Whatic is that counterpart in a chat box: it answers from the datasheets of the parts you actually chose, cites the section, and tells you which claims are document-verified and which are engineering practice.
Finish your next design, then ask it what you forgot: whatic.io