From a one-paragraph brief to a schematic and PCB, with an engineer stepping in three times

A USB-C serial adapter is a small board, but getting from a brief to a design you can hand over still takes about a dozen steps, and any change early on means rerunning the checks after it. Here a hardware engineer wrote the brief in one paragraph and gave it to Micius, Zhiman’s model, which did the design work in the PCB software. The engineer stepped in three times: to set the order of work, a way to diagnose a fault, and what counted as done.

Challenge

A small board still takes eleven steps

A USB-C serial adapter plugs into a computer at one end and brings out a serial port for a development board at the other. This one is a two-layer board of about 38 × 22 mm. The usual route is long: check each part’s datasheet and library footprint and draw any that are missing, draw the schematic and run the electrical rules check (ERC), place and route the board, pour copper and run the design rules check (DRC), then simulate and write up the files. Each step is done by hand in the design software, and each depends on the ones before. Change one wire on the schematic and the checks already run no longer count; a route that won’t close means moving parts or adjusting rules.

This time a hardware engineer wrote the board’s requirements as one paragraph and gave it to Micius, Zhiman’s model. Micius did the design work in the PCB software, 934 operations in all. The engineer didn’t draw anything and stepped in three times.

Brief

The brief, and the first time the engineer stepped in

The brief asked for a compact USB-C serial adapter with regulated power, a UART breakout, a header pinout matching common USB-to-serial cables, and selectable configuration, on two layers with a closed board edge, sensible placement and clear labels. It ended with three working rules: explain the purpose of each step before starting it; when the tools report a real problem, say so and fix it; and build in any requirement a person adds along the way.

Micius split the job into eleven steps, from choosing parts to delivery and a final 3D render. The conversation recorded every action and tool response, so the engineer could see where the run had got to at any moment.

Engineer

3

interjections, each at a fork

Micius

934

operations across 11 steps

1. 01Brief

2. Stop fetching datasheets; use library parts and get the whole flow through first.

   02Pick parts
3. 03Place symbols
4. 04Check schematic
5. 05Wire pins
6. 06Fix and recheck

   Finds a clash between two power outputs; moves one pin to its own net via a 0 Ω resistor. ERC passes.

7. 07Place on the board

8. Work out why routing failed, adjust and retry until DRC passes.

   08Route

   Routes the USB pair first, then reruns the autorouter around it.

9. 09Pour copper, check

   DRC flags one clearance error and two copper islands; adds vias, rechecks, passes.

10. Fix the test field and rerun; every analysis needs a clear verdict.

    10Simulate, deliver

    ERC, DRC and the declared analyses pass; full design acceptance stays “unknown”.

11. 113D render

    Assembly preview; two parts lack 3D body models, and the record says so.

Micius searched the part library for the USB-to-serial chip, USB-C receptacle, ESD protection, fuse, headers, switch and voltage regulator, and tried to download datasheets and footprints from the makers’ websites. It kept getting web pages back instead of files. The engineer stepped in: stop downloading datasheets; use the library’s symbol and footprint where a part exists; build the regulator, which wasn’t in the library, from a generic small-package symbol and note where it came from; and get the schematic, placement and routing through first. Micius carried on in that order. The finished board has 23 parts.

Layout

Fixing a power clash, and the second time the engineer stepped in

Once the pins were wired, Micius found a power conflict: one of the USB-to-serial chip’s power pins sat on the same net as the regulator’s 3.3 V output. Micius went back to the schematic, moved that pin onto its own net and bridged it to 3.3 V through a 0 Ω resistor, and ERC passed. The run kept one rule throughout: when the design changes, every earlier check is marked “stale” and doesn’t count as a pass until it runs again.

With the parts placed, the autorouter failed. The engineer stepped in a second time, giving a method and leaving the fix to Micius: work out whether the cause is clearance, too few channels or escaping the pads; adjust placement or rules until DRC passes; then rerun ERC, DRC and the whole-board acceptance check and save. Micius traced the bottleneck to the USB differential pair, routed that stretch by itself first, and reran the autorouter around it. The board ended up with 175 track segments and 30 vias.

After the copper pour, DRC flagged a 3.3 V track at 0.1164 mm clearance against a 0.15 mm rule, and two isolated copper islands on the top layer. Micius listed each problem, added a ground via to each island, fixed the rest and reran DRC until it passed.

Results

What was handed over, and what wasn’t checked

At the final step, the engineer stepped in a third time: correct one field in a geometry test, then rerun the analyses and the acceptance check so every declared analysis actually ran to a clear verdict. After the rerun, three circuit simulations and an EMC geometry pre-check all passed, the schematic passed ERC and the PCB passed DRC. Micius handed over the schematic, the PCB, the project files and an analysis report in the design software’s own format, plus a 3D assembly preview, which notes that the USB-C receptacle and fuse lack 3D models.

The line for full design acceptance reads “unknown”. Appearance, CAD checks, simulation and real-board qualification are recorded separately, and a check that didn’t run is never shown as a pass. No board was built or tested. The run is logged as human-assisted and doesn’t count towards autonomous results; the three interventions sit in the record beside every operation Micius made.

This was a two-layer board with 23 parts. Boards with more layers, denser routing or tighter signal-integrity requirements weren’t part of the run. The design still needs an engineer’s review, and whether it goes to fabrication is the engineer’s call.