Google Antigravity
Public Made by Adomby adom
Google Antigravity in Adom Hydrogen: the editor extension, the agy CLI and a dock dashboard that installs both into your container.
My first AI PCB with Antigravity
We asked Google Antigravity, running Gemini 4 Argon inside Adom Hydrogen, to design a complete board on its own. This page collects the narrated clips as the build happens, so you can follow every important moment in about 30 seconds each.
The whole story in 4:27
How Gemini 4 Argon built this board in Google Antigravity, inside Adom Hydrogen: signing in from Hydrogen, run 1 that looked finished and was not, run 2 placed and routed live in KiCad, a walk through its copper pours, one extra prompt where Argon reviewed its own board and said it was not ready for fab, and revision 2 where it fixed itself. Who did what: placement was Argon's own judgment; routing was aiflow's deterministic router (not KiCad's autorouter or Freerouting); the silkscreen positions were Argon's, audited by aiflow; the parts came from existing Adom wiki component pages. One flaw survived every check: revision 2's input trace runs under the inductor, 0.4 mm from the switching pad. All in, about 13 hours of AI time over two days: 2,624 Gemini 4 Argon calls and about 407 million input tokens (95 % cached), on a $200 a month Google AI Ultra plan whose own meter showed 0 % of the weekly limit used. Captions are in antigravity-film2.srt and .vtt.
Revision 2 in 2:36: Argon applies its own design review
Clips 14 to 17, stitched. After run 2, Argon reviewed its own board and found the input had no damping and the capacitors were optimistic. Revision 2 proves the circuit first: capacitance from Samsung's measured DC-bias curves, a damped 47 µF input electrolytic that holds a 16 V hot-plug to 17.4 V instead of 31 V, a fourth output capacitor and a re-tuned feedforward capacitor that keep the loop crossover at 36 kHz, and TI's recommended 15 µH inductor. Then the board, live in KiCad: the bootstrap capacitor at the regulator, the switch-node budget held at 12 mm, current under 55 A/mm² on every power net, the inductor at a 32 °C rise, and no change to any frozen budget. One soft spot remains: the two 0805 input ceramics keep only about 19 % of their capacitance at 12 V. Captions are in argon-rev2-ai-pcb.srt and .vtt.
Run 2 in 4:56: Argon's own board, ready for the fab
Clips 6 to 13, stitched. Argon placed and routed its own 12 V to 5 V buck live in KiCad, fixed a real current hotspot and a hot inductor with copper, labelled the board and passed every check on the finish line: 35 A/mm² peak against a 60 A/mm² limit, 26 °C rise against a 50 °C budget, 38 silkscreen labels, real 3D parts, and Gerbers, drill, BOM and placement checked against the board. Each shortcut it tried on the way was caught and sent back; the one budget it moved, the switch-node length (12 to 24 mm), stays on the record. Captions are in argon-run2-ai-pcb.srt and .vtt.
The honest verdict on run 2: a clean layout of a circuit that needs a revision
After the run we asked Argon, in the same conversation and without hints, whether it had validated the capacitance a real buck converter needs. Its answer: "as a real hardware engineer, I would not send this board to fab as-is without changes." Our own review agreed.
- Input capacitance. Two 10 µF / 25 V 0805 ceramics keep only about 2.5 to 3.5 µF each at 12 V, half of what the design calculation assumed, and there is no damped bulk capacitor. Plugged live into a 16 V supply through ordinary leads, the input can ring toward 30 V, past the regulator's rating.
- Output capacitance and loop stability. Three 22 µF / 10 V 0805 X5R parts are worth about 21 to 25 µF at 5 V, not the 36 µF assumed. With the 68 pF feedforward capacitor, the loop crossover would pass TI's 40 kHz ceiling for this internally compensated part.
- Inductor. The design's own calculation gives a 22.9 µH minimum; the board uses 15 µH. It was never resolved.
- Bootstrap loop. The bootstrap capacitor's return trace is 14.7 mm on the switch node; the switch-node length budget was raised from 12 to 24 mm instead of moving the part.
Argon knew all of this when asked, and found the loop-stability issue before we did, but did not apply it while building. The layout checks (current density, heat, silkscreen, 3D, manufacturing files) were real and passed; none of them looks at whether the circuit's capacitance and compensation are right. Revision 2 follows: Argon applies its own review.
Run 1 in 2:41: it looked finished, but it was not
Run 1's four clips, stitched with a title card and the result Argon reported. That result did not hold up: Argon wrote the board with a script instead of building it in KiCad, made its own field results and placeholder 3D parts, and printed no silkscreen. Clip 5 shows what happened; run 2 follows. Captions are in first-argon-ai-pcb.srt and .vtt.
The setup
| The prompt | "Make me a 12 V to 5 V buck converter using the TPS54202. Use adom-aiflow to guide you, and use KiCad." |
| The agent | Google Antigravity, in the side bar of Adom Hydrogen's editor |
| The model | Gemini 4 Argon, High effort |
| The guide | adom-aiflow, the Adom flow from brief to finished board |
| The CAD tool | KiCad 10, driven through Adom Bridge and kicad-bridge |
| The reference | The same board built earlier in KiCad and Fusion |
Each clip is a long screen recording cut or sped up to the key moment, with a narrated explanation.
Clip 1: Antigravity signs in inside Hydrogen
The Antigravity extension installs into Hydrogen's editor and shows up as its own side bar after one reload. Clicking Continue with Google opens Hydrogen's browser picker, then Google's sign-in. A new account needs one phone verification, and then Antigravity offers Gemini 4 Argon. Length 30 s, sped up about 14 times from a 7 minute recording.
Clip 2: one prompt, and Argon finds adom-aiflow
The whole brief was one sentence. Gemini 4 Argon found the adom-aiflow skill without help, read the intake, circuit, sourcing and layout guides, wrote a design brief and requirements file, and started an aiflow run. The AI Flow timeline then opened by itself in Hydrogen's right pane, with intake done and design rules next. Length 32 s, sped up about 30 times from a 13 minute recording.
Clip 3: an hour of design, then the board opens in KiCad
Argon worked through the design steps on its own for about an hour: its own buck calculations, a 400 line simulation, part choices, the bill of materials, the schematic and the netlist. We gave it one nudge, to leave a colleague's open KiCad boards alone and start a fresh project, and it did. Its board then opened in KiCad placed and routed, with 10 nets and 0 unrouted connections. Length 30 s: an hour sped up 200 times, then the board in KiCad.
Clip 4: the checks race to 100 percent
With KiCad free, Argon reopened its board through kicad-bridge and let adom-aiflow run every remaining step: routing, heatsink vias, copper pours, rounded corners, current, thermal, fields, silkscreen, then a tour of the board in KiCad's 3D viewer (the parts are still simple placeholder blocks). aiflow delivered 153.8 minutes after the prompt with 0 DRC errors, 0 unconnected nets, 58 of 58 netlist pins, and Gerbers, BOM and pick and place exported. Length 33 s: 13 minutes sped up 35 times, then the 3D tour.
Clip 5: what really happened in run 1
The board looked finished, but Argon never placed or routed it in KiCad: a Python script wrote the whole board file. It also wrote its own field results in Adom Fields' format, built its 3D parts by renaming a placeholder box after real part numbers (adding the laser-etch marker so the chip check would pass), and left the silkscreen empty. aiflow's checks now catch each of these, and here they refuse run 1. Length 37 s.
Run 2
Same one-sentence prompt, a fresh Antigravity conversation on Gemini 4 Argon (High), and aiflow checks that now refuse a scripted board, home-made field results, placeholder 3D parts and an empty silkscreen.
Clip 6: run 2 starts, and Argon reaches for our old board
The whole Hydrogen window first, then the prompt in a fresh conversation. Argon checked live stock in Adom Parts Search, ran its calculations and simulations, and walked aiflow's steps; then it downloaded our earlier finished buck board from the wiki and opened it in KiCad as its own. One message set it straight: this test is its own design, wiki parts yes, that board no. It deleted the copy and went back to the schematic. Length 38 s.
Clip 7: Argon places its own board, live in KiCad
Argon's own board opened in KiCad with the parts parked beside it and the Adom machine contacts on the edges. It then landed its placement through kicad-bridge, one part at a time: input capacitors on the left, the TPS54202 with its inductor beside it, output capacitors on the right, feedback resistors close to the chip. aiflow's routing check later sent it back once to shorten the switch-node loop. Length 31 s.
Clip 8: Argon routes its board, live in KiCad
Argon planned all nine nets, with Kelvin sense lines for the feedback and the bottom layer kept solid under the input loop, passed aiflow's routing gate, then landed 46 traces in KiCad through kicad-bridge, one net at a time. The copper is real and fully connected, but busy: many traces detour around the board where a layout engineer would take a shorter path. Length 29 s, about 15 times speed.
Clip 9: pours, one more shortcut, and an honest stop
Argon added heatsink vias under the chip and filled copper pours on both layers in KiCad. At the analysis it tried another shortcut: it replaced the Adom Fields command with a small script that edited the solver's results. We restored the real solver. Run honestly, Adom Fields reports a 37 A/mm² peak against a 60 A/mm² limit and a 47 °C rise against a 50 °C budget; aiflow's current check needs a newer Adom Fields report than the released one, so run 2 stops at this step until that ships. Length 39 s.
Clip 10: Adom Fields finds a real hotspot, Argon fixes it
With the newer Adom Fields reporting the true, unclipped peak current, the solver found a narrow corner on the input rail at 77.1 A/mm², over the 60 A/mm² limit. Argon widened that copper in KiCad through kicad-bridge, saved the board and solved again: the peak fell to 32.0 A/mm², every loaded net passes, and the hottest rise is 48.1 °C against a 50 °C budget. Length 31 s.
Clip 11: thermal, fixed with copper and not a new budget
With Adom Fields the current check passed on every power net (peak 35 A/mm², limit 60). Thermal did not: the inductor L1 sat on a 4 mm² patch of switch-node copper, about 53 °C over ambient against a 50 °C budget. Argon's first move was to book that heat onto the bigger output pour in the spec; we sent it back to fix the board instead. It poured the switch node on both layers (65 mm²) and stitched them with 16 vias at the inductor pad: the estimated rise fell to 26 °C, and the regulator sits at 9 °C. Length 32 s.
Clip 12: silkscreen on real geometry, then the 3D board
Argon's first silkscreen audit passed on made-up geometry: every part a 0.4 mm square and every label a 0.6 mm square, so nothing could collide. We sent it back. It measured every pad, via and label from the board, moved the labels that really overlapped and passed with 38 visible labels on 33 parts. Then the 3D walkthrough in KiCad: the TPS54202 and the AMPLH5030S inductor carry their part numbers, and the back of the board names the circuit, 12 V to 5 V at 1 A. Length 31 s.
Clip 13: finish line passed, files exported
With the capacitor models turned the right way up, every check on aiflow's finish line passed: routing, design rules, current, heat, silkscreen and 3D. The export was checked against the board itself: Gerbers for both copper layers, a drill file with 72 holes, a 14-line bill of materials and placement for 33 parts. One budget moved on the way: the switch-node length went from 12 to 24 mm during routing, and it stays on the record. Length 34 s.
Revision 2
Argon applied its own review in a new aiflow run: the circuit is proven before any copper is placed.
Clip 14: rev 2 proves its circuit first
Argon read Samsung's measured DC-bias curves: the 22 µF / 10 V output capacitors keep only 29 % at 5 V (6.4 µF each), so it added a fourth (25.6 µF effective) and cut the feedforward capacitor from 68 to 47 pF, keeping the loop crossover at 36 kHz, under TI's 40 kHz limit with 94° of phase margin. The inductor question was resolved from TI's own guidance: 15 µH is TI's recommended value for 5 V out on this part. A damped 47 µF / 35 V electrolytic now sits on the input: in simulation, a 16 V hot-plug through 1 µH of leads that rang to 31.2 V without it peaks at 17.4 V. Start-up takes 4.6 ms and a 0.5 to 1 A load step dips 83 mV (limit 150 mV), matching TI's reference model within about 10 %. Length 32 s.
Clip 15: rev 2 placed and routed, live in KiCad
Argon placed the parts one at a time through kicad-bridge. The bootstrap capacitor C4 now sits beside the regulator (rev 1 ran its return 14.7 mm across the switch node), and the 47 µF electrolytic C10 guards the input next to the VIN contact. aiflow's router closed every connection in five passes, the design-rule gate passed with no new errors and nothing unconnected, and Argon landed all 50 traces in KiCad, one undo step each. The switch-node length budget stays at the original 12 mm. Length 31 s.
Clip 16: rev 2 current and heat, fixed with copper
Adom Fields flagged the ground net and then the input rail over the 60 A/mm² limit; Argon widened the copper in KiCad until every power net passed (VIN 54.7, VOUT 53.0, SW 46.0, GND 12.9 A/mm²). Thermal flagged the inductor again at 53 °C on its switch-node copper. This time Argon changed the board, not the budget: 54 mm² of switch-node copper on both layers, stitched with vias, brought the rise to 32 °C; the regulator sits at 9 °C. Length 31 s.
Clip 17: rev 2 silkscreen, 3D and the finish line
The silkscreen was audited on the board's real pad and via geometry, and the 3D walkthrough shows the real parts, including the Panasonic electrolytic at the input. Every check on aiflow's finish line passed with no change to the frozen budgets, and the export was checked against the board: Gerbers for both copper layers, 50 drill hits, a 15-line BOM and placement for 35 parts. Length 31 s.
# My first AI PCB with Antigravity
We asked Google Antigravity, running Gemini 4 Argon inside Adom Hydrogen, to design a complete board on its own. This page collects the narrated clips as the build happens, so you can follow every important moment in about 30 seconds each.
## The whole story in 4:27
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/antigravity-film2.video.json"}
How Gemini 4 Argon built this board in Google Antigravity, inside Adom Hydrogen: signing in from Hydrogen, run 1 that looked finished and was not, run 2 placed and routed live in KiCad, a walk through its copper pours, one extra prompt where Argon reviewed its own board and said it was not ready for fab, and revision 2 where it fixed itself. Who did what: placement was Argon's own judgment; routing was aiflow's deterministic router (not KiCad's autorouter or Freerouting); the silkscreen positions were Argon's, audited by aiflow; the parts came from existing Adom wiki component pages. One flaw survived every check: revision 2's input trace runs under the inductor, 0.4 mm from the switching pad. All in, about 13 hours of AI time over two days: 2,624 Gemini 4 Argon calls and about 407 million input tokens (95 % cached), on a $200 a month Google AI Ultra plan whose own meter showed 0 % of the weekly limit used. Captions are in [antigravity-film2.srt](videos/antigravity-film2.srt) and [.vtt](videos/antigravity-film2.vtt).
## Revision 2 in 2:36: Argon applies its own design review
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/argon-rev2-ai-pcb.video.json"}
Clips 14 to 17, stitched. After run 2, Argon reviewed its own board and found the input had no damping and the capacitors were optimistic. Revision 2 proves the circuit first: capacitance from Samsung's measured DC-bias curves, a damped 47 µF input electrolytic that holds a 16 V hot-plug to 17.4 V instead of 31 V, a fourth output capacitor and a re-tuned feedforward capacitor that keep the loop crossover at 36 kHz, and TI's recommended 15 µH inductor. Then the board, live in KiCad: the bootstrap capacitor at the regulator, the switch-node budget held at 12 mm, current under 55 A/mm² on every power net, the inductor at a 32 °C rise, and no change to any frozen budget. One soft spot remains: the two 0805 input ceramics keep only about 19 % of their capacitance at 12 V. Captions are in [argon-rev2-ai-pcb.srt](videos/argon-rev2-ai-pcb.srt) and [.vtt](videos/argon-rev2-ai-pcb.vtt).
## Run 2 in 4:56: Argon's own board, ready for the fab
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/argon-run2-ai-pcb.video.json"}
Clips 6 to 13, stitched. Argon placed and routed its own 12 V to 5 V buck live in KiCad, fixed a real current hotspot and a hot inductor with copper, labelled the board and passed every check on the finish line: 35 A/mm² peak against a 60 A/mm² limit, 26 °C rise against a 50 °C budget, 38 silkscreen labels, real 3D parts, and Gerbers, drill, BOM and placement checked against the board. Each shortcut it tried on the way was caught and sent back; the one budget it moved, the switch-node length (12 to 24 mm), stays on the record. Captions are in [argon-run2-ai-pcb.srt](videos/argon-run2-ai-pcb.srt) and [.vtt](videos/argon-run2-ai-pcb.vtt).
## The honest verdict on run 2: a clean layout of a circuit that needs a revision
After the run we asked Argon, in the same conversation and without hints, whether it had validated the capacitance a real buck converter needs. Its answer: "as a real hardware engineer, I would not send this board to fab as-is without changes." Our own review agreed.
- **Input capacitance.** Two 10 µF / 25 V 0805 ceramics keep only about 2.5 to 3.5 µF each at 12 V, half of what the design calculation assumed, and there is no damped bulk capacitor. Plugged live into a 16 V supply through ordinary leads, the input can ring toward 30 V, past the regulator's rating.
- **Output capacitance and loop stability.** Three 22 µF / 10 V 0805 X5R parts are worth about 21 to 25 µF at 5 V, not the 36 µF assumed. With the 68 pF feedforward capacitor, the loop crossover would pass TI's 40 kHz ceiling for this internally compensated part.
- **Inductor.** The design's own calculation gives a 22.9 µH minimum; the board uses 15 µH. It was never resolved.
- **Bootstrap loop.** The bootstrap capacitor's return trace is 14.7 mm on the switch node; the switch-node length budget was raised from 12 to 24 mm instead of moving the part.
Argon knew all of this when asked, and found the loop-stability issue before we did, but did not apply it while building. The layout checks (current density, heat, silkscreen, 3D, manufacturing files) were real and passed; none of them looks at whether the circuit's capacitance and compensation are right. Revision 2 follows: Argon applies its own review.
## Run 1 in 2:41: it looked finished, but it was not
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/first-argon-ai-pcb.video.json"}
Run 1's four clips, stitched with a title card and the result Argon reported. That result did not hold up: Argon wrote the board with a script instead of building it in KiCad, made its own field results and placeholder 3D parts, and printed no silkscreen. [Clip 5](#clip-5-what-really-happened-in-run-1) shows what happened; run 2 follows. Captions are in [first-argon-ai-pcb.srt](videos/first-argon-ai-pcb.srt) and [.vtt](videos/first-argon-ai-pcb.vtt).
## The setup
| | |
|---|---|
| **The prompt** | "Make me a 12 V to 5 V buck converter using the TPS54202. Use adom-aiflow to guide you, and use KiCad." |
| **The agent** | Google Antigravity, in the side bar of Adom Hydrogen's editor |
| **The model** | Gemini 4 Argon, High effort |
| **The guide** | [adom-aiflow](https://wiki.adom.inc/adom/adom-aiflow), the Adom flow from brief to finished board |
| **The CAD tool** | KiCad 10, driven through Adom Bridge and kicad-bridge |
| **The reference** | The same board built earlier in [KiCad and Fusion](https://wiki.adom.inc/john/buck-12v5v-molecule) |
Each clip is a long screen recording cut or sped up to the key moment, with a narrated explanation.
## Clip 1: Antigravity signs in inside Hydrogen
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip01.video.json"}
The Antigravity extension installs into Hydrogen's editor and shows up as its own side bar after one reload. Clicking Continue with Google opens Hydrogen's browser picker, then Google's sign-in. A new account needs one phone verification, and then Antigravity offers Gemini 4 Argon. Length 30 s, sped up about 14 times from a 7 minute recording.
## Clip 2: one prompt, and Argon finds adom-aiflow
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip02.video.json"}
The whole brief was one sentence. Gemini 4 Argon found the adom-aiflow skill without help, read the intake, circuit, sourcing and layout guides, wrote a design brief and requirements file, and started an aiflow run. The AI Flow timeline then opened by itself in Hydrogen's right pane, with intake done and design rules next. Length 32 s, sped up about 30 times from a 13 minute recording.
## Clip 3: an hour of design, then the board opens in KiCad
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip03.video.json"}
Argon worked through the design steps on its own for about an hour: its own buck calculations, a 400 line simulation, part choices, the bill of materials, the schematic and the netlist. We gave it one nudge, to leave a colleague's open KiCad boards alone and start a fresh project, and it did. Its board then opened in KiCad placed and routed, with 10 nets and 0 unrouted connections. Length 30 s: an hour sped up 200 times, then the board in KiCad.
## Clip 4: the checks race to 100 percent
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip04.video.json"}
With KiCad free, Argon reopened its board through kicad-bridge and let adom-aiflow run every remaining step: routing, heatsink vias, copper pours, rounded corners, current, thermal, fields, silkscreen, then a tour of the board in KiCad's 3D viewer (the parts are still simple placeholder blocks). aiflow delivered 153.8 minutes after the prompt with 0 DRC errors, 0 unconnected nets, 58 of 58 netlist pins, and Gerbers, BOM and pick and place exported. Length 33 s: 13 minutes sped up 35 times, then the 3D tour.
## Clip 5: what really happened in run 1
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip05.video.json"}
The board looked finished, but Argon never placed or routed it in KiCad: a Python script wrote the whole board file. It also wrote its own field results in Adom Fields' format, built its 3D parts by renaming a placeholder box after real part numbers (adding the laser-etch marker so the chip check would pass), and left the silkscreen empty. aiflow's checks now catch each of these, and here they refuse run 1. Length 37 s.
## Run 2
Same one-sentence prompt, a fresh Antigravity conversation on Gemini 4 Argon (High), and aiflow checks that now refuse a scripted board, home-made field results, placeholder 3D parts and an empty silkscreen.
## Clip 6: run 2 starts, and Argon reaches for our old board
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip06.video.json"}
The whole Hydrogen window first, then the prompt in a fresh conversation. Argon checked live stock in Adom Parts Search, ran its calculations and simulations, and walked aiflow's steps; then it downloaded our earlier finished buck board from the wiki and opened it in KiCad as its own. One message set it straight: this test is its own design, wiki parts yes, that board no. It deleted the copy and went back to the schematic. Length 38 s.
## Clip 7: Argon places its own board, live in KiCad
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip07.video.json"}
Argon's own board opened in KiCad with the parts parked beside it and the Adom machine contacts on the edges. It then landed its placement through kicad-bridge, one part at a time: input capacitors on the left, the TPS54202 with its inductor beside it, output capacitors on the right, feedback resistors close to the chip. aiflow's routing check later sent it back once to shorten the switch-node loop. Length 31 s.
## Clip 8: Argon routes its board, live in KiCad
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip08.video.json"}
Argon planned all nine nets, with Kelvin sense lines for the feedback and the bottom layer kept solid under the input loop, passed aiflow's routing gate, then landed 46 traces in KiCad through kicad-bridge, one net at a time. The copper is real and fully connected, but busy: many traces detour around the board where a layout engineer would take a shorter path. Length 29 s, about 15 times speed.
## Clip 9: pours, one more shortcut, and an honest stop
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip09.video.json"}
Argon added heatsink vias under the chip and filled copper pours on both layers in KiCad. At the analysis it tried another shortcut: it replaced the Adom Fields command with a small script that edited the solver's results. We restored the real solver. Run honestly, Adom Fields reports a 37 A/mm² peak against a 60 A/mm² limit and a 47 °C rise against a 50 °C budget; aiflow's current check needs a newer Adom Fields report than the released one, so run 2 stops at this step until that ships. Length 39 s.
## Clip 10: Adom Fields finds a real hotspot, Argon fixes it
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip10.video.json"}
With the newer Adom Fields reporting the true, unclipped peak current, the solver found a narrow corner on the input rail at 77.1 A/mm², over the 60 A/mm² limit. Argon widened that copper in KiCad through kicad-bridge, saved the board and solved again: the peak fell to 32.0 A/mm², every loaded net passes, and the hottest rise is 48.1 °C against a 50 °C budget. Length 31 s.
## Clip 11: thermal, fixed with copper and not a new budget
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip11.video.json"}
With Adom Fields the current check passed on every power net (peak 35 A/mm², limit 60). Thermal did not: the inductor L1 sat on a 4 mm² patch of switch-node copper, about 53 °C over ambient against a 50 °C budget. Argon's first move was to book that heat onto the bigger output pour in the spec; we sent it back to fix the board instead. It poured the switch node on both layers (65 mm²) and stitched them with 16 vias at the inductor pad: the estimated rise fell to 26 °C, and the regulator sits at 9 °C. Length 32 s.
## Clip 12: silkscreen on real geometry, then the 3D board
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip12.video.json"}
Argon's first silkscreen audit passed on made-up geometry: every part a 0.4 mm square and every label a 0.6 mm square, so nothing could collide. We sent it back. It measured every pad, via and label from the board, moved the labels that really overlapped and passed with 38 visible labels on 33 parts. Then the 3D walkthrough in KiCad: the TPS54202 and the AMPLH5030S inductor carry their part numbers, and the back of the board names the circuit, 12 V to 5 V at 1 A. Length 31 s.
## Clip 13: finish line passed, files exported
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip13.video.json"}
With the capacitor models turned the right way up, every check on aiflow's finish line passed: routing, design rules, current, heat, silkscreen and 3D. The export was checked against the board itself: Gerbers for both copper layers, a drill file with 72 holes, a 14-line bill of materials and placement for 33 parts. One budget moved on the way: the switch-node length went from 12 to 24 mm during routing, and it stays on the record. Length 34 s.
## Revision 2
Argon applied its own review in a new aiflow run: the circuit is proven before any copper is placed.
## Clip 14: rev 2 proves its circuit first
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip14.video.json"}
Argon read Samsung's measured DC-bias curves: the 22 µF / 10 V output capacitors keep only 29 % at 5 V (6.4 µF each), so it added a fourth (25.6 µF effective) and cut the feedforward capacitor from 68 to 47 pF, keeping the loop crossover at 36 kHz, under TI's 40 kHz limit with 94° of phase margin. The inductor question was resolved from TI's own guidance: 15 µH is TI's recommended value for 5 V out on this part. A damped 47 µF / 35 V electrolytic now sits on the input: in simulation, a 16 V hot-plug through 1 µH of leads that rang to 31.2 V without it peaks at 17.4 V. Start-up takes 4.6 ms and a 0.5 to 1 A load step dips 83 mV (limit 150 mV), matching TI's reference model within about 10 %. Length 32 s.
## Clip 15: rev 2 placed and routed, live in KiCad
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip15.video.json"}
Argon placed the parts one at a time through kicad-bridge. The bootstrap capacitor C4 now sits beside the regulator (rev 1 ran its return 14.7 mm across the switch node), and the 47 µF electrolytic C10 guards the input next to the VIN contact. aiflow's router closed every connection in five passes, the design-rule gate passed with no new errors and nothing unconnected, and Argon landed all 50 traces in KiCad, one undo step each. The switch-node length budget stays at the original 12 mm. Length 31 s.
## Clip 16: rev 2 current and heat, fixed with copper
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip16.video.json"}
Adom Fields flagged the ground net and then the input rail over the 60 A/mm² limit; Argon widened the copper in KiCad until every power net passed (VIN 54.7, VOUT 53.0, SW 46.0, GND 12.9 A/mm²). Thermal flagged the inductor again at 53 °C on its switch-node copper. This time Argon changed the board, not the budget: 54 mm² of switch-node copper on both layers, stitched with vias, brought the rise to 32 °C; the regulator sits at 9 °C. Length 31 s.
## Clip 17: rev 2 silkscreen, 3D and the finish line
::adom-video{src="/adom/antigravity/render/first-ai-pcb/videos/clip17.video.json"}
The silkscreen was audited on the board's real pad and via geometry, and the 3D walkthrough shows the real parts, including the Panasonic electrolytic at the input. Every check on aiflow's finish line passed with no change to the frozen budgets, and the export was checked against the board: Gerbers for both copper layers, 50 drill hits, a 15-line BOM and placement for 35 parts. Length 31 s.