AI-Assisted Circuit Design: Workflows for the Idea → KiCad Schematic Handoff

Research for a solo hobbyist building a USB-PD modular-synth power supply in KiCad, driven by an LLM coding agent (Claude Code). Three approaches were already tried and rejected: letting the AI write .kicad_sch directly, schemdraw, and improving ASCII-art schematics. This page documents why those failed and what the field actually does instead.

The Core Finding: LLMs Can't Do 2-D, So Stop Asking Them to Draw

All three rejected approaches failed for the same root cause, unfixable by better prompting: LLMs cannot reliably encode geometry. Tokenization flattens 2-D into 1-D and self-attention has no native 2-D inductive bias. The literature measures a 13–27 point gap between an LLM's ability to read a spatial layout vs. produce one.

• ASCII art is unattainable by construction (a generation problem, not a prompt problem).

• Raw .kicad_sch is the worst possible target — explicit x/y/rotation, inverted Y-axis, 1.27 mm grid. The 2026 HWE-Bench benchmark for exactly this goal (spec → board schematic) has the top model passing only 8.15%.

• schemdraw is still a drawing API, so the layout burden stays on the human/model.

KiCad's lead developer Seth Hillbrand called natural-language circuit generation "one of the least helpful applications for LLMs out there." The whole field has converged on one architecture: the LLM emits coordinate-free connectivity; a deterministic engine (or the human, in KiCad) does the placement.

The real handoff is not "AI draws a schematic" — it is "AI emits a structured, checkable connectivity artifact; you render it."

The LLM's job is to emit connectivity; geometry is always the deterministic engine's (or the human's) job.

Family 1 — AI Drives KiCad Directly (MCP / IPC / In-Editor Plugins): Dead End for Generation

• The official KiCad 9 and 10 API is PCB-only; KiCad 10 shipped (April 2026) still without an Eeschema (schematic) API. No sanctioned way to author a .kicad_sch.

• kicad-sch-api is the one library that can create wires/symbols from scratch, but its own issues admit "LLMs struggle with inverted Y-axis" and placement is "unreliable." An independent test called a Claude-generated KiCad schematic "penguin on a bicycle level."

• In-editor chat plugins exist (ALT TAB Circuit Copilot; KiC-AI with local Ollama; a community KiCad AI Assistant) — but they are Q&A/assist, not reliable generators.

Useful subset: read-only MCP servers for interrogating a design; kicad-sch-api for scripted edits to schematics you already drew (bulk rename, stamp a repeated block).

Family 2 — Code-Defined Hardware That Exports to KiCad: Contains the Best Candidate

• circuit-synth (top pick) — Python (@circuit, net += pin) → a real hierarchical .kicad_sch + .kicad_pcb + netlist, then stops before layout. Mirrors the AI→schematic→manual-PCB split exactly; bidirectional, JLCPCB-aware, ships Claude Code slash-commands. Caveat: auto-placement is the weak link (topologically correct, not pretty); its ERC is incomplete and would not have caught the v3 STUSB4500 pin-18 bug.

• SKiDL — mature Python→netlist+ERC; its author says skip its schematic gen and go netlist→PCB. Best use: an ERC oracle for the repeated 78xx/79xx LDO stages.

• atopile — most mature, but produces no readable schematic (netlist/ratsnest only) — bypasses the step we care about.

• Diode pcb/Zener — best verification story (compiler + TestBench + SPICE, Anthropic-partnered) but exports .kicad_pcb, not .kicad_sch.

• tscircuit — architecturally right but React/TS stack mismatch.

Family 3 — Better Text Representations + the Netlist Handoff: Lowest-Effort, Highest-ROI

The winner: a net table + a Mermaid block diagram. Drop ASCII art, keep a | Net | Connected pins (Ref.Pin) | Value/Note | table (one per sheet) plus a Mermaid flowchart TD for topology. Pure connectivity = zero geometry to get wrong = the LLM's most reliable output.

The repo's own v3-pd-failure-diagnosis.md proved this: a multi-agent pass produced false findings by guessing pin numbers from screen position, while the connectivity table made the analysis verifiable.

Optional docs upgrade: lcapy (1-D netlist + one direction word per part → clean rendered figure, guaranteed correct-to-netlist) — better than ASCII/schemdraw for figures, but needs a LaTeX toolchain and has no KiCad path.

Family 4 — Commercial AI EDA: Mostly Not Composable with a Terminal Agent

Flux Copilot / Circuit Mind / CELUS are walled-garden chat UIs (replace the agent, don't compose). The AI autorouters (Quilter, DeepPCB) target the layout stage done by hand.

Worth adopting regardless: Octopart (free KiCad parts + queryable API) and SnapMagic (free native-KiCad footprint/symbol generation).

Family 5 — Academic State of the Art: Steal the Patterns, Not the Tools

• A paper on switched-mode power supply design with LLMs (the exact domain): value/parameter tuning went 14% (GPT-4o alone) → 85% (o3 + SPICE) → 91% (o3 + RAG + SPICE), but topology adaptation stays ~50%. Lesson: let the LLM size values for a topology you choose, in a SPICE loop; don't let it invent topology.

• "Understanding" is now strong (96%+ netlist-from-schematic-image, even hand-drawn) → sketch → digitize is a more reliable bet than generate-from-scratch.

• EEschematic (MIT) is a runnable implementation of the "generate → render → critique" visual-chain-of-thought loop.

Comparison of Top Candidates

Recommendations (A–D)

A — Adopt Now (Zero Setup): Net Table + Mermaid, Replacing ASCII Art

Per circuit, emit a Mermaid flowchart TD (topology) + a markdown net table (one per hierarchical sheet). Bake the convention into CLAUDE.md. Transcribe into KiCad from a diffable, reviewable, geometry-free artifact. Highest-ROI change.

B — One-Hour Spike: circuit-synth on the Linear-Regulation Sheet

Write it as circuit-synth Python, generate the .kicad_sch, and time placement tidy-up vs. hand-drawing ~27 parts. Keep it for well-understood sheets if it wins.

C — Standing Safety Net: a Datasheet-Aware Review Pass Before Every JLCPCB Order

Export the schematic to compact text, feed it + datasheets to the LLM with a power-supply checklist (VBUS sense paths, EN/feedback dividers, dropout headroom, polarity, abs-max). A datasheet-aware pass plausibly would have caught "pin 18 → R14 → GND instead of sensing VBUS." Reviewer beats generator.

D — High-Leverage: a SPICE Loop for Value-Sizing

The academic 14%→90% lever, runnable today in ngspice/LTspice, independent of any KiCad tool. For a noise-sensitive multi-rail supply this is arguably more valuable than circuit-synth.

Corrections and Caveats

• The 13–27 point read-vs-generate gap is from the Schemato paper (arXiv 2411.13899); the HWE-Bench 8.15% figure is from arXiv 2603.18102. Both measure a specific task difficulty, not a permanent ceiling — models will improve.

• circuit-synth's auto-placement produces topologically correct schematics but the visual result usually needs manual tidying. Time savings depend heavily on circuit complexity.

• The SMPS-with-LLMs paper (arXiv 2507.10639) measured value-sizing accuracy on a specific benchmark; real-world SMPS design involves additional constraints not captured in the benchmark.

• lcapy requires a working LaTeX toolchain (TeX Live or MiKTeX) — non-trivial on some platforms.

• kicad-sch-api is the only library that can round-trip .kicad_sch files reliably; it is maintained by the circuit-synth team.

Hands-on trial results (A–D)

These are the actual outcomes of running approaches A–D on this repo (wave-1 of the AI circuit design experiment, June 2026). Each subsection links to the artifact produced.

A — Net Table + Mermaid Convention (adopted)

What was tried: Defined a repo-wide convention for documenting circuit connectivity using a net-connectivity table (one row per net) paired with a Mermaid flowchart TD, then applied it to the linear-regulation sheet as the first worked example.

What worked: The convention was baked into the root CLAUDE.md (§ Schematic Documentation Conventions) and a dedicated how-to page was created: Net-Table + Mermaid Convention. The worked example covers all 27 parts of the three LDO stages, derived from a kicad-cli netlist export — no GUI, no geometry guessing. The table + Mermaid diagram accurately represent the linear-regulation sheet and are diffable in version control.

What was blocked: Nothing. Zero-setup, clean end-to-end.

Verdict: Adopted. Highest-ROI change of the four approaches. Replaces ASCII art as the canonical AI→human connectivity handoff.

B — circuit-synth spike (not adopted)

What was tried: Installed circuit-synth 0.12.1 on Python 3.14, modelled the linear-regulation sheet as a Python circuit description, generated a hierarchical .kicad_sch project, and validated it headlessly with kicad-cli v10.

What worked: Install succeeded on the first try (no compile, all wheels, ~1 min). generate_kicad_project() produced a valid hierarchical project (root sheet + three sub-sheets). kicad-cli v10 parsed each sub-sheet cleanly (exit 0); it read the v9-era format token without error. Netlist/connectivity is correct — each 9-part stage has the right refs and nets.

What was blocked / didn't work: The generated schematics have zero drawn wires. Connectivity is expressed entirely through 22 hierarchical-label stubs per sheet. Auto-placement is a flat left-to-right row of symbols — no signal-flow layout, no GND rail grouping. All symbols are generic (Regulator_Linear:*, Device:*), not the repo's zudo-pd: library parts with real LCSC part numbers. The format token is v9, not the repo's v10. Result: converting the output to repo quality requires drawing all the wiring by hand, re-symboling every part, and re-laying the placement — roughly the same effort as hand-drawing (~2–3 h), plus the re-symbol overhead.

Full findings: experiments/ai-circuit-design/circuit-synth-spike/FINDINGS.md

Verdict: Not adopted for this repo. The pipeline works end-to-end and is useful as a netlist/connectivity generator or ERC cross-check, but for ~27 highly-repetitive LDO parts it provides no time saving over hand-drawing in Eeschema.

C — Datasheet-aware schematic review skill (adopted)

What was tried: Ran a read-only netlist-topology review of the full zudo-pd schematic using kicad-cli export, parsed the 87-component netlist, and applied a power-supply checklist (VBUS sense paths, EN/feedback dividers, dropout headroom, polarity, abs-max). The workflow was then codified as a reusable .claude/skills/pd-schematic-review/ skill.

What worked: The review confirmed the v3 fix is present: VBUS_IN → R14 (470 Ω) → U1.18 with no GND on the pin-18 net (F-1). Feedback-divider math recomputed from the actual resistors matched the target rails to within 0.22 % (F-7). Load-switch polarity, VDD decoupling, CC Rd pull-downs, and VREG decaps all checked out.

What was flagged:

• Both the +12V (L7812, F-11) and −12V (CJ7912, F-13) LDOs run at only ~1.5 V dropout against a ~2 V typical spec — marginal at rated current. The existing components/l7812cv.md doc already flags this.

• Documentation discrepancy (F-9): CLAUDE.md and the architecture text describe the negative rail as "−15V (LM2586SX-ADJ inverted SEPIC)," but the actual schematic uses a third LM2596S-ADJ in an inverting buck-boost topology — a stale description that should be corrected.

Full findings: experiments/ai-circuit-design/schematic-review/REVIEW-REPORT.md

Verdict: Adopted as a standing safety net. A datasheet-aware net-topology pass would plausibly have caught the v3 pin-18→GND bug before ordering. The skill is now available to re-run before any future JLCPCB order.

D — SPICE value-sizing loop (delivered, blocked on ngspice availability)

What was tried: Demonstrated the generate → simulate → read-back → adjust value-sizing loop on the LM2596-ADJ feedback divider (the +7.5 V intermediate rail, R3 = 5.1 kΩ / R4 = 1 kΩ). Two SPICE decks were written: a .op forward + loop-closed check, and an R3 sweep deck that prints the resulting rail voltage for a range of E24 values.

What worked: The algebra was verified and the worked result is documented. For the as-designed +7.5 V rail the divider gives 7.503 V (error +0.04 %). The re-targeting demo (13.5 V rail) shows the E24 pick: R3 = 10 kΩ → 13.53 V (+0.22 %). The decks are complete and ready to run.

What was blocked: No command-line ngspice was available on the host. The KiCad bundle ships ngspice as a shared library for GUI use only (libngspice.0.dylib); there is no ngspice CLI executable. A brew install ngspice attempt was killed by the 180 s guardrail timeout and was not retried. The hand-computed read-back values are exact and serve as the documented expected output.

Full findings: experiments/ai-circuit-design/spice-value-sizing/FINDINGS.md

Verdict: The workflow is sound and the deck is production-ready. Run brew install ngspice once to unlock it. Extending to ripple/regulation or dropout-headroom checks uses the same loop structure with a behavioural model or the LM2596 SPICE subckt.

Sources

• HWE-Bench (spec→schematic, 8.15% pass rate): arXiv 2603.18102

• CircuitLM (ERC-clean vs. functionally-correct; fails analog/power): arXiv 2601.04505

• Schemato (position coords problematic; netlists carry no geometry): arXiv 2411.13899

• SMPS-with-LLMs (14%→91% with SPICE+RAG): arXiv 2507.10639

• AnalogCoder / AnalogCoder-Pro (simulator-in-the-loop): arXiv 2508.02518

• SINA / Masala-CHAI (schematic-image → netlist, incl. hand-drawn): arXiv 2601.22114, arXiv 2411.14299

• EEschematic (runnable VCoT loop, MIT): github.com/eelab-dev/EEschematic

• Hillbrand quote (KiCad devlist): groups.google.com/a/kicad.org/g/devlist/c/4hjJIFbBtXU

• circuit-synth: github.com/circuit-synth/circuit-synth

• kicad-sch-api: github.com/circuit-synth/kicad-sch-api

• SKiDL: github.com/devbisme/skidl

• atopile: github.com/atopile/atopile

• Diode pcb: github.com/diodeinc/pcb

• tscircuit: github.com/tscircuit/tscircuit

• lcapy: lcapy.readthedocs.io

• Octopart: altium.com/octopart

• SnapMagic: snapeda.com