Guide

Text to schematic, command by command

What actually happens between typing a description of a circuit and getting an editable schematic — the exact commands an AI agent issues, and how it checks its own work.

Last updated 26 September 2026

"Text to schematic" sounds like a single leap. It is not. It is a sequence of small, ordinary editor operations, each of which either succeeds or refuses, and the quality of the result comes almost entirely from what happens between them.

This page opens the box. Everything below is a real command from the Schema command reference, and the agent has no private ones.

The loop, not the leap

The agent runs a perceive / act / verify / recover loop rather than emitting a finished design in one shot:

Perceive — describe returns the sheet as JSON: every symbol, pin, net and connection. This is the agent reading the design rather than remembering it.
Act — place_symbol, connect, draw_wire, place_power and the rest change the sheet. Each reports what it actually did.
Verify — validate and nets say whether the change had the effect that was intended.
Recover — undo steps back when a check fails, so a wrong step does not become the foundation of the next one.

Why refusals matter more than successes

Commands refuse rather than approximate. place_symbol will not silently pick a different reference designator if the one you asked for is taken, and connect refuses a net name that matches more than one net instead of guessing. A refusal is information the agent can act on; a silent wrong guess is a bad netlist nobody notices.

A worked example

The prompt:

> add a 100nF decoupling cap on the 3V3 rail next to U1

What the agent issues:

describe
  → U1 at (100, 80), 3V3 net present, nearest free space (112, 74)

place_symbol Device:C C7 auto 112 74
  → placed C7

set_property C7 Value 100nF
  → set

connect C7 1 3V3
  → C7.1 joined 3V3 via power symbol

connect C7 2 GND
  → C7.2 joined GND via power symbol

validate
  → 0 errors, 0 warnings

Six commands. The first is a read, the last is a check, and only four of them change anything. That ratio is the point.

Where things end up

Coordinates are in millimetres on a 2.54mm grid, with the canvas origin at the centre and positive Y downward. The agent reads existing positions out of describe rather than estimating them.

One exception worth knowing

place_net_label takes X and Y in grid units rather than millimetres — they are multiplied by 2.54 before use. It is the only command in the reference that does this, and it is documented on its own page for exactly that reason.

What this is good at, and what it is not

Good: conventional blocks with well-known topology — decoupling networks, crystal load capacitors, linear regulators with their caps, pull-ups on a bus, a debug header.
Good: the tedious accurate work — wiring a 32-pin part to the right rails without transposing two of them.
Not good: novel analogue design, tight tolerance stacks, anything where the right answer depends on a datasheet curve.
Not good: deciding what the circuit should be. It builds what you describe.

Frequently asked questions

Is a text-to-schematic result actually editable?

Yes. The agent issues the same commands you would, so what you get is an ordinary schematic — you can move parts, rewire nets, change values and extend it by hand. There is no generated-mode to exit.

What is the difference between this and an AI diagram tool?

A diagram tool draws a picture that resembles a schematic. This produces a netlist: real symbols with real pins and connections that can be checked with an electrical rules check and turned into a board.

Can I see what the agent did?

Yes — the commands are visible as it works, and each one reports its result. That is also why you can undo any of it.

Which parts can it use?

Whatever is in your loaded libraries. Schema reads KiCad symbol libraries and Altium schematic libraries, so the agent draws on the parts you have already vetted rather than a fixed catalogue.

More guides

GuideHow to make a PCB schematic with AIA step-by-step walkthrough: describe a circuit in plain English, have an AI agent place the parts and wire the nets, check the result, assign footprints and push it through to a board.GuideHow to import a KiCad libraryPoint Schema at KiCad symbol and footprint libraries you already maintain — how discovery works, how to add directories by hand, and how load order resolves name clashes.GuideAI PCB routing explainedHow an AI agent actually routes a PCB — the commands it issues, how collisions are handled, why it never rewrites copper it was not asked to touch, and where it hands back to you.

Comparisons

Try it on your own board

Free to start on macOS, Windows and Linux. Reads your existing KiCad and Altium libraries.