daishuge/pcb-skill/pcb

PCB

Take a hardware idea to a board a human can actually order, solder and bring up.

This skill is written from a real project that went from a blank page to a released 4-layer board. Most of what is here is not PCB theory — it is the specific set of failures that theory does not catch, and the gates that catch them.

The three principles this skill optimises for

1. 性价比 — cost-effectiveness. Every part is chosen on measured price and real

availability, netted against stock the user already owns. A price nobody has read off a live page is an ESTIMATE and must be labelled as one.

1. 好装配 — assemblable by hand. The user solders this. No QFN/DFN/exposed-pad-only

packages unless the user says otherwise. Every pad reachable with an iron, in an order that does not trap a joint behind a tall part.

1. 免费 PCB — the board should cost nothing. JLC's free-prototype coupon has hard

geometric and process limits. Design inside them from the first sketch, not at the end.

Before you start — the environment must be ready

Do not begin design work until these are true. Check, do not assume.

• EasyEDA Pro running, with the MCP bridge installed and answering (server_info).

See setup/easyeda-mcp.md.

• A browser with the agent extension, logged in to the sites the project needs:

Taobao, JD, JLC/LCSC. See setup/browser-logins.md.

• An approval watcher if the platform pops permission cards for every browser action,

or the agent blocks on a card it cannot press itself. See setup/auto-approve-*.md. Carts and payment stay manual — the watcher is deliberately bypassed there.

• A notification channel for long runs (scripts/notify/), so silence is never

ambiguous.

The workflow — four phases, each with a gate

Read references/01-workflow.md for the full version. In outline:

Never place an order or pay. Drive to the pre-payment page and stop.

The rules that actually caught the bugs

These are load-bearing. references/02-review.md explains each with the failure it caught.

• A green board is not a correct board. Every phase ends with a one-pass, read-only,

adversarial review by a fresh reviewer that re-derives the numbers instead of reading them. Write its termination criterion before it starts, or it will loop.

• Measured, not inferred. "X probably does not support Y" is not a finding. If a claim

decides whether work continues, prove it. Say which numbers were measured and which were computed.

• A wait condition must answer: if this crashed right now, would my filter emit a line?

Watchdogs cover four terminal states — output present, crash signature, process gone, wall-clock timeout — plus a progress-rate floor.

• Verify the tool before trusting the result. Three separate checkers on this project

silently measured the wrong thing (a cached 3D model, a bounding box instead of a pad, a merged copper label). Calibrate every geometric transform against a known asymmetric object.

• The last honest step is the user's. Logins, CAPTCHAs, and payment are theirs. Prepare

everything else so their part is two minutes.

References — load what the current phase needs

Scripts

scripts/ holds the checkers this workflow depends on. They are deliberately independent of the EDA: they parse exported documents and Gerbers, so they can contradict the tool that produced them. Each directory has its own README.

• scripts/placement/ — courtyards from real pads, via-lane widths, module-body clearance
• scripts/routing/ — the DSN/SES autoroute chain with a four-state watchdog and a rate floor
• scripts/verify/ — netlist assertions, Gerber parsing, drill census, 3D interference
• scripts/notify/ — a progress relay so a long run is never silent