17 prompts, 53 minutes

Last week I wrote that I had built an image-processing tool for an audience of one — a solder paste stencil pre-processor, built in a field I have never worked in. I said domain knowledge had become the scarce half of the job.

That was an assertion. This is the evidence.

I still have the transcript. The first session runs from 07:36 to 08:29 — fifty-three minutes, seventeen typed prompts, empty directory to working tool with seventeen commits in between. The median prompt is nine words long. A second session the following week added one more rule.

The speed is not the interesting part. The interesting part is what is missing from those seventeen prompts.

What I never typed

I never typed “solder bridge.” I never explained that fine-pitch leads sit fractions of a millimetre apart, that a 1:1 aperture deposits enough paste to short them on reflow, or that the fix is to reduce the aperture rather than remove it.

I also never gave a single dimension for a SOT-23. Not a pad size, not a pitch, not how many pads sit on which side.

Here is the entire prompt that added SOT-23 support:

one more shape group we should shrink, SOT-23-X (X=3/5/6)

Thirteen words, one of which is a variable. What came back was a detector that knew a SOT-23-3 is two pads side by side with a third centred opposite, that the -5 is three and two, that the pitch is around 0.95 mm, and — the part I would never have thought to specify — that a three-pin transistor never forms a row of three collinear pads, so the existing fine-pitch rule was structurally incapable of seeing it. It needed its own rule.

I was not describing a problem. I was handing a parameter to an understanding that was already there.

Six prompts, in order

Step through the ones that mattered. Five are from the first session; the last is from a follow-up the next week. Each shows what I typed, what the tool gained, and — the panel worth opening — what it had to already know for those words to mean anything.

Typos are left in. They are evidence: these were typed quickly, between looking at a picture and looking back at the screen.

The shape of the split

Read those in order and the division is consistent enough to state plainly.

Everything I contributed was judgment about my own situation: which packages are on my boards, what a hatch cell should measure for my laser, which pitch counts as fine. Plus the reviewing eye — prompt 12 exists only because I looked at an image and something seemed off.

Everything else came from the other side. Pad geometries, package conventions, why through-hole pins take no paste, the failure mode at each end of the pitch range, and the structural insight that some footprints cannot be found by the rule that finds the others.

I could not have written that second list. I did not need to.

The part that does not automate

Two of the seven rules exist because I caught something by eye. Prompt 12 is one. The other was a false positive on the action map — pads being removed that should not have been.

Which means the colour-coded review image, which I asked for in the very first rules prompt almost as an afterthought, was load-bearing. Without something that showed me every decision at a glance, I would have had no way to notice the two-pad side was wrong, and no way to trust the output enough to cut with it.

If you take one practical thing from this: build the thing that shows you what it did, first. It is what makes the rest correctable.

Seventeen prompts

That is the real barrier now — not skill, and not time. Seventeen short messages with a median length of nine words, eleven of them under fifteen, several containing typos, none containing a specification.

I did not start with a plan. I started with three rules and a screenshot of a board, and corrected it when it was wrong.

If you have been putting something off because it seemed like it would need a field you have never studied — that is no longer the thing standing in the way. Try three rules and a photo.

---

The tool is on GitHub at MansukhPatidar/stencil — seven rules, 90 tests, and the sample board it was developed against. All twenty-five prompts, including the nineteen not stepped through above, are in PROMPTS.md.