The Same Conversation, Twelve Times

About six months into the renovation, I asked the AI to review my conversation history and tell me what I keep asking about. Not out of curiosity — out of frustration. I'd noticed that I was spending 10 to 15 minutes in a chat window every time I needed to figure out how many slats fit across a span, or what 476 millimeters is in fractional inches, or what angle to set a miter gauge for a railing joint. The math was not hard. The conversations were not deep. But I kept having them, because I didn't have the formulas memorized and my phone's calculator doesn't do fractional inch display.

The AI went through about sixty of my past conversations and came back with a ranked list of recurring patterns. Number one, by a wide margin: woodworking math. Spacing calculations, unit conversions, angle computations, cut lists. The inputs changed every time. The logic never did.

I had been using the most powerful language model available to humanity to repeatedly compute what 3/4 plus 2-1/2 is, twenty times across a 67.5-inch span, with edge offsets.

The pattern

The conversations followed the same shape every time. I'd describe the physical situation: a header that's 67.5 inches long, slats that are 3/4 of an inch thick, a desired gap of 2.5 inches, with 2.5 inches of space at each end. The AI would do the arithmetic, return the number of pieces, the gap size, the position of each element, and sometimes a story stick diagram showing where every mark goes.

Then I'd change a number — 68 inches instead of 67.5, or 21 slats instead of 20 — and ask again. The AI would recompute everything from scratch, because every conversation starts from zero, and I'd copy the numbers into a note on my phone and walk back to the garage.

For the Gridfinity drawer baseplates, the conversation was the same: 476mm wide, 558mm deep, 42mm grid, how many columns and rows fit, what's the remainder, where does the padding go. For the stair rail angles: 36-degree pitch, what's the miter cut. For the sewing desk cut list: how many 4-by-8 sheets do I need for a torsion box with ribs at 6-inch spacing.

Same structure. Same math. Different numbers. Every time.

When a conversation should become a tool

There's a principle in software engineering that if you do something three times, you should automate it. The threshold is lower than people think — it's not really about the labor saved on the third repetition. It's about the error rate on the twelfth. Every time I re-entered the same kind of problem into a chat window, I was re-introducing the possibility of describing it wrong, interpreting the output wrong, or transcribing the numbers wrong onto the piece of wood.

A conversation is a terrible interface for a solved problem. Conversations are good for ambiguous situations where you need to think through trade-offs, explore options, and refine your understanding. Spacing twenty slats across a header is not ambiguous. It's arithmetic with five inputs and one output. The AI was doing the work of a calculator while consuming the bandwidth of a consultant.

So I stopped having the conversation and started writing the spec for the tool that would replace it.

What the AI built for itself

I asked Claude to write the PRD and implementation specification for a Raycast extension — a local app launcher that lets you type a few words and get a result — that would handle the four most common calculation types: spacing, unit conversion, joint angles, and cut lists.

The irony of this was not lost on me. I was using the AI to design the tool that would eliminate the need to talk to the AI. The conversations about slat spacing became the test cases for the spacing calculator. The drawer baseplate math became the validation data. The stair rail angle conversation produced the reference values that the angle module would need to match.

The spec that came back was precise. Four pure calculation modules in TypeScript, zero dependencies on the host application, callable from either a form-based UI or natural language input. The key architectural decision: all math lives in a library layer that has no knowledge of how it's being invoked. The same function that powers a Raycast command can power a web app, an iOS shortcut, or a CLI tool. The interface changes. The math doesn't.

This is exactly the insight the AI helped me reach about jigs, six months earlier in the stair slat project. A jig is the decision to stop measuring and start indexing — to do the hard thinking once and encode it into a physical object that makes the answer repeatable. ShopCalc is a digital jig. The hard thinking was done in the conversations. The tool encodes it so I never have to re-derive it.

The test cases tell the story

The spec included a validation table. Every test case was pulled from a real conversation about a real project:

67.5-inch span, 20 slats at 3/4-inch width, 2.5-inch edge offset: expected gap of 2.5 inches. Source: stair slat spacing, November.

476mm span, 11 columns at 42mm width: expected standard Gridfinity grid. Source: drawer baseplate, February.

36-degree stair pitch: expected miter angle of 54 degrees, complement of 36 degrees. Source: railing-to-post joint, November.

476mm converted to inches at 1/16 precision: expected 18-3/4 inches. Source: Gridfinity drawer measurement, February.

Each line in the validation table is a conversation I had, a project I built, and a number I copied onto a piece of wood or typed into a 3D printing slicer. The test suite for the tool is the history of the renovation.

What this means for AI and making

The discourse about AI tends to assume a stable relationship: either AI does the work or you do. The reality, at least for physical making, is that the relationship evolves. You start by having conversations because you don't know enough to ask a precise question. The AI helps you learn. You learn enough to recognize the pattern. Then you extract the pattern from the conversation and encode it as a tool that runs without the AI.

The AI's role shifts over time from consultant to specification. The first conversation about slat spacing was exploratory — I didn't know what a story stick was, I didn't know the right formula, I needed the AI to walk me through it. The twelfth conversation was rote. I knew exactly what I needed; I just needed the numbers. The tool exists because the conversations taught me enough to know what the tool should do.

This is the opposite of the replacement narrative. The AI didn't replace my skill. It built my skill to the point where I could replace the AI — for this specific class of problem, in this specific domain. The hard problems, the ambiguous ones, the ones where I need to think through trade-offs and ask questions I don't know how to ask — those stay in conversation. The solved problems graduate to tools.

The residue

MeasureTwice — the name the project eventually took — now lives as a web app. Four calculators. Natural language input. Fractional inch display to 1/64 resolution. An interface designed for a phone screen held in one hand while the other hand holds a pencil, standing in a garage, with sawdust on the screen.

It's a small tool. It does simple math. Nobody looking at it would guess that behind each calculator is a conversation about a specific piece of wood in a specific room of a specific house, and that the conversation taught me something I didn't know before I had it, and that the thing I learned is now encoded in a function that runs in three seconds and never forgets.

The conversations are gone. The tool remains. The slats are on the wall. The baseplates are in the drawer. The railing is at the right angle. And I don't need to ask anymore — not because the questions got easier, but because the answers became mine.