Digitizing Machine Shop Work Instructions and Drawing Notes with Claude Code

A third-year kid was standing at the lathe, gripping his phone, staring at it. In the margin of the drawing were pencil notes the old shift lead had left: “extra oil here,” “chamfer about 0.3, see how it goes.” He could read the words. What he couldn’t figure out was why. The shift lead had retired and walked out the door last month.

When I started helping small machine shops with floor-level improvements, this was the complaint I heard more than any other: “All the setup knowledge lives in people’s heads.” Work instructions technically existed. But they’d been typed up ten years ago in some ancient word processor, and nobody knew which folder the file was in. So in practice, the floor ran on paper and the memory of a few veterans.

Every time a veteran quits, one more setup disappears from the shop. This isn’t a problem you fix with a pep talk. It’s a records problem. Today I’m going to walk through how to take that “knowledge in someone’s head” and those “notes in the drawing margin” and turn them into something that survives, using Claude Code and generative AI, based on what I’ve actually tried on the floor.

Key takeaways

• Machine shop work instructions never get written down not because there’s no time to write, but because turning spoken explanation into clean prose is heavy lifting. Hand that part to AI.
• Record a veteran doing the job, transcribe it, then have Claude Code shape it into a work-instruction template. That flow trips people up the least.
• Margin notes on a drawing become durable knowledge in three steps: photograph, transcribe, then ask “why do you do it that way?” and append the answer.
• Let AI handle formatting, first drafts, and flagging gaps. A human always checks anything touching dimensions, tolerances, or safety.
• Keep drawings and customer names off the public internet by design. Skip this and it becomes a trust problem.

What’s actually happening on the floor

The reader I have in mind runs a shop with somewhere between 5 and 50 employees, working as a plant manager or floor lead. Metal cutting, stamping, plastic molding, assembly, whatever. The common thread: Figma and the cloud feel like another planet, and the floor runs on Excel, paper, and the occasional fax.

Here’s the typical workflow:

1. A drawing comes in from the customer (PDF, sometimes paper)
2. A veteran reads the drawing and decides the setup (material, fixtures, machining order, things to watch)
3. That setup gets passed to the floor verbally or as a handwritten note
4. The part gets machined. If something goes wrong, the veteran rushes over
5. Done. Next time the same part number comes in, someone has to remember the setup all over again

The problems are steps 3 and 5. The setup evaporates into “a conversation that happened once.” So when you make the same part six months later, you reconstruct it from scratch. Or the only person who could reconstruct it is already gone.

Common rework and headaches

Here are the specific complaints I hear on the floor:

• “We ran this part number before, right?” — and nobody saved the procedure
• The margin notes are written in handwriting so fancy that only the author can read them
• Every time a new hire comes in, someone explains the same thing out loud
• When a veteran takes a day off, production stalls on a setup only they know
• The customer asks “what were the machining conditions back then?” and nobody can answer

Every one of these collapses into a single point: there’s no record. And there’s no record not because anyone is slacking. It’s because you don’t have time to write prose the same day with oil all over your hands. Start by admitting that honestly.

What to delegate to AI, and what a human must decide

Let me draw the line up front. Leave this fuzzy and you end up with the worst outcome: an AI-written procedure goes straight to the floor and causes an accident.

Step	Delegate to AI	A human must decide
Interviewing the worker	Transcribing the recording, pulling out key points	What to record, who to have explain it
Writing the procedure	Draft, structure, flagging gaps	Whether dimensions, tolerances, machining conditions are correct
Organizing margin notes	Transcribing handwriting, cleaning it up	Verifying the reading is right
Safety warnings	Suggesting generic cautions	The real hazards of this machine and this material
Approving the final version	Never	The floor lead always reviews it

The point fits in one line: AI is the “put it into words” department; the human is the “decide whether it’s right” department. Get a dimension wrong and you make scrap. Misread a tolerance and you cause the customer grief. So anything involving numbers gets a human’s eyes last. That part I don’t negotiate on.

Use case 1: Record a veteran and turn it into a work instruction

This is the one with the biggest payoff and the least friction. The method is simple.

Have the veteran narrate what they’re doing while they actually do the job. Record it with the voice memo app on a phone — that’s it. “First I chuck the stock, then I dial in the center…” Just have them say out loud the thing they always do. A 10-minute job is a 10-minute recording.

Transcribe that audio and hand it to Claude Code. You can use this prompt as-is:

You are responsible for writing technical documentation for a manufacturing floor.
Below is a transcript of a veteran machinist talking through a job as they worked.
Shape it into a "work instruction" template.

Rules:
- Number each step
- For each step, separate "what to do," "things to watch," and "fixtures/tools used"
- Do NOT change any number the veteran said (dimensions, RPM, feed, etc.). Keep it verbatim and tag it [VERIFY]
- Where the wording is vague, do NOT fill it in by guessing. Write "* needs confirmation"
- Keep technical terms as-is, and add a one-line note for new hires
- Output in plain text

Transcript:
(paste the transcript here)

The two things that matter here are “do not change numbers” and “do not fill in vague spots.” Out of helpfulness, AI will pad missing numbers with “probably around this much.” Let it do that on a manufacturing floor and it’s dangerous. So force the [VERIFY] tag every time, and a human clears them one by one afterward.

What changes before and after

Before: writing a work instruction was a job where “someone carves out time and writes prose from scratch.” So it stayed at the bottom of the list forever.

After: the veteran just “narrates while working as usual.” The heavy part — turning it into prose — is done by AI. All a human does is check and correct the numbers in the draft that comes out. Writing from zero versus correcting a draft are psychologically worlds apart in weight.

Use case 2: Digitize the notes in the drawing margins

“Extra oil.” “Chamfer, see how it goes.” Those handwritten margin notes are the most fragile knowledge there is. Swap the drawing for a new revision and the notes vanish with it.

Here’s the workflow:

1. Photograph the drawing with the margin notes on a phone (customer name and drawing number may be in frame — at this point it stays on your device only)
2. Feed that image to a multimodal Claude (via Claude.ai or Claude Code) and transcribe the handwriting
3. While reviewing the transcription, ask the veteran “why extra oil here?”
4. Append the reason and save it as a per-part-number “machining note”

Step 3 is the heart of it. A note that only says “extra oil” tells a new hire nothing. Add the reason — “this material is gummy and the chips tend to cling, so we run more oil” — and now it’s knowledge. The time you save by having AI do the transcription is exactly the time you reinvest in “asking why.”

Here’s a prompt template for cleaning up margin notes:

This is a handwritten note in the margin of a part drawing.
Write out exactly what you can read, with no guessing.
Where a character is illegible, mark it clearly as "(illegible)."
Then, for each note, add one "question to confirm with the operator."
Never fill in or complete any dimension or tolerance value.

Put “exactly as read” and “do not complete” in every time. Without them, AI will quietly convert fancy handwriting into “plausible correct characters,” and you’ll never catch the error.

Use case 3: Take inventory of the scattered old procedures

The third one is cleaning up the old work instructions you already have. In a lot of shops, Excel and word-processor procedures are scattered across folders. First you want to know what’s where.

A small check script helps here. It just lists the procedure files under a folder you specify and prints them in a table with their last-modified date. It runs if you have Node.js installed.

import { readdir, stat } from "node:fs/promises";
import path from "node:path";

// Point this at the folder where the procedures live
const root = process.argv[2] || ".";
// Extensions we treat as procedure files
const targets = [".xlsx", ".xls", ".doc", ".docx", ".pdf", ".txt"];

async function walk(dir) {
  const rows = [];
  for (const name of await readdir(dir)) {
    const full = path.join(dir, name);
    const info = await stat(full);
    if (info.isDirectory()) {
      rows.push(...(await walk(full)));
    } else if (targets.includes(path.extname(name).toLowerCase())) {
      const updated = info.mtime.toISOString().slice(0, 10);
      rows.push({ file: full, updated, kb: Math.round(info.size / 1024) });
    }
  }
  return rows;
}

const rows = await walk(root);
// Oldest first (the oldest procedures most need review)
rows.sort((a, b) => a.updated.localeCompare(b.updated));
console.log("Updated\tSizeKB\tFile");
for (const r of rows) console.log(`${r.updated}\t${r.kb}\t${r.file}`);
console.log(`\nFound ${rows.length} procedure files in total.`);

To run it, just pass the target folder as an argument: node list-sop.mjs "C:\procedures". Once the list prints, work down from the oldest modified date and ask the floor “are we still using this one?” Just deleting the procedures nobody uses cuts a surprising amount of confusion.

This script only builds a list — it never changes any file’s contents. So it’s safe to run as many times as you like. If touching a terminal is new to you, the basics of Claude Code are covered in the Claude Code getting started guide, so read that first.

Security and confidentiality you can’t skip

Take this lightly and you lose trust itself. The drawings a machine shop handles are the customer’s intellectual property. Part numbers, dimensions, and machining conditions are a serious problem if they reach a competitor.

Three lines to hold:

1. Don’t casually upload images or drawing PDFs with customer names or drawing numbers to a free consumer AI. Some terms of service state “input may be used for training.” For business use, choose a service with a contract or setting that does not use your data for training.
2. Run things so customer-identifying information is stripped from transcripts and procedures. Standardize the small step of swapping part numbers for internal codes and masking drawing numbers.
3. Before handing anything to AI, pause and ask “would I be okay if this got out?” When in doubt, don’t send it. That alone prevents most accidents.

This mindset — what to show AI and what to keep back — also comes up in Claude Code for non-engineers. It’s the first instinct a floor lead should develop. For a deeper grounding in how to give AI clear instructions so it leaks less and asks more, see advanced prompt engineering for Claude Code. For how enterprises think about protecting data in generative AI, public references like the NIST AI Risk Management Framework are a useful starting point when you write internal rules.

A rough ROI estimate

Without numbers, neither the floor nor the boss moves. Here’s a back-of-the-envelope estimate.

Writing one work instruction from scratch — interviewing plus cleaning up the prose — takes half a day (4 hours), in my experience. Replace that with “veteran talks for 10 minutes + AI formatting + 1 hour of human checking” and you’re at roughly 1.5 hours. That’s 2.5 hours saved per procedure.

Item	Before	After
Time per work instruction	~4 hours	~1.5 hours
For 50 procedures a year	200 hours	75 hours
Time freed up	—	~125 hours

125 hours a year is close to three weeks of one person’s labor. And on top of the time savings, you get something bigger: procedures that never got written before now get written. The cost is roughly the transcription and AI usage fees, and you can start from a few dollars a month. Payback should come fast.

FAQ

Q. Can a veteran who’s bad with computers still use this? A. The veteran never has to operate the AI. All they do is talk. The recording and formatting can be handled by the floor lead or an office staffer. Split the roles and you can include people who hate machines.

Q. I’m worried about transcription accuracy. A. You’ll get misfires on jargon and regional dialect. So don’t chase perfection — treat it as “a draft foundation.” Assume a human fixes the misfires, and it’s still dramatically faster than writing from zero. That’s the realistic way to use it.

Q. The handwriting is so messy even AI can’t read it, right? A. Have it honestly report “illegible” when it can’t read something. At that point you just have to ask the author. But the big win is getting to ask while they’re still on the job. After they’ve quit, it’s too late.

Q. Is it worth it for a small shop? A. Small shops benefit more. Fewer people means a bigger blow when one person leaves. The smaller you are, the more urgent it is to break the single-person dependency.

Q. I want to get better at giving instructions. A. Sharper prompts mean less rework. The Claude Code productivity tips post collects concrete ways to build your instructions.

What I found when I actually tried this

At a metal-cutting shop run by someone I know, we tried Use case 1 — “record it and turn it into a work instruction” — together. The target was a fixture setup done by a veteran who was about to hit retirement.

The recording ran 12 minutes. Hand the transcript to Claude Code, and the work-instruction template came out in a few minutes. The draft had [VERIFY] stamped on every RPM and feed number, and we cleared those with the veteran in about an hour. What I’d assumed was a half-day job finished before lunch.

The interesting part: AI flagged a gap — “this step has no explanation for why it’s in this order.” The veteran himself said, “Oh, yeah, I just did that on autopilot.” Setup so obvious it had never been put into words got pulled to the surface by the AI’s question. That was a payoff I hadn’t seen coming.

If you want to run this as a company-wide knowledge-transfer system, lock down the rules and data handling first. If you’d like to work that out in team training or a one-on-one consultation, take a look at the training and consulting page. But the fastest first step is to try just one procedure in your own shop — one where it’s fine if it goes wrong.