It's 6:15 AM on second-shift training day. Your new operator is at Station 3 with three sources of "truth": a print marked up in red pen from last month's engineering change, a Slack thread where the process engineer typed "same as PN-4412 but skip the second deburr," and a notebook page someone photocopied in 2023. The lead walks over, shrugs, and says "just watch me once." Twenty minutes later the lead is pulled to a quality hold. The new hire guesses.

This is how work instructions get made in a lot of 30–80 person shops: notebooks, Slack, and tribal walkthroughs. Auditor SOPs are a different stack. Floor operators don't open those binders when the traveler hits their station — they open whatever the last engineer left behind.

That gap costs scrap, rework, and slow onboarding. One wrong setup can burn a shift of machine time. Temps take days longer when instructions live in someone's head. When the process engineer is out, Station 3 becomes a fire drill.

This post is about fixing that with Claude — not quality-system SOPs, but turning messy engineer notes into stepwise, versioned work instructions operators can follow without a live coach.

Work Instructions Are Not SOPs

If you've already seen our post on writing SOPs and quality docs with Claude, this is a different lane.

	SOPs / quality docs	Work instructions
Audience	Quality, auditors, customers	Floor operators, leads, temps
Tone	Formal, compliance-ready	Plain language, station-level
Trigger	Audit, shipment, NCR	Job traveler hits the station
Failure mode	Missing cert / incomplete form	Wrong setup, skipped gate, scrap

Work instructions answer: What do I do at this station, in what order, with what checks, before I hit start? One or two pages. Tools, fixtures, safety and quality gates, and what "done" looks like. Engineer notes are the raw material. Claude structures them — it should not invent process you never wrote down.

What Engineer Notes Usually Look Like

Before you paste anything into Claude, gather what you already have. In shops we work with, the "source pack" for one station often includes:

• Handwritten notebook pages or photos of whiteboard markups
• Slack or Teams messages after an engineering change ("Rev C — change torque to 18 ft-lb, remove step 4")
• A marked-up drawing with arrows and "SEE NOTE 3"
• An old Word doc last updated two revs ago
• A verbal walkthrough the lead gave once and never wrote down

None of that is useless — it's unstructured. Dumping a notebook photo into Claude and publishing the first draft produces confident-sounding wrong steps. The workflow below forces structure, flags ambiguity, and keeps a human in the publish path.

The Claude Workflow: Notes → Floor Instructions

Use Claude Pro ($20/month) and a Project dedicated to work instructions — separate from your quality-docs Project if you have one. Project memory is what makes this stick: Claude keeps your station naming, part-number format, and instruction template across conversations.

1. Set Project instructions once

In Project settings, write 4–6 sentences about how your floor runs. Example:

We're a 45-person CNC / assembly shop. Part numbers look like XXXXX-Rev-XX. Stations are named by cell (Cell A / Lathe 2 / Weld Booth 1). Work instructions are operator-facing: numbered steps, tools/fixtures listed up front, safety and quality gates called out in bold, no auditor jargon. Never invent torque values, speeds, or hold points — flag unknowns as [NEEDS ENGINEER]. English at 8th-grade reading level.

Upload one good example of a finished work instruction (even if you only have one) and your blank template if you have one.

2. Structure the raw notes first

Paste or upload the engineer notes. Ask Claude to reorganize only — no polished prose yet:

Reorganize these engineer notes into:
1. Purpose / when this instruction applies (part + station + rev)
2. Tools, fixtures, PPE
3. Numbered process steps (one action per step)
4. In-process checks / quality gates
5. Safety callouts
6. Ambiguities or missing values as [NEEDS ENGINEER]

Do not invent numbers. Do not polish language yet.

You want a skeleton with holes labeled. If Claude invents a torque or a time, send it back: Remove invented values. Mark unknowns.

3. Clarify ambiguity with the engineer — not the model

Take the [NEEDS ENGINEER] list to the process engineer (or the lead who owns the station). Ten minutes of answers beats an hour of guessing. Common holes:

• Which fixture for Rev C vs Rev B?
• Is the second deburr still required after ECO-1847?
• What is the go/no-go on the critical dim?
• Where does scrap go vs rework hold?

Paste the answers back into the same Claude chat. Ask it to fill only the resolved holes and leave anything still open flagged.

4. Write operator-facing steps

Now ask for the floor version:

Rewrite as a work instruction for a new operator.
- Title: part number, station, rev, effective date
- Tools/fixtures/PPE at top
- Numbered steps, one action each
- Bold safety and quality gates
- End with "Done when…" acceptance checks
- Keep every [NEEDS ENGINEER] that is still open

Push Claude toward short sentences and shop language: "Load blank in Fixture F-12," not "Utilize the designated holding apparatus."

5. Add safety and quality gates explicitly

Ask for a pass that inserts gates where a wrong move scrapes a part or hurts someone:

• PPE before power-on
• Lockout / air dump if required
• First-piece check or critical dim before running the lot
• Stop conditions (noise, chatter, missing hardware)

These should interrupt the step list, not hide in a footer nobody reads.

6. Human review before anything hits the floor

The process engineer or cell lead owns publish. Review checklist:

• Every number matches the current ECO / print rev
• No leftover [NEEDS ENGINEER] on a live instruction
• Station name and tooling match what's on the floor today
• Version + effective date on the header
• Old photocopies / Slack screenshots pulled from the station binder

Claude drafts. Your people approve. That line is non-negotiable on the floor.

What a Good Output Looks Like

A solid work instruction for one station usually includes:

1. Header — part, rev, station, author, effective date, supersedes
2. Setup — tools, fixtures, PPE, material condition at arrival
3. Steps — 8–20 numbered actions max for one station
4. Gates — bold stop-and-check moments
5. Done when — measurable acceptance (dims, torque marks, visual criteria)
6. If something's wrong — who to call, hold tag location

That's enough for a temp to run without inventing process. It's not enough to replace training — but it stops "watch me once" from being the only system.

The Math on One Station

A process engineer often spends 3–4 hours turning messy notes into a clean one-page instruction the old way. With a structured Claude pass, that often drops to 45–90 minutes of review and gap-filling once the Project and template exist.

Across 12 high-churn stations, cutting ~2 hours per refresh saves roughly 24 hours per cycle — about $1,500–$2,000 at a $65–$85/hour burdened engineer rate — before counting scrap avoided when operators stop guessing. Onboarding improves too: "shadow a lead for a week" becomes "read the instruction, run under supervision for a day" on standard jobs.

What Still Needs a Human

Claude does not:

• Own process ownership. If the notes are wrong, the instruction will be wrong faster.
• Replace ECO control. Rev changes still need a human to update the header and pull the old sheet.
• Substitute for training on hazardous ops. Gates help; they don't certify an operator.
• Integrate your ERP traveler automatically unless you build that next (n8n + your document store is a later Quick Win).

Start with one cell and one engineer who already writes notes. Prove the template. Then expand.

Soft Ladder If You Want Help Standing This Up

You can run the workflow above yourself this week. If you'd rather not invent the Project setup alone:

• Free assessment — map which stations still run on notebook / Slack truth
• Tier 1 AI Audit ($999) — score the gaps and pick the first 3–5 stations
• Quick Win — Claude Project + template + one station published end-to-end
• Sprint — roll the pattern across a cell with versioning and a named owner

One station that stops guessing is enough proof. If Station 3 still runs on a Slack screenshot and a photocopied notebook page, book a discovery call at cloudbeast.io/schedule. We'll walk your notes pack and sketch the first work-instruction Project before you spend a dollar on a build.