How we use AI as a photo lab

Vivid Metal Prints · Lexington NC · Field notes from the floor

We are a metal print lab in North Carolina. Over the last year we have rebuilt a large part of how this shop runs using AI, and we have been asked about it enough times by other labs that it is worth writing down properly.

This is the long version. What we use it for, what we deliberately do not use it for, the exact method, and what it actually replaced. If you run a lab or a print shop, everything here is reproducible for about thirty five dollars a month. None of it required us to write code.

Start with the part everybody actually wants answered.

Where AI touches a file, and where it does not

When a lab says "we use AI," the first thing a photographer hears is you are letting a machine change my picture. That is a fair thing to worry about and it deserves a straight answer rather than a marketing one.

Here is exactly where we stand.

That is the whole boundary and we would rather state it precisely than make a broad promise we would have to walk back. Upscaling is an AI process. We use it, we tell you we use it, and we turned off the part that phones home. Everything past this point is about running a business, not about touching pictures.

The reframe that makes the rest of it work

Almost everyone who bounces off this stuff bounces for the same reason: they are picturing a search engine.

You are not paying twenty dollars for software. You are paying twenty dollars for a person.

And every separate conversation you open is a separate person. Not a session, a person. The one you told about your pricing knows your pricing. The one you told about your CNC files knows your CNC files. They do not compare notes.

That single fact explains nearly every mistake we made early on. Opening a new chat and being annoyed that "it forgot". It did not forget, that is somebody who has never met you. Dumping four unrelated projects into one conversation is one employee juggling four jobs badly. Give each conversation one job and it will be excellent at that job.

The second reframe is about how you talk to it. Google takes keywords. This takes conversation, and the single most useful sentence we have found is this one:

The stack, and what it costs

Two accounts. That is the whole floor.

Thirty five dollars a month and every method in this article works. We eventually added Netlify for hosting what we build, Monday as the hub, and a handful of connectors, but none of that is month one and none of it is required to start.

Context is the whole game

The difference between shops that get something remarkable out of this and shops that get mediocrity is not intelligence and it is not clever prompting. It is how much context they bothered to hand over before asking for anything.

Here is the trap, and we fell in it repeatedly before we figured it out. You have an idea. You open a chat. You type three sentences. It builds the wrong thing. You spend the next two hours correcting it, and the entire time it is carrying your original three thin sentences around as the foundation of everything it does.

The fix is to separate working out what you want from building it, and to do them in two different conversations.

1. Open a chat and describe the idea badly. At length, out of order, contradicting yourself. Then end with an explicit instruction: do not build this, interview me, and give me a prompt I can paste into a fresh chat.
2. It asks you ten to thirty questions. Every one of them will be a question a professional in that field would have asked you.
3. It hands you a complete brief. Dense, organized, containing everything you actually meant.
4. Paste that into a brand new chat. Now the real work starts from a full picture instead of three sentences.

The loop

This is the method. Everything else in this article is a footnote to this section.

Your job: talking  /  Its job: everything else

Notice what is not in there. No typing long messages, no pasting transcripts, no code, no configuration. You talk, and you tell it where to listen.

The exact words we use

Hey, I just left you a recording in Otter about [project]. Go listen to it.

Don't build anything yet. I want you to ask me as many questions as you
need so we can really flesh this out together. Assume I'm a zero out of
ten at coding and explain anything technical like you're guiding a
five-year-old.

I answered all of your questions in the Otter recording I just made.
Go scan it, then build me version 1.0.

If anything I said was contradictory or unclear, flag it instead of
guessing.

Why talking beats typing

This is the part that sounds like a gimmick and is not. It is not about speed. Speed is a side effect.

Watch yourself the next time you type a message. You start a sentence, delete half of it, tighten it, cut the tangent, drop the caveat, and send something clean and short. Every one of those cuts was context. The tangent you deleted was the thing that would have made version one right. You edit your own brief down to a stub and then wonder why the output is generic.

When you talk you do not edit. You go sideways. You circle back. You say "actually, no, more like" and then describe the thing properly on the second attempt. And it understands all of it.

Fifteen minutes of talking produces more usable context than an hour of typing, at a fraction of the effort. Some of our recordings run two hours. Nobody would type that much.

It feels ridiculous for about three sessions. Then it flips, and you catch yourself doing design work in the car on a drive you were going to spend on the radio.

Reviewing out loud is where it earns its keep

Most people talk their way into version one and then start typing again for the revisions. That is backwards. You get version one, it is 70% there, you type "make the header bigger and change the colors," and you get something 72% there. Forty rounds of that is a week spent reaching 85%.

Instead: open the thing, hit record, and narrate your reaction in real time as you scroll. Unfiltered, detailed, ranting where it is warranted.

• Say what is good, too. "Do not touch the pricing block" is as valuable as a complaint, because it stops good work from being rewritten.
• Change your mind mid sentence. "I love the banner. Actually, no, I hate it." That is not a mistake to clean up, it is information about what you actually feel.
• Describe motion and feeling, not settings. "Rotates in as I scroll" beats "add an animation." You are the client here, not the implementer.
• Go overboard, then pull back. Far easier to say "that is too much, dial it back 40%" than to coax life out of something flat.
• Do not fix, complain. Your job on that recording is reaction. Its job is solutions.
• Stutter freely. Say "um" fifty times, start the same sentence three ways, do not re-record. The transcript can be a disaster and the output will still be clean.

All you are doing is pretending you are criticizing a developer's work to his face. And then he goes and fixes it.

Runway and handoffs

Every conversation has a finite amount of memory, and nobody tells you when you are running out. As you approach the limit it does not crash, it starts compressing. It takes the rich version of everything you discussed and quietly turns it into a bulleted list to save room. So it does not fail. It gets dumber. It loses details from early on, repeats decisions you already made, contradicts something it said an hour ago. By the time you notice, you have usually wasted a session.

So you ask.

How's our context looking? How much runway do we have left in this chat?

If the answer is "we're getting loaded," stop. Do not start the next feature. Ask for a handoff document, read it, copy it, open a brand new chat, and paste it as the very first message. Nothing before it, nothing after it.

That is how a serious project runs across twenty five separate conversations chained into one continuous piece of work. Every link in that chain is a handoff.

We're getting close on runway. Give me a complete handoff document I can
paste as the first message of a fresh chat. Include:

  1. What we're building and who it's for, in plain language.
  2. Every decision we've made and WHY we made it.
  3. Things we tried that didn't work, so the next chat doesn't retry them.
  4. Current state: what files exist, what's done, what's half-done.
  5. Exactly what the next step is.
  6. My constraints and preferences, and my skill level.
  7. Anything I've said that you think is important that I might not
     realize is important.

Write it for someone who has never seen this project.
Also tell me anything you think I've been getting wrong or avoiding.

Two rules that took us a while to learn. Hand off early, not at the buzzer, because a handoff written by a chat that has already degraded is itself degraded. And read it before you paste it, because that is your one clean look at what it thinks the project is. If it has drifted you will see it there, and it is ten times cheaper to correct in a handoff than three hours into the next chat.

Then keep every handoff. That is your project's save file. If a chat goes sideways you restart from the last good one and lose an hour instead of a month.

Pointing it at real files without wrecking anything

This is the part that makes people nervous, correctly. Claude has an Add folder button that gives it access to an entire folder on your machine. It can read everything in there, reorganize it, and write new files into it. That is enormously useful and it is exactly the capability you want guardrails on.

Keeping the cost down

On a twenty dollar plan your monthly allowance is the real constraint, not your ideas. Three habits roughly double how far a month goes.

Model and effort

There are two dials. The model list is ordered sensibly, most capable at the top. Effort is how long it thinks before it answers. Workhorse model at high effort is where you live. Low and medium effort are for execution, formatting, cleanup, and small edits you have already specified, and they are very cheap and perfectly capable.

The senior and junior split

This is the single biggest cost saver we use. Real work has two very different kinds of task in it. Deciding what to do is hard and benefits enormously from a strong model. Doing it is mostly typing, and a cheaper configuration handles it fine. Paying premium rates for the typing is where most people's allowance disappears.

So we run two chats in the same project folder. The senior is on the frontier model at high effort: it understands the project, makes the plan, reviews what came back, decides what is next. The junior is on the workhorse model at medium effort: it takes a detailed plan and executes it exactly. It does not need to be brilliant, it has been handed brilliance in the form of a plan. You are the courier between them.

You're the senior developer on this project. Don't write the
implementation yourself, I'm handing it to a junior. Give me an
implementation plan detailed enough that someone with no context could
execute it correctly. Flag anything you'd expect them to get wrong.

This is not only about money. A dedicated reviewer that did not write the thing catches problems a single chat will not, because a chat reviewing its own work is naturally generous with itself. Better output and a cheaper month at the same time.

Do not pay premium rates for pictures

Any time a project needs a visual asset, we do not let the paid account generate it. We ask it to write the prompt, run that on a free tool, and bring the result back. Free ChatGPT and free Gemini live permanently in browser tabs here for images, throwaway questions, second opinions, and explaining error messages. Every one of those you do not spend on the paid account is another real build you get that month.

What this actually replaced

None of the following required us to write code. All of it required knowing what we wanted and being willing to describe it out loud, repeatedly, until it was right.

The CNC program

We cut metal in house. Running the machine and creating the cut file are two completely different skills. We had operators, we did not have a designer. So for years we paid an outside contractor somewhere in the range of a thousand to fifteen hundred dollars a month, plus a hundred and fifty to two hundred per drawing, and it routinely took a week or two to get a file back. Clients waited. Our warehouse guy could not so much as duplicate a shape without going back through him.

Our machine is old, so it runs on G-code, which is raw text. We could not even see our own shapes. They were number files we loaded and hoped ran correctly.

We had about 130 of those G-codes sitting in a folder. We handed over the folder, talked through the problem for fifteen or twenty minutes, and about two hours later we had a tool: every shape we have ever cut, rendered visually and clickable. Click one and it drops onto a board and tells you how big a sheet you need. Right click to duplicate. Drop in a PNG outline and it traces it into a cut program.

The whole thing is a 250 KB file. We can email it to anyone. It has never been on the internet.

The lead list

One of the biggest furniture markets on the East Coast lists its exhibitors ten per page across roughly 175 pages. About 1,700 businesses. Our front office is two people. We were never going to call all of them.

So we asked for a spreadsheet: business name, a one sentence description of what they do, contact details where findable, and a low, medium or high rating on whether there was enough product adjacency to bother. If they sell sinks, we are not selling them metal prints.

Partway through, the site started blocking it. We watched it reason through that in real time and write a script to keep going. Fourteen minutes. A color coded spreadsheet, all 1,700 companies evaluated, narrowed to about 62 genuine matches. We mailed them flyers, met several at the show, and did business with roughly 15% of them.

The pricing calculator

Live on our site now. Pick a product, add items, unlock reseller pricing, generate a white labeled PDF with your own markup on it, with live UPS rates pulled straight into the tool. It turned a twenty minute quoting process into seconds, for our team and for our customers.

Fourteen years of files

Two folders, work and personal, holding well over half a million files. Thousands of images named Screenshot 2019-04-11 at 3.42.11 PM. Hundreds of documents called Untitled.

We handed over the work folder and asked it to organize everything in a way that would make sense to us: open every screenshot, look at it, rename it based on what it is a screenshot of, file it by date. Any document called untitled, open it and name it from what is inside. Then produce a spreadsheet logging every change and the methodology, and put a "how to read this folder" document at the top level.

Somewhere between forty five minutes and an hour. Fourteen years, reorganized, with a manual.

The sales automation

Our salesperson runs this every day. Start a recording, call the customer on speaker, have the conversation, hang up, then debrief out loud for thirty seconds. At the end of the day an automation reads all of her recordings, works out which were sales calls from the debrief, pulls the details into the CRM, and sets the follow up reminders. If she said the words "I want to send them a sample kit," it writes the order up automatically. It lands in our store as a zero dollar sample order and production prints and ships it.

Three people used to touch that. Now nobody does.

If you run a lab and want to start

You are going to want to do all of it at once. Do not.

The best first project is not the most exciting one. It annoys you, it is small, and nobody is counting on it. The real test: what did you do this week, more than once, that made you sigh? A quote you assemble by hand. A report you rebuild every Monday. A folder you keep meaning to organize. That is the project.

Why we wrote this down

We taught a version of this at IPIC and ran out of clock both times, which is how this article exists. The labs in that room were dealing with the same things we were: not enough hands, quotes taking too long, a contractor holding up a job, a folder nobody has opened in six years.

None of what we built is proprietary to metal. A lab doing canvas, framing, albums or signage could rebuild every one of these in a month for thirty five dollars, and we would rather this industry get good at it than not.

And the boundary at the top of this page holds regardless. We automate our business. We do not automate your photograph.