Print Shop Workflow: Speed Up Preflight Checks and Quotes with Claude Code

It’s Friday afternoon, and a customer just emailed a PDF with the note “this one’s final, go ahead and print.”

In my experience (and across the print shops I trade notes with), Friday rush jobs are exactly when accidents happen. The flyer looked clean at a glance. The bleed seemed to be there. We ran it Monday, and a thin white line showed up along the trimmed edge. On closer inspection the bleed was only 2mm, and it landed right on the border of the most prominent photo. The customer insists they sent it “exactly as designed,” and we already printed it. Before anyone argues about whose fault it is, the paper, the ink, and the time are already gone.

Preflight checks and quote replies. These two tasks quietly eat into a print shop’s margin. Today I want to walk through how we rebuilt them so a generative AI (Claude Code) does the grunt work, and the human only makes the judgment calls.

Key takeaways

• Preflight and quote replies always collapse on a Friday afternoon when you rely on human focus alone. Put a machine gatekeeper in front of them.
• Hand the AI the “read it, pull it out, line it up, draft it” work. Keep the final sign-off, color proofing, and final pricing in human hands.
• This article includes a ready-to-use prompt template, a preflight checklist, a quote-reply template, and a verification script that mechanically checks a PDF’s layout.
• For a shop doing 20 quotes a day, drafting the replies alone frees up roughly 20 hours a month. Prevent one reprint and you save the paper cost alone, which can run into hundreds of dollars.
• Submission data contains personal information and unreleased product details. Decide how to redact it and what your rules are for sending data outside the building before you start.

First, who is this for?

I’m writing for the small-to-midsize print shop, where the same person runs both DTP and sales. You touch Illustrator, InDesign, and PDFs every day, and you write your own quotes. You handle everything from intake to arranging the color proof, solo.

In a shop like that, there’s no room to add more tools. No budget approval cycle for a new system, no time to learn one. So I think the honest goal is “keep using the PDFs and spreadsheets I already have, just stop missing things.” This article stays in that lane. No talk of giant ERP systems.

Lay out the print shop flow, from intake to delivery

Before you decide what to delegate, you have to break your own job into steps. It probably looks something like this.

Step	What you do	Where accidents happen
1. Intake / discovery	Ask size, quantity, paper, finishing, deadline	Missed details. The late “oh, it’s double-sided”
2. Quote	Calculate per-quantity unit price, plate cost, finishing cost, and reply	Math errors, lost jobs from slow replies
3. Submission intake	Receive the data from the customer	Old version submitted, fonts not embedded
4. Data check	Bleed, resolution, color, crop marks, garbled text	A reprint caused by an oversight
5. Proof / color proof	Get the customer’s confirmation	Vague verbal “OK”
6. Platemaking / printing / finishing	Actually print it	Upstream mistakes surface here
7. Shipping / billing	Deliver and invoice	Wrong quantity

Generative AI is useful for step 2 (drafting the quote) and the first pass of step 4 (data check). The color proof in step 5 and everything after step 6 stay off the machine. I’ll explain why below.

The reprints you keep seeing, and what causes them

The patterns that force a reprint are almost always the same cast of characters.

• The bleed (the margin past the trim) is too small, so white shows after trimming.
• Low image resolution. A 72dpi photo grabbed off the web, submitted as-is.
• Solid black set as rich black, so misregistration leaves type fringed green or red.
• Fonts not embedded, so the type swaps to a different face when opened on another machine.
• Submitted in RGB, so the color goes dull in print.
• No crop marks, or the finished size doesn’t match the spec.

The tricky part is that none of these are obvious at a glance. A human can pick them out one by one, but when ten jobs land on a Friday afternoon, focus doesn’t hold. This is the textbook case of a boring check that a machine is good at.

Use case 1: let the machine do the first preflight pass

Turn the items a human tends to miss into a checklist and run the AI over it first. The aim is to change the human’s final review from “look at every item from scratch” to “only verify the spots the AI flagged red.”

Reuse the checklist below on every submission.

• Does the finished size match the spec?
• Is the bleed at least 3mm on all four sides?
• Is each placed image’s effective resolution 300dpi or higher at actual size?
• Is the color mode CMYK (no stray RGB)?
• Is small black type (especially) single-channel K100, not rich black?
• Are all fonts embedded or outlined?
• Are crop marks and center marks present?
• Do the page count and imposition orientation match the spec?
• Are there any unintended transparency or overprint settings?

Hand the AI the PDF properties or the text of a preflight report, have it cross-check against this list, and it will point out the missing rows directly.

Use case 2: draft a quote reply in 30 seconds

For quotes, the lost-job cause isn’t usually the math. It’s that writing the reply is a chore, so it gets put off. Hand over the discovery notes as bullet points and have the AI produce a draft of a polite reply.

Teach the AI your own price table (your unit-price logic) up front. A human does the final check on the numbers, but you can leave the wording and the order of the line items to the AI.

Use case 3: surface missing details from the inquiry email

Customer inquiries are almost always short on information. “I need 100 business cards” is all that comes in. Have the AI build the list of follow-up questions first: single- or double-sided? what paper? what deadline?

Fewer gaps in what you ask means fewer redone quotes. It’s unglamorous, but cutting one round-trip lands the job half a day sooner.

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

This is the most important part. Get the line wrong and you get the accident where the AI is “confidently wrong.”

Task	Delegate to AI	Human decides
Cross-check against the checklist	Yes, catch the gaps	The final OK
Reading off resolution / bleed numbers	Yes, line up the numbers	Whether a photo’s edge is visually conspicuous
Drafting quote wording	Yes, write the prose	Final pricing, discount calls
Building the follow-up questions	Yes, list them out	Phrasing that fits the customer relationship
Color proofing / final color call	No	Yes, always with human eyes
Sending to the press	No	Yes, a human pushes the button

Color is the one thing you never hand over. Monitor color and ink color are different beasts, and they shift with the paper and even the weather. That’s a job for a veteran’s eye. Draw a firm line: the AI goes as far as “are the settings in the data wrong,” and it never judges whether the printed result is good or bad.

This way of drawing the line is continuous with using Claude Code as a non-engineer. The trick is to start with a narrow scope of delegation.

A copy-paste prompt template

This is the prompt I use for the first pass of a preflight check. Replace what’s inside the brackets with your own job.

You are a veteran DTP operator at a print shop.
Cross-check the submission data I'm about to give you against the checklist below.

[Job info]
- Finished size: A4 (210x297mm)
- Paper / quantity: 90lb coated / 1000 copies
- Sides: double-sided, color

[Checklist]
1. Does the finished size match the spec?
2. Is the bleed at least 3mm on all four sides?
3. Is each placed image's effective resolution 300dpi or higher at actual size?
4. Is the color mode CMYK (no stray RGB)?
5. Is small black type single-channel K100 (not rich black)?
6. Are fonts embedded or outlined?
7. Are crop marks present?

[Output rules]
- Mark each item OK / Needs review / NG and add a one-line reason.
- For any item where the number can't be read, write "Insufficient data" and do not fill it in with a guess.
- At the end, list the questions to confirm with the customer as bullet points.

Always include “do not fill in with a guess” and “write insufficient data as insufficient data.” Without that, the AI plausibly fills in what it doesn’t know and hands you a false sense of safety. For the finer points of tuning a prompt, see advanced Claude Code prompt design.

Quote-reply template

Pour the reply draft into this mold.

Dear [Customer name],

Thank you for your inquiry. Here is our proposal.

[Specs]
- Item: [business cards / flyer / booklet, etc.]
- Size: [size]
- Paper: [paper type / weight]
- Quantity: [quantity]
- Finishing: [PP lamination / foil / folding, etc.]

[Price (excl. tax)]
- Printing: [amount]
- Plate / data adjustment: [amount]
- Total: [amount]

[Lead time]
[number] business days after data is finalized

Based on the data check, [note any items to confirm].
Please don't hesitate to reach out with any questions.

A verification script you can actually run

Opening every PDF to measure “is the bleed there” is a chore. Here’s a Node.js script that reads the PDF page size (MediaBox / TrimBox) and roughly judges whether the margin from the trim is at least 3mm. It uses pdf-lib.

npm init -y
npm install pdf-lib
node check-bleed.mjs sample.pdf

import { PDFDocument } from "pdf-lib";
import { readFile } from "node:fs/promises";

const PT_PER_MM = 72 / 25.4; // 1mm = about 2.835pt
const MIN_BLEED_MM = 3;

const file = process.argv[2];
if (!file) {
  console.error("Usage: node check-bleed.mjs <PDF file>");
  process.exit(1);
}

const bytes = await readFile(file);
const pdf = await PDFDocument.load(bytes);

pdf.getPages().forEach((page, i) => {
  const media = page.getMediaBox();
  // If there's no TrimBox, treat MediaBox as the finished size (likely no bleed).
  const trim = page.getTrimBox?.() ?? media;
  const bleedPt = Math.min(
    trim.x - media.x,
    trim.y - media.y,
    media.x + media.width - (trim.x + trim.width),
    media.y + media.height - (trim.y + trim.height)
  );
  const bleedMm = bleedPt / PT_PER_MM;
  const ok = bleedMm >= MIN_BLEED_MM - 0.05;
  console.log(
    `P${i + 1}: bleed about ${bleedMm.toFixed(1)}mm -> ${ok ? "OK" : "Needs review"}`
  );
});

This prints the bleed for every page in one list. A human only needs to open the pages flagged “Needs review” in Illustrator. That’s far faster than measuring every page from scratch. The script is just a first filter, and the final call is still a human’s. If you’re not comfortable on the command line, start with the Claude Code getting-started guide so you don’t get stuck.

What changed, before and after

Rough numbers. These are felt estimates from the small print shops around me, nothing more.

Item	Before	After
Check time per submission	about 15 min	about 6 min (verify only the red spots)
Writing a quote reply	about 12 min	about 3 min (just fix the draft)
Round-trips from missing details	30% of jobs	under 10%
Reprints	1-2 a month	aiming for near zero

Turn the saved time into ROI: at 20 quotes a day, shaving 9 minutes off each reply is 3 hours a day, on the order of 60 hours a month. Even conservatively, 20 hours a month is solid. Prevent one reprint on top of that and you keep hundreds of dollars from vanishing into paper, ink, and press setup. You didn’t buy new software, so the time you freed up lands straight on the profit side.

Security and personal-information cautions

Skip this part and you don’t get efficiency, you lose trust. Almost all the data a print shop handles is someone else’s confidential material.

• Business cards and direct mail carry names, addresses, and phone numbers. A membership-roster print job is even heavier.
• A new-product flyer is unreleased information before launch. A leak can end the business relationship.
• So what you hand the AI is limited to layout data like size, resolution, color settings, and bleed. Redact the personal info and product names in the body before you send it.
• Processing that stays local, like the verification script, never sends data outside, so the safety level is high. Handle the sensitive personal data on your own machine.
• If you do send to an outside AI service, always confirm whether the input is used for training and how the contract treats it. Write a one-page internal rule.

The line is simple. Before you send anything, ask yourself: “is this data I could show the customer without getting yelled at?” When in doubt, don’t send it. For building the in-house rules on what’s OK to show the AI, you can apply the thinking in CLAUDE.md best practices — the idea of writing each project’s agreements into a file. The same discipline of small, repeatable habits shows up in these Claude Code productivity tips.

If you want to nail down the PDF technical spec itself, go to primary sources like Adobe’s official information (Adobe’s official site). Recommended bleed and crop-mark values differ by print shop, so your own submission guidelines take priority.

FAQ

Q. Can I print data the AI said was “OK” as-is? No. The AI’s verdict is a first filter. Color in particular, and how a photo’s edge reads, must be checked with human eyes. The AI only catches “gaps in the data settings”; it can’t take responsibility for the printed result.

Q. My operators aren’t comfortable with the command line. Is the verification script mandatory? No. The checklist and prompt alone are effective. The script is extra gear for people who find “measuring bleed every time” a chore. Start with the prompt first.

Q. Is it safe to let the AI read a customer’s personal information? As a rule, no. Hand over only the layout data (size, resolution, color) and redact names and addresses. If you want it to read the body, use a local setup that doesn’t send data outside, or swap in dummy text first.

Q. Can the AI decide the quote amount too? A human makes the final call on the amount. The AI only drafts in line with your price table. Discounts and rush fees can only be judged by someone who knows the relationship.

Q. How many days does it take to set up? If it’s just preparing the checklist and reply template, half a day. Tune the prompt to your own submission guidelines over a week or two of real use, and the accuracy of the verdicts settles into your shop.

Wrap-up: keep the judgment, let go of the boredom

For both preflight and quoting, the chore isn’t the “judgment,” it’s the “pulling out and writing up.” Let the AI shoulder that, and the human only looks at the red spots, the color, and the price. Re-divide the roles this way and accidents drop even on a Friday afternoon.

If you want to rethink how submissions get stopped and who owns which role across the whole company, we can work out a setup tailored to your shop’s flow together through training and consulting. If you’d rather try it on your own first, grab the checklist and prompts from the free PDF and learning materials.

What I actually found when I tried it

I prepared a few sample PDFs and ran the verification script above. A page built with 0mm bleed was correctly caught as “P1: bleed about 0.0mm -> Needs review,” and a version built with 3mm came back “OK.” A PDF that only has a MediaBox gets judged as having no bleed, so I saw a limit right there: unless it’s an export with a TrimBox (a PDF/X with crop marks), the script can’t measure accurately.

For the preflight prompt, I deliberately mixed in “information where the resolution can’t be read.” The payoff was that it followed instructions and returned “Insufficient data” instead of guessing. On the flip side, whether something was rich black couldn’t be fully determined from the text I handed over, so in the end a human had to open Illustrator and look. The boundary between what you can hand the AI and what you can only judge by eye only came into sharp focus once I ran it. Don’t try to delegate everything from the start; hand over only the boring pulling-out. That distance feels just right for a print shop.