AI Automation for Graphic Design: Resizing, Handoffs and Proofing

AI and plain scripts can reliably handle the mechanical side of studio production: exporting size variants from an approved master, naming files, building handoff folders, converting colour profiles and turning client comments into a numbered change list. A designer still checks every crop, brand colour, font, small-size legibility, print specification and whether each comment was really applied.

Order matters more than the choice of tool. Automate only after a designer has approved the master artwork, because every variant inherits its mistakes: a wrong logo version or an off-brand blue is copied into fourteen files in a minute. Resizing within one aspect ratio, naming and packaging are safe to hand over. Judging a crop, a colour or a client's meaning is not.

Follow me on Instagram@sagnikteaches

Three tiers: what runs alone, what gets a glance, what stays manual

Sort every production job into a tier before choosing a tool. The tier decides who looks at the output, and that matters more than which software produced it. A useful test is the cost of a miss multiplied by how hard it is to spot. A wrongly named file is obvious to a script. A slightly wrong crop is invisible at thumbnail size and embarrassing at full size.

Connect on LinkedInSagnik Bhattacharya

Two cautions about the tiers. A job moves up (towards more human checking) after any miss the client sees. It moves down only after a run of clean jobs, and I'd want about ten before trusting it. That figure is my rule of thumb, not a measured one.

Subscribe on YouTube@sagnikteaches

One campaign, fourteen variants, and the method beside each

Take a campaign for a specialist food importer's new tasting box. The studio has an approved portrait master for social posts (1080 × 1350 pixels), and the client wants fourteen further outputs. The studio sorts them by method before opening any software. The sizes here are an illustration; confirm each channel's current specification before the matrix is locked.

Straight exports are the only group where a preset can run unattended. The layout does not change, so the failure modes are limited to resolution and file type. Reflows move text and images inside a template, so a person looks at each one. The three recomposed sizes are fresh design decisions: 1080 × 1920 is about 42% taller than the master at the same width, 728 × 90 has room for little more than a logo and four words, and the flyer carries bleed and a print profile.

One sum catches most straight-export failures. At 300 pixels per inch, a photo spanning the full width of an A5 flyer needs about 1,748 pixels across, while the A3 version of the same photo needs about 3,508. A supplier image 2,000 pixels wide is fine on the A5 and drops to roughly 171 pixels per inch on the A3. Many printers ask for 300 at final size (use your printer's own figure), so the A3 file fails a test the A5 passed, and no export preset will warn you. Put the check in your sheet as a formula, effective ppi = pixel width ÷ placed width in inches, because arithmetic belongs in a formula, not in a chat window.

Six resizing rules, and an assistant that plans the reflow

These rules live in the template, the preset or the checklist, never in anyone's memory.

1. Change the aspect ratio only by hand. Photoshop's Image Processor has a Resize to Fit option that keeps the original proportions inside the width and height you enter, so it scales but never recomposes. If the ratio changes, the job is a reflow or a recompose, not a resize.
2. Never upscale a raster file past its native pixels. Use the effective-resolution sum above, and ask the client for a larger original instead of enlarging a small one.
3. Lock the zones that must not move. Logo clear space, price, legal line and any face are locked in the template. Figma's Buzz page describes brand templates where you set which parts are editable and which are not, which is this rule in tool form.
4. Set a minimum text size for each output. As example thresholds, 28 pixels on a 1080-wide social image and 14 on a leaderboard ad. Check them on a real phone screen and adjust from your own complaint history.
5. Screen files in sRGB, print files in the printer's profile. Image Processor offers Convert Profile To sRGB and an Include ICC Profile tick box, so screen conversions can run unattended. The print profile comes from the printer's spec sheet and stays a manual step.
6. Originals stay untouched. Exports go to a new folder. Image Processor saves each processed file under a unique name instead of overwriting, and puts results in JPEG, PSD or TIFF subfolders of the destination you choose.

A realistic miss shows why rule one exists. A square variant came out of a preset that centre-cropped the portrait master. Going from 1350 to 1080 pixels of height removes 270 pixels, 135 from the top and 135 from the bottom. The headline started 90 pixels from the top and the strapline ended 100 from the bottom, so both were clipped. The preset reported success because it did what it was told. The fix is a rule, not a better preset: any ratio change is a reflow, and reflows get a glance.

An assistant is useful before the work starts, when you ask it to find the risks rather than make the layout. Attach the approved master and use a prompt like this one.

I'm a studio designer. The attached image is an approved master: portrait, 1080 x 1350 px, headline top left, product photo in the centre, strapline and logo along the bottom.
Plan reflows to 1080 x 1080, 1080 x 1920 and 728 x 90.
For each size give a table row with: elements at risk of cropping or crowding; the smallest text size in px that stays legible; whether a straight scale is safe or the layout needs recomposing.
Do not redesign anything. Do not invent brand colours or fonts.

Here is an illustrative answer, the kind you will actually get back.

The first row is wrong. A square is 270 pixels shorter than the master (1350 minus 1080), so the bottom band is cut unless the layout moves. The verdict should read reflow, with a glance. The smallest-text figures are the assistant's guesses, so treat them as a starting point and test on a phone. The other two rows earn their place because they name what to look at, and that is all this prompt is for.

The same step covers copy trims. Asked to shorten "Eight small-batch cheeses in one tasting box" for the leaderboard, an assistant may return "Our best cheeses", which loses the number and adds a superlative nobody approved. Trimmed copy goes through the same glance, with the approved line beside it.

Naming and folders: a convention a script can police

When nobody is around to explain a handoff, the file names carry the explanation. The convention below has six parts, all lower case and hyphen-separated with no spaces, because spaces and mixed case cause the most trouble when files move between systems.

Put together, one file reads fab-tasting-box-social-1080x1350-v03-review.jpg. The script may write the word review. Only the designer, holding a recorded client approval, renames a file to approved. The word final never appears, because final is how a folder ends up with final-v2-really-final.

Tidying legacy folders is the first job people hand to an assistant, and it needs a dry run. Below is an illustrative rename plan an assistant returned from a pasted list of old file names.

The rule for the dry run is simple. The plan is printed and never applied until a person has read it, and the script refuses to run if two rows share a new name or any row says approved without an entry in the approval log.

The checker below is a short Python script of the kind an assistant can write from the table above. It lists names that break the convention, more than one version of the same asset and size in a pack, and sizes missing from the campaign matrix.

import re
import sys
from pathlib import Path

# client-job-asset-WxH-vNN-status.ext, all lower case
NAME = re.compile(
    r"^([a-z]{3,5})-([a-z0-9]+(?:-[a-z0-9]+)*)-([a-z0-9]+)"
    r"-(\d{2,5})x(\d{2,5})-v(\d{2})-(review|approved)\.(jpg|png|pdf|svg)$"
)
EXPECTED = {"1080x1080", "1080x1920", "1920x1080", "1200x628", "600x200",
            "300x250", "300x600", "728x90"}

def check(folder):
    problems, sizes, versions = [], set(), {}
    for f in sorted(Path(folder).iterdir()):
        if not f.is_file():
            continue
        m = NAME.match(f.name)
        if not m:
            problems.append("bad name: " + f.name)
            continue
        size = m.group(4) + "x" + m.group(5)
        sizes.add(size)
        versions.setdefault((m.group(2), m.group(3), size), []).append(m.group(6))
    for (job, asset, size), v in versions.items():
        if len(v) > 1:
            problems.append(f"{asset} {size}: several versions in the pack ({', '.join(v)})")
    for s in sorted(EXPECTED - sizes):
        problems.append("missing size: " + s)
    return problems

if __name__ == "__main__":
    out = check(sys.argv[1])
    print("\n".join(out) if out else "names valid, one version per size, all sizes present")
    sys.exit(1 if out else 0)

You do not need to write Python to use this. Ask ChatGPT or Claude to adapt it to your convention, ask it to explain each line in plain English, and run it on a copy of one finished job. If you can read the report and confirm it against the folder, you can maintain it. The same convention needs a folder tree that a script can create for every new job.

2026-11_fab_tasting-box/
  00_brief/
  01_source/          native files, font list, linked assets
  02_working/
  03_review/R1/       proofs for round 1, then R2, R3
  04_approved/        only files with a recorded approval
  05_handoff/         print/  web/  social/
  99_archive/

What goes into the handoff pack for print, web and social

A handoff pack differs by destination, and the receiver's specification wins over the studio's habits.

A one-page note travels with every pack. The assistant drafts it from the job record, not from the files, and the designer verifies each line against the source. Here is an illustrative draft for the tasting-box job.

HANDOFF NOTE (draft from job record, not yet verified)
Job: fab-tasting-box    Version: v03
Approved by: [client approver] on [date], via [proof tool or email thread]
Files: 14 variants in 05_handoff (6 straight, 5 reflow, 3 recomposed)
Colour: screen files sRGB; flyer in the printer's profile
Fonts: [font family] outlined in print PDFs; live in native files
Images: 3 supplier photos (licence on file); no AI-generated content
Known compromises: 728 x 90 uses the short lockup, not the full logo

Two lines in that draft need a human. "No AI-generated content" appeared because the job record had no field for it, yet the vertical story used a generative expand to extend the background. The fix is structural: add an "AI-assisted edits: where, which tool, which model" field to the job record so the note cannot leave it out. The font line is a claim to test, not accept. Open the exported PDF's font list in your PDF reader's document properties and confirm that the fonts you said were outlined are not listed.

Proof rounds: rules a tool enforces and a designer owns

Most proofing trouble is a rules problem rather than a software problem. These six rules make the process mechanical enough for a tool to hold up its end.

1. The number of rounds is written into the proposal, together with the price of an extra one.
2. Each round starts from one consolidated comment list from one named approver, pinned to a numbered version.
3. Every comment gets an ID and one of four outcomes: done, done differently (with the reason), query, or out of scope.
4. Approval is recorded with who, when, which version and which channel.
5. After approval the files lock, and later changes become a new quote.
6. When comments contradict each other, the studio sends back a question before touching the file.

Wording you can adapt for the proposal:

Each proof round starts from a single consolidated list of comments from [approver's name]. We reply with a numbered change list, and work begins once you confirm it. Comments that change the agreed brief are quoted separately.

An assistant earns its keep in triage. Paste the client's comments and the agreed brief into the prompt below. Never paste unreleased artwork into an account where chats may be used for training; the section on confidentiality covers that.

Below are a client's comments on proof v02 of the "tasting box" social post. The agreed brief and brand guide are attached.
Turn them into a numbered change list. For each comment give: ID, the format it refers to, the requested change, type (copy, layout, colour, new element), whether it is inside the brief, and a question for the client if it is unclear.
Do not merge comments that refer to different formats. Do not suggest hex codes, fonts or copy that are not in the brand guide. Flag contradictions.

Comments:
1. Can the logo be bigger?
2. The logo looks too big on the banner.
3. The green looks different from our website.
4. Move the offer line up a bit.
5. Can we try a different photo?
6. Not sure about the headline.

This illustrative answer is typical in shape and in errors.

Four corrections before this goes anywhere. Comments 1 and 2 were merged, but one asks for a bigger logo on the social post and the other a smaller one on the banner, so they are separate items for separate formats. The hex code in row 3 appears nowhere in the brand guide, so delete it and measure the website green and the proof green in the same colour space before replying. Comment 5 is marked in scope although the brief fixed three supplier photos, so a replacement is a quote. In row 6 the assistant wrote a headline when the client only said they were unsure; the right output is a question about what bothers them.

Before the list is sent, a designer reads it against the original comments. After the changes, the designer ticks each ID against the new version of the file, not against the list, because a tidy list can describe work that was never done. A proofing tool helps by keeping comments attached to a version. Frame.io, for example, describes feedback on videos, images and docs, and its pricing page shows a free plan.

Fourteen variants, before and after: the clock for one campaign

These are illustrative figures for a studio where a designer takes 15 to 35 minutes per variant by hand. Time your own jobs for two weeks before trusting any sum, including this one.

The saving is 147 minutes, about two and a half hours, on one campaign. At four campaigns a month that is nearly ten hours. Notice where it comes from: the straight exports, the reflows and the admin. The recomposed group barely moves, and it should not, because that is design work. Setup is a one-off cost. Allow a day for presets, the checker script and the note template, then keep adding to them from the miss log described below.

The checks only a designer can make before anything leaves

Whatever the tiers say, these seven checks stay human. Together they take roughly half an hour on a fourteen-variant campaign.

Generated fills, client files and confidentiality

Generative tools belong in Tier 3 because they invent pixels. In Photoshop, the Crop tool's handles are dragged outwards, Generative Expand is chosen in the Contextual Task Bar, and a model is picked from the picker. Adobe's help page, last updated in February 2026, lists Firefly Fill & Expand, Firefly Image 3 and Firefly Image 1, and says that leaving the prompt blank lets Photoshop extend the image. Choose a variation in the Properties panel, then zoom to 200% and inspect the seams, edges, repeated textures and anything that looks like lettering.

The model choice affects what you can promise. Adobe's Firefly page labels its own models "commercially safe" and lists partner models separately as "models created by others", and Photoshop's Generative Fill picker now offers partner models such as Gemini and FLUX beside the Firefly ones. Note which model made each fill, because the safety label does not travel to the partner models. The studio's own contract has to say whether generative tools may touch client files, and the handoff note should record where they did. What designers can promise clients about Firefly output goes into the detail, and AI clauses for agency contracts covers disclosure, ownership and approvals.

Unreleased artwork is confidential. Consumer ChatGPT and Claude plans let you switch off the use of chats for model training in the privacy settings, while business plans do not train on business content by default. Check your NDA and which account the studio uses before uploading a client's proofs. Scripts on your own machine involve no upload at all, which is one more reason to prefer them for Tier 1 work.

Tools that fit these rules, and what I could verify

Choose tools after the rules exist, and expect each to cover one slice. This table lists what the vendors' own pages say as of November 2026.

For a wider look at software and licences, AI graphic design software for a small studio compares workflow fit, and brand kits and templates for client work covers the locked-template side of rule three.

A miss log that rewrites the rules

Every miss, however small, goes into a log with five columns: when, which job, what was missed, which tier it came from and what changed. Review it monthly. When the same miss appears three times, change a rule, a template or the script. This illustrative log shows the loop using the cases above.

Within two months the log becomes the studio's real production manual, because every line in it is something that went wrong here, with these clients and this software. For handling briefs and revisions upstream of production, briefs, mock-ups and revisions with AI is the natural companion to this workflow.

Questions studios ask before automating production

Can ChatGPT or Claude resize a folder of images for me?

Not reliably through chat. An assistant can plan the work or write a short script, but batch resizing belongs in presets, actions or a script you can rerun and audit. Ask the assistant to write it, run it on a copy of one finished job, then compare the output dimensions and file count with what you expected.

Do I have to tell clients I used AI for resizing?

Resizing, renaming and packaging change no content, so most studios treat them like any other software feature. Generated pixels are different, because a generative expand invents image content. Whether you must disclose depends on your contract and the client's rules, so put a clause in the proposal, record each use in the handoff note and ask a solicitor to word it.

How many proof rounds should a proposal include?

There is no standard number. I'd start with two and state exactly what counts as a round: one consolidated list from one named approver, on a numbered version. Add what an extra round costs. Templated work may need fewer rounds, and a brand identity project more, so judge by how many decisions are still open.

What should a small studio automate first?

Start with the first tier: folder creation and file naming, plus a checker script. A miss is cheap to catch, the saving is immediate (about twenty minutes per campaign in the sums above) and the habits carry over into resizing and proofing. Leave generative fills and print-critical files to a designer.

Further reads

• ChatGPT Prompts for Graphic Designers: Briefs, Concepts, Client Emails — Prompt wording for briefs, concepts and client emails.
• AI for Graphic Design Layouts: Presenting Options for Client Approval — How to present layout options so clients sign off sooner.
• AI Image to Vector: Clean Logos for Print and Signage Clients — Turning client logos into clean vectors for print and signage.
• Midjourney for Graphic Designers: Client Work, Licences and Limits — Licence limits when concepts start in Midjourney.
• AI Content Workflow for Agencies: From Brief to Approved Draft — A brief-to-approved-draft workflow for written content.
• AI for Printing Companies: Proof Approvals and Job Status Updates — Proof email, approval log, reminder and ready-for-collection message: the message flows a print shop can draft with AI, and the checks that keep them accurate.
• Design Studios Using AI: Briefs, Moodboards and Client Revisions — A studio-side workflow for moving from a messy brand brief to approved visual directions and controlled revision rounds.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Adobe Photoshop help (Image Processor, Generative Expand), Adobe Illustrator help (export), Figma Buzz page, Frame.io plans and pricing pages, vendor pages on Firefly commercial use and privacy defaults.