---
name: graphic-design
description: "Your professional process for making a flyer, poster, menu, social graphic, sign, or any visual — gather the brief, classify every attached image, then design it with create_image (ChatGPT Images 2.5: fast drafts, then a sharp final), which preserves the owner's real attached photo and logo at high fidelity inside a professionally designed piece. Load this BEFORE you design any flyer or graphic."
version: 2.6.0
metadata: {"bizer":{"name":"Graphic Design"}}
---

GRAPHIC DESIGN SKILL — your process for making a flyer, poster, menu, social
graphic, sign, certificate, or any designed piece. Load this BEFORE you design.
The bar is simple: it should look like a real designer made it — built from the
owner's REAL assets, never inventions. You beat the competition with PROCESS.

THE ENGINE: create_image runs on ChatGPT Images 2.5, the same models behind
ChatGPT's image work, and EVERY image the owner attaches this turn is
sent to it at HIGH INPUT FIDELITY — so their real photo and real logo are
preserved and woven into a professionally designed piece. 'Make a flyer with
this photo and our logo' is ONE create_image call; the design is free-form and
beautiful AND keeps their actual assets. This is the default for ALL design
work, including requests that use the owner's attached images.

DRAFT FAST, FINALIZE ONCE: 2.5 is two models. stage='draft' (the default) is
Flare: quick, good enough to judge layout, wording and direction. Every first
look and every round of changes is a draft, so the owner iterates in seconds.
stage='final' is Sunburst: slower, sharper detail, cleaner text, bigger canvas.
When the owner approves a draft ('that's it', 'perfect') or is about to post,
print or send it, render the final: stage='final', editLastDesign=true, the
SAME aspectRatio, instruction 'Render the final, high-detail version of this
design; keep everything exactly as it is.' The final is the approved draft
re-rendered, never a new design. Offer it in one line after a draft the owner
likes; don't spend a final on a design they're still changing. Go straight to
final only when they ask for the finished piece in one go. A print PDF
(format='pdf') is always a final. Change ONE thing per draft round when you
can, so the owner sees what each change did, and restate the exact text and
constraints every round: nothing carries over unless you say it again.

MULTIPLE ATTACHED IMAGES — do this every time before designing:
1. LOOK at every image attached this turn and understand each one: business
   logo, product/food/scene photo, or style reference. They ALL ride to the
   model automatically — reference each in your instruction ('use the attached
   butcher photo as the background scene; place the attached Bizer logo at the
   top, reproduced exactly').
1a. GIVE EACH IMAGE ITS JOB. Whenever two or more images are attached, set
   imageNotes — one short line per image, in the order attached, saying what it
   is and what to do with it: ['the new sign to hang on the wall behind the
   counter', 'the drink that replaces the cup on the table']. Several images
   doing DIFFERENT jobs is one call, not one call per change. Without those
   per-image jobs the model can use a couple of them and quietly ignore the
   rest, and the owner gets half the changes they asked for.
1b. CHANGES TO THE DESIGN YOU JUST MADE go with editLastDesign=true EVEN WHEN
   they attach new images for those changes — the previous design is the base
   and the attached images are the elements to put into it. Never make the
   owner re-attach the design or apply their changes one at a time.
2. If the owner says 'use my logo' but attached only a photo, their saved
   Branding logo is what they mean: set includeLogo=true. An attached logo
   ALWAYS wins over the saved one — never substitute a different logo than the
   one they gave you.
3. If no usable logo exists anywhere, say so plainly and point them to
   Marketing → Branding — never design around an invented logo.

ONLY WHAT THEY ASKED FOR: the design contains the owner's subject and the
elements they requested — nothing more. When they attach a photo, THAT photo
is the hero of the piece; build the background, text, and styling around it.
NEVER insert your own persona, mascot, avatar, or any character or scene the
owner didn't ask for ('background + caption' means exactly that). Uninvited
extras are the #1 complaint after a design turn.

REFINEMENT TURNS (critical): when the owner asks for changes to the design
you just made — 'remove the extras', 'change the background', 'bigger text',
'swap the sign for this one' — call create_image with editLastDesign=true,
whether or not they attached anything new. That
loads the previous design and applies the changes TO it, keeping the real
people, product photos, and logos from earlier turns intact. NEVER regenerate
a refinement from text alone: their photo would silently vanish from the redo.
Describe the requested change precisely and state what must stay unchanged.

PHOTOS FROM EARLIER TURNS / RETRIES (critical): attachments only ride the
message they came with. When the owner references photos sent earlier in this
chat — 'use my photo', 'the pictures I sent' — or says 'try again' after a
render that failed and there is no previous design to revise, set
useRecentPhotos=true so their recent photos are pulled from this thread and
fed to the model.

LIKENESS = PHOTO REQUIRED (hard rule): a design that depicts a real, specific
person or pet — a caricature or portrait of the owner, 'me and my puppy', a
card with their family on it — MUST set requiresRealPhotos=true. With no real
photo available the tool then refuses to render, and that refusal is the
correct platform behavior, not an error: without a photo the model can only
invent a lookalike stranger, and an invented likeness must never be created,
let alone presented as them. Relay the refusal, ask for the photo, and never
re-call with the flag off to force a render. (Generic, unnamed people — 'a
chef tossing pizza' — need no photo and no flag.)

HONESTY (absolute): report what went into the piece using the tool result's
own composition report. If the owner says the result altered or replaced their
image or logo, do not argue — acknowledge it and regenerate with a sharper
preserve instruction ('reproduce the attached logo EXACTLY as given, do not
redraw it'). A fabricated or lookalike logo on a business's marketing is a
serious failure. Same rule for likenesses: a render made with no photo of the
owner or their pet is an artistic invention — say so; never call it them.

STEP 1 — GET THE BRIEF (be inquisitive; ask, don't assume). The owner usually
isn't a designer, so do the translating. In ONE short message ask only the few
things that change the design, each with a smart default they can just confirm:
- SIZE: "I'll make a standard US Letter flyer (8.5x11", portrait) unless you'd
  rather a postcard, half-page, or a square social post." (If they just say
  "flyer," assume US Letter portrait.)
- REFERENCE: "Got a look you like — your own past design, a brand you admire, or
  something you saw online? Upload it and I'll match the style." (If they attach
  one, use it as the look to emulate or edit from.)
- PURPOSE + CALL TO ACTION: "When someone sees this, what's the ONE thing you want
  them to do — call, visit, scan a code, come to an event, buy? And the must-have
  details: date, time, price, offer, phone, address?"
- BRAND: use their colors/logo if you know them (check their profile / Branding);
  if not, ask for brand colors and whether to include a logo.
Don't interrogate — 3-4 quick questions, then design. If they already gave the
answers, skip ahead and design.

CHECK THEIR SAVED LOOK FIRST. If the owner has a saved design theme (read
pages/learned/design-theme.md), use it: match that palette, typography, imagery,
and mood so this piece looks like it belongs with their other work — don't ask for
what's already recorded there. If they have NO saved theme, this is a great moment
to offer the Design Theme skill: "want me to learn your style? Upload a few designs
you like and I'll match them from now on."

STEP 2 — RELAY, DON'T REWRITE (the most important rule). The image model is
SMART — it designs beautifully from plain words, the same way ChatGPT does.
Your instruction is the owner's request, passed through nearly verbatim, plus
only what's mechanically needed: the EXACT text in quotes (correctly spelled)
and which attached image is which ('the attached photo of the girl', 'the
attached logo'). Do NOT expand their request into an elaborate brief. Do NOT
invent subjects, characters, scenery, styles, colors, or moods they didn't
mention — every detail you add is something they didn't ask for, and uninvited
detail is the #1 design complaint. 'Patriotic background with the caption X'
goes to the model as almost exactly that. Trust the model with everything the
owner left unspecified.

Add design direction ONLY when the owner asked for it ('make it pop', 'you
pick the style') or their saved design theme provides it — then a short line
or two (brand colors, mood) is plenty. Set aspectRatio to fit the use — "3:4"
US Letter flyer, "1:1" square social post, "9:16" story, "16:9" banner — and
when the owner will PRINT or download it, set format="pdf".

STEP 3 — REVIEW & REFINE (this is what separates good from bad). Before you hand it
over, look at the result critically: Is every word spelled right? Is the hierarchy
clear and the CTA obvious? Is it aligned and uncluttered? Does it deliver the
brief? If anything's off, say what you'll improve and regenerate with a sharper
prompt. Offer one refinement round — "want it bolder, different colors, more room
for the photo?" Treat the first output as a draft, the way a real designer would.
When they're happy, render the final (see DRAFT FAST, FINALIZE ONCE).
