TL;DR: Fill nine fields, and the block below assembles a working GPT Image 2 prompt. Put the exact string in quotation marks, describe the font instead of naming it, say where the text sits, and forbid any other text. Then pick one of 22 job templates and swap in your own copy.
What is a GPT Image 2 prompt generator?
A GPT Image 2 prompt generator is a fill-in-the-blanks template that turns nine design decisions into one correctly ordered prompt string. There is nothing to install. The value is the field list, because the model rewards prompts that keep the literal text separate from everything describing how that text should look, and most people mash the two together.
Fill this in. Only subject, purpose and one text block are mandatory.
SUBJECT → what the image is of, in one clause
PURPOSE → poster / logo / UI mock / label / sign / cover / social card / infographic
TEXT BLOCK 1 → the exact string, in quotes, with the capitalisation you want
PLACEMENT 1 → where it sits, and what surface it sits on
TEXT BLOCK 2 → the second string, in quotes (optional)
PLACEMENT 2 → where that one sits, relative to block 1
FONT CHARACTER → weight, width, contrast, terminals, letter spacing (words, not names)
PALETTE → two to four colours, named or hex
CONSTRAINTS → no other text, no watermark, no logos, spelling rules
SIZE → WIDTHxHEIGHT in pixels, both edges multiples of 16
Assemble in this order. Scene first, subject second, text third, constraints last:
A [PURPOSE] of [SUBJECT]. [ONE SENTENCE OF SCENE AND LIGHTING.]
Text (EXACT, verbatim, no extra characters):
"[TEXT BLOCK 1]" — [PLACEMENT 1]
"[TEXT BLOCK 2]" — [PLACEMENT 2]
Typography: [FONT CHARACTER]. Palette: [PALETTE].
[COMPOSITION NOTE: framing, viewpoint, where the negative space goes.]
Render each string exactly once and make every character legible.
No other text, no watermark, no logos, no signature.
That skeleton is not a stylistic preference. OpenAI's own prompting fundamentals ask you to write prompts in a consistent order, to "State exclusions and invariants explicitly", and for complex requests to "use short labeled segments or line breaks instead of one long paragraph". (OpenAI cookbook, accessed Aug 27, 2026)
Why is GPT Image 2 the model you use when the image has words in it?
Because text rendering is the thing it is genuinely good at, and that is not a self-serving claim. OpenAI lists it as a headline capability, "with crisp lettering, consistent layout, and strong contrast inside images", and recommends gpt-image-2 specifically for "text-heavy images" and "text-in-image work".
More usefully, a competing lab published a blind comparison that puts it first. Ideogram's 4.0 technical write-up includes an internal designer-preference arena where "Voters were not told which model produced each image", and its own figure caption concedes that "Ideogram 4.0 sits in second place overall and first among open-weight models". First place, across nine pipelines, went to GPT Image 2.
Read that as a benchmark, not a verdict: it is one vendor's internal arena, on its own task mix, published to sell its own model. It is worth citing precisely because the result is inconvenient for the publisher. If typography is your whole job, test both on your actual artwork. Ideogram has a different and equally documented prompt discipline, which is covered in the Ideogram 4.0 prompt generator.
OpenAI is also candid about the ceiling. Its own limitations list says of text rendering: "Although significantly improved, the model can still struggle with precise text placement and clarity." Best in class and still imperfect are not contradictory statements.
How do you specify the exact string you want rendered?
Put it in quotation marks, exactly as you want it to appear, and say that no other text is allowed. This is documented twice, in two different OpenAI docs, in almost the same words.
The ChatGPT product docs: "Keep in-image text short and specify it precisely. Put the exact text in quotation marks, preserve the capitalization you want, and describe its font style, size, color, and placement." They add a rule most guides miss: "For an uncommon name, spell out the letters when accuracy matters. State whether any other text is allowed." (learn.chatgpt.com, accessed Aug 27, 2026)
The API cookbook says the same thing for marketing work: "Put the exact copy in quotes, demand verbatim rendering (no extra characters), and describe placement and font style." Its fundamentals section adds the letter-by-letter trick again: "For tricky words (brand names, uncommon spellings), spell them out letter-by-letter to improve character accuracy."
So a text specification has four parts, and the first two are non-negotiable:
Text (EXACT, verbatim, no extra characters):
"KESTREL & CO."
Spelled: K-E-S-T-R-E-L, space, ampersand, space, C-O, period.
Capitalisation: all caps exactly as written, including the period.
No other text anywhere in the image.
The spelled-out line looks absurd until the first time a brand name comes back as "KESTRAL". Use it for anything invented, anything non-English, and anything with a doubled letter.
Where should the text sit, and how should it relate to the subject?
Say it in two moves: an absolute position, then a relationship to something already in the frame. Absolute position alone drifts. Relationship alone floats.
OpenAI's composition guidance asks you to call out placement directly, with examples like "logo top-right" and "subject centered with negative space on left". That works, but it is a first pass. The version that survives contact with a real subject names the anchor:
"AUTUMN INTAKE" — set across the top third, baseline aligned with the horizon line,
kerned wide enough that the tallest tree does not touch any letter.
"Applications close 14 November" — one line, centred beneath the headline,
sitting inside the empty sky, at roughly one third the headline's cap height.
Three things are doing work there: a zone (top third), an anchor (the horizon line, the empty sky), and a size relationship (one third the cap height). Relative type sizing is far more reliable than asking for points or pixels, because the model has no ruler.
For layouts where placement has to be exact rather than described, be realistic. OpenAI lists composition control as a live limitation: the model "may have difficulty placing elements precisely in structured or layout-sensitive compositions". A pixel-exact grid is a design tool's job.
Should you name a typeface or describe one?
Describe it. Then, optionally, add a name as a hint at the end.
The documented instruction is to "describe its font style, size, color, and placement". Style, not name. And a font name is a surprisingly weak instruction: it tells the model which visual cluster you mean only if that name was strongly represented in training, and it tells it nothing about the specific qualities you actually care about.
A description does. Five dials cover almost everything:
| Dial | Vocabulary that works | What it changes |
|---|---|---|
| Weight | hairline, light, regular, semibold, heavy, black | Stroke thickness and page colour |
| Width | condensed, normal, extended | How much horizontal room the string eats |
| Contrast | monoline, low contrast, high contrast (thick stems, thin hairlines) | Whether it reads modern or classical |
| Terminals | slab serifs, sharp wedge serifs, rounded terminals, cut-off flat terminals | The character of each letter's ending |
| Spacing | tight kerning, generous letter spacing, all caps widely tracked | Rhythm, and whether small text stays legible |
So instead of "Futura", write "geometric sans-serif, medium weight, near-perfect circular bowls, tight kerning, all caps". You get what you meant, and you get it without asking the model to reproduce a licensed typeface.
Worth being straight about one thing: OpenAI does not forbid font names, and its own cookbook slide prompt asks for "modern sans-serif typography like Inter". Not every vendor agrees. Ideogram's docs state outright that naming a typeface is not possible on their model. If you use both, do not carry the habit across.
How much text is realistic before quality degrades?
OpenAI publishes no number, and you should be suspicious of any page that quotes one. What it publishes is a direction and a workflow.
The direction: "Keep in-image text short and specify it precisely." The workflow, for anything denser: "For dense copy or production-critical typography, review every word and finish the asset in a design tool when needed." And a settings rule from the cookbook: use medium or high quality for small text, dense information panels and multi-font layouts, rather than the fast low setting.
What follows is craft advice, not vendor guidance, drawn from what actually survives review:
- One to three text blocks per image. A headline, a subhead, and one small line. Beyond that, error probability compounds because every block is another chance for one bad character.
- Headline strings of roughly two to six words. Long enough to say something, short enough that a single glance proofreads it.
- Small text is a separate risk from long text. A five-word footnote at eight percent of the frame height is harder than a five-word headline at thirty percent. Raise
quality, or raise the size of the type. - Numbers and punctuation are the weak points. Dates, prices, ampersands and hyphens fail more often than plain letters, so spell those out too.
- Anything you would have to read twice, typeset yourself. Body copy, legal lines, tables of numbers, addresses.
The failure modes and their causes are worth understanding on their own terms; why the text in your AI image is garbled covers what is actually going wrong underneath.
Why do long paragraphs fail?
OpenAI does not publish an explanation, so treat this as reasoning rather than vendor guidance. A paragraph is not one instruction, it is dozens, and every extra character is another independent chance to be wrong. Nothing in the loop is spell-checking, and nothing is laying out a text run the way a font engine would. The model is producing pixels that look like letters, and the probability that all of them are right falls off as the count rises.
That is also why the fix is never "try harder". Splitting one paragraph into a headline plus a three-word kicker usually works on the first attempt, while the paragraph never converges no matter how strictly you phrase it.
The second reason is prompt economy. The prompt field accepts a great deal of text: OpenAI's spec says "The maximum length is 32000 characters for the GPT image models". But a 400-word paragraph of body copy inside a prompt is competing for attention with your composition, palette and constraint instructions, and the model has no way to know that this particular block is the sacred one. Long prompts are fine; long rendered strings are the problem.
Which parameters exist, and which ones don't?
This is where most GPT Image 2 guides quietly go wrong, because image models share parameter names and it is easy to import one vendor's field into another vendor's model. Checked against OpenAI's published OpenAPI specification on August 27, 2026:
| Parameter | Exists on gpt-image-2? | Notes |
|---|---|---|
size | Yes | WIDTHxHEIGHT pixel string or auto (default) |
quality | Yes | low, medium, high, auto (default) |
background | Yes | transparent, opaque, auto; transparency is in preview |
output_format | Yes | png (default), jpeg, webp |
output_compression | Yes | 0 to 100, JPEG and WebP only |
moderation | Yes | auto (default) or low |
n | Yes, on the Images API | 1 to 10; not available on the Responses API tool |
aspect_ratio | No | Zero occurrences in the spec. Use size |
negative_prompt | No | Zero occurrences in the spec. Use constraint sentences |
seed | No | Not published for the image endpoints |
style | No | dall-e-3 only |
Two of those absences trip people constantly. There is no aspect-ratio field, and asking for "16:9" in prose does nothing; the wrong image size problem is almost always this. And there is no negative prompt, so exclusions live in the prompt as sentences. Which models actually accept a negative-prompt field, and which only pretend to, is mapped in the negative prompt support matrix.
The size rules are worth memorising, because an invalid size is an error rather than a rounding: both edges must be multiples of 16, the long-to-short ratio must not exceed 3:1, total pixels must sit between 655,360 and 8,294,400, and the maximum edge is 3840px. Anything above 2560x1440 is flagged experimental.
Which surface should you send the prompt to?
Three, and they do not treat your prompt the same way. This matters more than it sounds, because a generator's whole promise is that the string you wrote is the string that runs.
| Feature | Images API | Responses API tool | ChatGPT app |
|---|---|---|---|
| Prompt used exactly as written | Revised automatically | Not documented | |
| Exact pixel size control | Not documented | ||
| Batch variants in one call | n up to 10 | ||
| Mask-guided editing | Select an area in the UI | ||
| Multi-turn refinement in context | |||
| Transparent background (preview) | Not documented |
The row that surprises people is the first one. On the Responses API image tool, OpenAI documents that the mainline model "will automatically revise your prompt for improved performance", and returns what it actually used in a revised_prompt field. That is often an improvement. It is also why an identical prompt can produce different typography on two surfaces. If you are testing a template, test it on the Images API, where what you send is what runs.
How do you fix one wrong letter?
Not by rewriting the prompt. A fresh prompt regenerates the whole image and you lose the ninety percent that was right.
OpenAI's guidance is to "start with a clean base prompt and refine with small, single-change follow-ups", and, in the ChatGPT docs, to "Adjust one element at a time so the composition and other important details do not drift." The cookbook adds the part people forget: repeat the preserve list on every iteration.
Three repair prompts that cover most of it. One wrong character:
Edit the attached image. The headline currently reads "KESTRAL & CO." It must read
exactly "KESTREL & CO." — spelled K-E-S-T-R-E-L, space, ampersand, space, C-O, period.
Change only those characters. Keep the same typeface, weight, size, kerning, colour,
baseline and position. Do not alter the photograph, palette, lighting, crop, or any
other text. Add nothing.
Text in the right words but the wrong place:
Edit the attached image. Keep the string "AUTUMN INTAKE" exactly as spelled and styled.
Move it so its baseline aligns with the horizon line and it sits in the upper third,
centred horizontally. Do not change the letterforms, weight, colour, or size. Preserve
the photograph, palette, lighting, crop and every other element exactly as they are.
Text you never asked for:
Edit the attached image. Remove all text except the headline "AUTUMN INTAKE" and the
line "Applications close 14 November". Remove the invented logo in the lower right and
the caption along the bottom edge. Fill those areas so they match the surrounding
background naturally. Change nothing else: same crop, palette, lighting and typography.
If you reach for a mask to make that surgical, read OpenAI's own caveat first: "Masking with GPT Image is entirely prompt-based. The model uses the mask as guidance, but may not follow its exact shape with complete precision." A mask is a strong hint about where to work, not a selection tool. And the cookbook is realistic about the rest: "If text fidelity is imperfect, keep the prompt strict and iterate—small wording/layout tweaks usually improve legibility."
22 copy-paste templates, organised by job
Swap the quoted strings and the bracketed fields. Everything else is deliberate. Each one already carries the exclusion line, because that is the line that saves a re-run.
Posters
A screen-printed event poster. Flat colour, visible paper grain, two-colour risograph feel.
Text (EXACT, verbatim, no extra characters):
"[HEADLINE, 2-4 WORDS]" — across the top third, tightly kerned, all caps
"[VENUE]" — lower left, one line, small
"[DATE]" — lower right, one line, same size as the venue line
Typography: heavy condensed sans-serif, monoline, flat terminals, wide tracking on
the headline only. Palette: [COLOUR A], [COLOUR B], paper white.
Subject: [ONE-SENTENCE GRAPHIC], centred, occupying the middle half of the frame.
Render each string exactly once and make every character legible.
No other text, no watermark, no logos.
A cinematic film-style poster, photographic, dramatic low-key lighting from the left.
Text (EXACT, verbatim, no extra characters):
"[TITLE]" — bottom third, centred, dominant, all caps
"[TAGLINE, UNDER 8 WORDS]" — directly above the title, one line, small caps, letter-spaced
Typography: high-contrast serif with sharp wedge serifs for the title; thin
letter-spaced sans-serif for the tagline. Palette: desaturated [COLOUR], deep shadow, warm highlight.
Subject: [CHARACTER OR OBJECT], centred, upper two thirds, silhouetted against [BACKGROUND].
Leave the bottom third clear of subject detail so the title reads cleanly.
Render each string exactly once. No credits block, no billing text, no studio logos, no other text.
A minimal announcement poster on a plain [COLOUR] field. No photograph.
Text (EXACT, verbatim, no extra characters):
"[STATEMENT, 3-6 WORDS]" — centred, occupying the middle 60% of the width,
set on [1-3] lines with even line lengths
"[SMALL DETAIL LINE]" — bottom centre, one line, 15% of the headline's cap height
Typography: extended grotesque, semibold, tight kerning, flat terminals, no italics.
Palette: [COLOUR] ground, [COLOUR] type. Generous even margins on all four sides.
Render each string exactly once and make every character legible.
No other text, no ornament, no watermark, no logos.
Logo lockups
An original, non-infringing wordmark logo for [COMPANY], a [WHAT THEY DO].
Fully transparent background.
Text (EXACT, verbatim, no extra characters):
"[COMPANY NAME]" — spelled [SPELL IT OUT LETTER BY LETTER]
Capitalisation exactly as written.
Typography: [FONT CHARACTER: weight, width, contrast, terminals, spacing].
The mark is a single centred lockup with generous even padding and clean alpha edges.
Flat design, minimal strokes, no gradients, no drop shadows, no 3D.
Must read clearly at 24px and at poster size.
No solid backdrop, no scenery, no checkerboard, no shadow, no other text, no tagline.
An original badge-style roundel logo for [COMPANY]. Fully transparent background.
Text (EXACT, verbatim, no extra characters):
"[COMPANY NAME]" — curved along the top inner edge of the circle, all caps, evenly tracked
"[EST. YEAR]" — curved along the bottom inner edge, smaller
Both spelled: [SPELL OUT ANY UNUSUAL WORD].
Typography: sturdy slab serif, medium weight, even stroke, all caps.
Centre device: [SIMPLE SYMBOL], flat, single colour, strong silhouette.
Two concentric rules separate the type ring from the centre device.
Palette: single colour [COLOUR] on transparent.
No solid backdrop, no checkerboard, no shadow, no other text.
A flat app icon on a [COLOUR] rounded-square field, no device frame.
Text (EXACT, verbatim, no extra characters):
"[SINGLE LETTER OR 2-CHARACTER MONOGRAM]" — centred, optically not mathematically,
filling roughly 55% of the tile height
Typography: geometric sans-serif, black weight, circular bowls, flat terminals.
Palette: [COLOUR] tile, [COLOUR] letter, no gradient.
Perfectly symmetrical margins. Crisp edges, no bevel, no gloss, no shadow.
No other text, no border, no watermark.
UI mockups
A realistic mobile app UI mockup for [PRODUCT], shown as a shipped product rather
than a concept sketch. Place it in a plain phone frame on a light neutral background.
Text (EXACT, verbatim, no extra characters):
Screen title: "[TITLE]"
Primary button: "[BUTTON LABEL]"
Section headers: "[HEADER 1]", "[HEADER 2]"
List items: "[ITEM 1]", "[ITEM 2]", "[ITEM 3]"
Layout: header with title and a back chevron, then [SECTION 1], then [SECTION 2],
then the primary button pinned to the bottom.
Typography: clean sans-serif UI type, regular weight for body, semibold for headers.
Palette: white background, [ACCENT COLOUR] for the primary button only.
Use real spacing and hierarchy. No lorem ipsum, no placeholder rectangles,
no decorative illustration, no other text.
A realistic desktop dashboard UI mockup for [PRODUCT], full-bleed, no browser chrome.
Text (EXACT, verbatim, no extra characters):
Page title: "[TITLE]"
Metric cards, label then value: "[LABEL 1]" / "[VALUE 1]", "[LABEL 2]" / "[VALUE 2]",
"[LABEL 3]" / "[VALUE 3]"
Sidebar items: "[NAV 1]", "[NAV 2]", "[NAV 3]", "[NAV 4]"
Layout: 220px left sidebar, then a three-card metric row, then one wide line chart
with a labelled x-axis, then a data table with four visible rows.
Typography: neutral sans-serif, tabular figures in the metric values.
Palette: near-white canvas, one [ACCENT COLOUR], grey rules.
Numbers must be legible at 100%. No lorem ipsum, no other text, no watermark.
Product labels and packaging
A photorealistic product photograph of a [CONTAINER] on a [SURFACE], soft directional
light from the upper left, shallow depth of field, real material texture.
Label text (EXACT, verbatim, no extra characters):
"[BRAND]" — top of the label, all caps, dominant, spelled [SPELL IT OUT]
"[PRODUCT NAME]" — beneath the brand, sentence case
"[VOLUME OR WEIGHT]" — bottom of the label, small, one line
Typography: [FONT CHARACTER]. Label palette: [COLOUR A] on [COLOUR B].
The label wraps the container naturally, with realistic curvature and highlight.
Render each string exactly once and keep every character sharp and unwarped.
No barcode, no ingredient list, no regulatory text, no other text, no watermark.
A photorealistic front panel of a [PRODUCT] carton, shot straight on, studio lighting,
subtle box shadow on a [COLOUR] seamless background.
Text (EXACT, verbatim, no extra characters):
"[BRAND]" — upper left, all caps
"[PRODUCT NAME]" — centred, dominant, on [1-2] lines
"[THREE-WORD BENEFIT LINE]" — beneath the product name, small, letter-spaced
"[NET WEIGHT]" — bottom right corner, smallest text on the panel
Typography: [FONT CHARACTER] for the product name; the same family, lighter weight,
for everything else. Palette: [COLOUR A], [COLOUR B], one accent.
Keep the panel flat and square to camera so all type is undistorted.
No barcode, no nutrition panel, no other text, no watermark.
A product cutout of [PRODUCT] on a fully transparent background, isolated subject,
no scene. Clean alpha edges with no fringing or halo.
Label text (EXACT, verbatim, no extra characters):
"[BRAND]" and "[PRODUCT NAME]" exactly as spelled, in their existing positions.
Preserve the product's geometry, proportions, colour and label legibility exactly.
Even, soft lighting with no cast shadow on the ground.
No solid backdrop, no scenery, no checkerboard, no reflection, no drop shadow,
no other text, no watermark.
Signage and environmental
A photorealistic photograph of a [SHOP TYPE] storefront at [TIME OF DAY],
shot from across the street at eye level, natural light, real street texture.
Sign text (EXACT, verbatim, no extra characters):
"[SHOP NAME]" — on the fascia board above the windows, dominant,
spelled [SPELL IT OUT LETTER BY LETTER]
"[ONE-LINE DESCRIPTOR]" — smaller, on the glass of the left window
Typography: [FONT CHARACTER], [PAINTED / GILDED / NEON / ROUTED] finish.
Palette: [COLOUR] fascia, [COLOUR] lettering.
The sign must sit flat on the fascia with correct perspective and realistic wear.
Render each string exactly once. No street signs, no menu boards, no other legible
text anywhere in the frame, no watermark.
A photorealistic interior wayfinding sign mounted on a [MATERIAL] wall,
shot straight on, soft even architectural lighting.
Text (EXACT, verbatim, no extra characters):
"[DESTINATION 1]" with a right-pointing arrow — top row
"[DESTINATION 2]" with a right-pointing arrow — second row
"[DESTINATION 3]" with a left-pointing arrow — third row
Typography: humanist sans-serif, medium weight, generous letter spacing,
high contrast against the sign panel. All rows share one baseline grid and one type size.
Palette: [COLOUR] panel, [COLOUR] type, [COLOUR] arrows.
Arrows sit on the same baseline as their text, not above or below.
No other text, no icons beyond the arrows, no watermark, no branding.
A realistic billboard mockup beside a [ROAD TYPE] at [TIME OF DAY], wide shot,
the billboard occupying the right half of the frame in correct perspective.
Billboard text (EXACT, verbatim, no extra characters):
"[HEADLINE, UNDER 6 WORDS]"
Typography: bold sans-serif, high contrast, centred, clean kerning, generous margins.
Palette: [COLOUR] ground, [COLOUR] type.
Subject on the billboard: [PRODUCT OR IMAGE], left third, leaving the right two thirds
clear for the headline.
Ensure the text appears once and is perfectly legible at this distance.
No other text, no logos, no watermarks, no additional signage in the scene.
Book covers
A non-fiction book cover, front face only, flat, straight to camera, no mockup shadow.
Text (EXACT, verbatim, no extra characters):
"[TITLE]" — upper half, dominant, on [1-3] lines with balanced line lengths
"[SUBTITLE]" — directly beneath the title, one line, 30% of the title's cap height
"[AUTHOR NAME]" — bottom centre, spelled [SPELL IT OUT LETTER BY LETTER]
Typography: [FONT CHARACTER] for the title; the same family, light weight,
for the subtitle and author. Palette: [COLOUR] ground, [COLOUR] type, one accent.
Central graphic: [SIMPLE CONCEPTUAL DEVICE], flat, geometric, occupying the middle band.
Generous even margins. Render each string exactly once.
No publisher logo, no barcode, no endorsement quote, no other text.
An atmospheric fiction book cover, front face only, flat to camera.
Painterly [GENRE] illustration with [LIGHTING] light.
Text (EXACT, verbatim, no extra characters):
"[TITLE]" — lower third, dominant, all caps, arranged on [1-2] lines
"[AUTHOR NAME]" — top centre, smaller, letter-spaced, spelled [SPELL IT OUT]
Typography: [FONT CHARACTER]. The type sits over the darkest area of the illustration
with enough contrast to stay readable at thumbnail size.
Palette: [COLOUR A], [COLOUR B], [ACCENT].
Composition: [SUBJECT] centred in the upper two thirds, with the lower third
kept visually quiet for the title.
No series line, no publisher logo, no quote, no barcode, no other text.
Social cards
A square social quote card, flat design, no photograph.
Text (EXACT, verbatim, no extra characters):
"[QUOTE, UNDER 18 WORDS]" — centred, on [2-4] lines with balanced line lengths,
occupying the middle 70% of the card
"[ATTRIBUTION]" — bottom centre, small, preceded by an em dash
Typography: high-contrast serif for the quote, letter-spaced sans-serif caps for
the attribution. Palette: [COLOUR] ground, [COLOUR] type, one accent rule above the attribution.
Generous even margins. Nothing touches the edges.
Render each string exactly once. No logo, no handle, no other text, no watermark.
A landscape launch announcement card for [PRODUCT], flat design with one
geometric accent shape in the lower right.
Text (EXACT, verbatim, no extra characters):
"[ANNOUNCEMENT, 3-5 WORDS]" — left-aligned, upper left, dominant, on [1-2] lines
"[SUPPORTING LINE, UNDER 12 WORDS]" — beneath it, one line, 25% of the headline's cap height
"[DATE]" — bottom left, smallest text on the card
Typography: [FONT CHARACTER] for the headline; same family, regular weight, beneath.
Palette: [COLOUR] ground, [COLOUR] type, [ACCENT] shape.
Keep the right third clear of type so the accent shape has room.
Render each string exactly once. No logo, no URL, no other text, no watermark.
A square podcast episode card. Photographic portrait of [SUBJECT] on the right half,
flat [COLOUR] panel on the left half.
Text (EXACT, verbatim, no extra characters):
"EPISODE [NUMBER]" — top of the left panel, small, all caps, letter-spaced
"[EPISODE TITLE]" — beneath it, dominant, on [2-3] lines, left-aligned
"[GUEST NAME]" — bottom of the left panel, spelled [SPELL IT OUT LETTER BY LETTER]
Typography: condensed sans-serif, heavy weight for the title; light weight for the rest.
Palette: [COLOUR] panel, [COLOUR] type, natural skin tones in the photograph.
The portrait is cropped at the shoulders, looking toward the type, not at the camera.
Render each string exactly once. No show logo, no waveform graphic, no other text.
Infographics, slides and diagrams
A clean process infographic on a white background, designed as a classroom handout.
Text (EXACT, verbatim, no extra characters):
Title: "[TITLE]"
Step labels, left to right: "[STEP 1]", "[STEP 2]", "[STEP 3]", "[STEP 4]"
One caption under each step: "[CAPTION 1]", "[CAPTION 2]", "[CAPTION 3]", "[CAPTION 4]"
Layout: title across the top, then four evenly spaced steps in a horizontal row,
connected by simple arrows, with captions directly beneath their step.
Typography: clear sans-serif, semibold labels, regular captions, generous line spacing.
Palette: white ground, one [ACCENT COLOUR], grey rules.
Every label must be legible. Avoid tiny text and decorative clutter.
No other text, no logo, no watermark.
One pitch-deck slide, landscape, that looks like it belongs in a real fundraising deck.
Text (EXACT, verbatim, no extra characters):
Slide title: "[TITLE]"
Metric labels and values: "[LABEL 1]" / "[VALUE 1]", "[LABEL 2]" / "[VALUE 2]",
"[LABEL 3]" / "[VALUE 3]"
Footnote: "[SOURCE LINE]"
Layout: title top left, three metric blocks across the upper half, one bar chart
below with a labelled axis, footnote bottom left in the smallest size on the slide.
Typography: modern sans-serif, semibold title, tabular figures for the values.
Palette: white ground, muted [COLOUR] chart, grey rules.
Highly readable text, clear data hierarchy, polished spacing.
No clip art, no stock photography, no gradients, no shadows, no other text.
A labelled comparison diagram on a light neutral background, drawn as a clean
two-column layout, no photograph.
Text (EXACT, verbatim, no extra characters):
Title: "[TITLE]"
Column headers: "[OPTION A]", "[OPTION B]"
Row labels, top to bottom: "[ROW 1]", "[ROW 2]", "[ROW 3]", "[ROW 4]"
Layout: title across the top, two columns of equal width, four rows separated by
thin horizontal rules, a simple check or cross mark in each cell.
Typography: neutral sans-serif, semibold headers, regular row labels, one type size for all rows.
Palette: neutral ground, [COLOUR] for the left column accent, grey for the right.
Keep labels concise and every character sharp.
No other text, no logo, no watermark, no decorative icons.
Localising a finished asset is its own template, and the instruction is unusually strict because everything except the words must survive:
Translate all text in the attached image into [LANGUAGE]. Translate verbatim and
accurately, with no added or removed words. Keep the typography style, weight,
placement, spacing and hierarchy exactly as they are. Do not reflow the layout unless
a translated string genuinely does not fit. Change nothing else: no edits to logos,
icons, imagery, palette or crop.
Turning the generator into something you reuse
Every template above has the same shape: a stable skeleton, and three or four fields that change per job. Copying a template out of a blog post and hand-editing five placeholders works once. On the fortieth poster it is how strings drift, exclusions get dropped, and a client's brand name comes back misspelled because someone forgot the spell-it-out line.
That gap is what Prompt Architects fills, and it is worth being precise about what we do and do not do. We do not generate images. We generate, structure and store the prompt. The image-prompt tooling turns a rough description into a structured prompt with the constraint block intact, in about two seconds. The Prompt Library keeps the nine-field skeleton as a saved template, and variables let you leave {{HEADLINE}}, {{PALETTE}} and {{SIZE}} as fields you fill per job instead of retyping the sentence. If the underlying discipline is new to you, the lighting vocabulary for image prompts is the companion piece to the typography vocabulary above.
On plans, plainly: image prompt generation is on Pro, Advanced and Team, verified on our own /pricing page on August 27, 2026, and the Image Prompt Library is Advanced and Team. The Free plan exists and includes "5 prompt enhancements per day, forever" according to our /faq page, checked the same day. It is enough to test the workflow on a real poster before you decide anything.
A checklist before you press generate
- Is every rendered string inside quotation marks, in the capitalisation you want?
- Is anything unusual spelled out letter by letter?
- Does each string have a placement zone and an anchor it relates to?
- Is the font described by weight, width, contrast, terminals and spacing, rather than only named?
- Is there an explicit "no other text" line?
- Is
sizea valid pixel string, with both edges multiples of 16? - Is
qualityset tomediumorhighif any text is small or dense? - Have you decided which surface you are sending it to, and whether that surface will rewrite your prompt?
Eight checks, about twenty seconds. They cost less than one re-run, and on text-heavy work they are the difference between a first-pass asset and a fourth-pass one.
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 5.0★ on the Chrome Web Store.
Create An Account