How to Prompt AI Design Tools: Prompt-to-Outcome Tutorial
Most AI design results disappoint for a plain reason: the prompt described a mood and the tool guessed everything else. Every open slot - subject, style, composition, light, color, format, limits - gets filled with averages. The output looks generic because it is an average.
This tutorial teaches prompting job by job, each prompt shown in full with the outcome next to it. Every example starts with a weak prompt and what goes wrong, then gives an improved prompt to copy and adapt, then explains why each element changed the result. These are worked examples: they describe the result a prompt is written to produce, across most image tools. Settings differ between tools, and one slot shows it most: aspect ratio, which may be a setting, size words, or a parameter. Check your tool's controls for that slot; keep the rest the same.
The Seven-Slot Prompt Formula
Every improved prompt here fills the same seven slots. Fill all seven and the tool has few guesses left - and fewer guesses means fewer surprises.
| Slot | What goes in it | Why it changes the result |
|---|---|---|
| 1. Job and placement | What the image is for and where it runs | Sets detail level, empty space, and finish |
| 2. Subject | The one main thing, in concrete nouns with physical details | Stops the tool picking a subject for you |
| 3. Style and medium | The visual language: flat vector, watercolor, photo, 3D render | Steers texture, edges, and detail density |
| 4. Composition | Viewpoint, framing, where the subject sits, what stays empty | Decides whether the image fits your layout |
| 5. Light and color | Lighting type and direction, plus a named palette | Prevents flat default light and random colors |
| 6. Format | Aspect ratio and orientation: 16:9, 1:1, vertical | A wrong ratio crops your subject |
| 7. Constraints | What must not appear: text, watermark, extra objects | Removes common defects before they happen |
Write the slots as short comma-separated phrases, not a paragraph of wishes: they are easier to edit one at a time, which is the whole iteration method here. Keep a copy of every prompt.
Job 1: A Logo Concept for a Coffee Brand
The job. A small coffee roaster needs logo concepts to react to: not finished files - a working logo needs redrawing as vector paths - but strong directions to refine.
The weak prompt.
PROMPT: a logo for a coffee company
What goes wrong: a busy picture, usually a steaming cup and scattered beans, plus a made-up word in a decorative font, often misspelled. Nothing survives shrinking to favicon size.
The improved prompt.
PROMPT: Logo concept for a small-batch coffee roaster. Subject: a geometric fox head built from two curved shapes, one ear shaped like a coffee leaf. Style: flat vector, bold edges, no gradients, no shading. Composition: mark centered on a plain off-white background, empty space around it, nothing else in frame. Color: two colors only, deep forest green and warm cream. Format: square. Constraints: no text, no letters, no coffee cup, no beans, mark must read at small size.
The outcome, and why. A single flat mark: a fox head readable at thumbnail size, two flat colors, no lettering to garble. The job slot pushed the tool from scenes toward a symbol; the subject slot swapped the predictable cup for an animal with one twist, the leaf ear. Flat style removed the shading that makes AI logos muddy. An empty frame lets the mark be lifted and traced; two colors allow a one-color print test. The text ban matters most: lettering is unreliable, so the designer sets the wordmark later.
Iteration step: run it three times, keep the strongest silhouette, then change one slot - fox to heron, or green to rust red.
Job 2: A Landing Page Hero Image
The job. A project management app for construction firms needs a hero image: headline and signup button on the left, picture on the right. Palette and image rules: see the brand style guide.
The weak prompt.
PROMPT: happy construction team using software, modern, professional
What goes wrong: a staged group photo, five people smiling, harsh light, a frame filled edge to edge. The headline has nowhere to sit, and it could belong to any software company.
The improved prompt.
PROMPT: Landing page hero image for a construction project management app. Subject: one site manager in a hi-vis vest and white hard hat holding a tablet, seen from behind at a three-quarter angle, looking across a building site at golden hour. Style: realistic photograph, shallow depth of field. Composition: subject in the right third; left half soft blurred warm background, empty for headline text. Light and color: warm late-afternoon sun from the left; palette of warm sand, slate gray, safety orange. Format: wide 16:9. Constraints: no readable text on screen, no logos, no extra people, no watermark.
The outcome, and why. A wide image with one figure on the right and a calm blurred left side where the headline sits. Because the job slot named the placement, composition could request empty space on a specific side: soft background. One person instead of a team removes most hand and face errors, and the back three-quarter angle makes facial defects stop mattering. Named light replaced the flat default; the screen constraint blocks garbled text.
Iteration step: if the empty side returns busy, strengthen the composition phrase only and touch nothing else.
Job 3: A Social Media Post Graphic
The job. A bakery announces a Saturday sourdough class as a square feed post. Date and price are added as real text afterward, as in Job 1.
The weak prompt.
PROMPT: sourdough bread baking class poster, delicious, cozy
What goes wrong: a finished poster with invented text in a wobbly font - a date along the lines of "Satuday Octobr" - and bread under yellow light that makes it look stale. Baked-in text cannot be corrected, only regenerated.
The improved prompt.
PROMPT: Social media graphic background for a bakery announcing a sourdough class. Subject: one rustic sourdough loaf with a scored crust, on a linen cloth beside a small flour bowl, photographed from overhead. Style: editorial food photography, crisp crust detail. Composition: loaf in the lower right; upper left third plain warm cream background, empty for text added later. Light and color: soft window light from the upper left; palette of warm cream, crust brown, muted sage. Format: square 1:1. Constraints: no text, no letters, no hands, no people, no watermark.
The outcome, and why. A clean overhead shot with one loaf and a calm empty corner. The deciding move sits in the job and constraint slots: this is a background for text, not a finished poster. The overhead angle reads well at feed size, and the reserved empty third gives the date and price a quiet field. Stated window light gives the crust its texture.
Iteration step: hold the prompt fixed, generate a small set, and pick on crust detail alone. Texture varies more between runs than between prompt edits.
Job 4: An Icon Set That Matches
The job. A budgeting app needs eight small icons - wallet, chart, calendar, bell, shield, card, coin, gear - that look like one family. Sets are hard: each icon is generated separately, and the tool remembers nothing unless you repeat the same words.
The weak prompt.
PROMPT: a wallet icon
...repeated eight times, once per object. What goes wrong: eight unrelated pictures - a 3D render, a thin line icon, a colored cartoon, one with a background, one without. A mismatched set looks broken even when each icon is fine alone.
The improved prompt. Write one style block and reuse it word for word across the set. Only the subject slot changes.
PROMPT: Icon for a budgeting app. Subject: a wallet with one card sliding out. Style block, identical for every icon in the set: flat vector icon, bold outline in near-black, one accent fill of muted teal, rounded corners, thick even stroke weight, no gradients, no shadows. Composition: icon centered on a plain white background, filling about 70 percent of the frame. Format: square 1:1. Constraints: no text, no letters, no second object.
The other prompts are identical, subject swapped: a bar chart with three bars, a calendar page with two rings, and so on.
The outcome, and why. Icons that read as siblings: same outline weight, same accent, same frame fill, same background. The style block fixes every variable except the subject. Naming the stroke and corners stops drift between sharp and soft icons. One fixed accent stops each generation picking its own palette, the most common reason sets clash. The stated 70 percent frame fill keeps icons the same size, and the white background makes cutting out easy. For interface use, see the accessible color systems guide on testing that accent for contrast.
Iteration step: fix the style block on the wallet first, then run the rest with no style edits.
Job 5: A Product Mockup Scene
The job. A candle maker needs a lead image per listing: the candle on a bathroom shelf, styled but believable.
The weak prompt.
PROMPT: luxury candle product photo, beautiful bathroom, spa vibes
What goes wrong: the candle is one of nine objects on a marble counter, its label carries invented script, and the scene reads as a hotel ad. "Luxury" and "vibes" have no visual content, so the tool spends them on marble and gold.
The improved prompt.
PROMPT: Product photograph for an online candle listing. Subject: one amber glass candle jar, plain blank cream label, lit flame, on a light oak shelf. Style: realistic product photography. Composition: candle large in the center foreground; behind it, blurred, a folded towel and a eucalyptus stem in a vase; plain plaster wall. Light and color: soft morning window light from the right; palette of warm white, light oak, amber, muted green. Format: vertical 4:5. Constraints: label completely blank, no text, no logo, no people, no hands, no extra candles, no marble, no watermark.
The outcome, and why. One candle, large and centered, on a believable shelf, label blank for real art added later. Mood words were replaced with objects, materials, and light. The subject slot fixed the count at one, removing clutter and fake script. Composition built depth in a stated order - product sharp, props blurred, plain wall last - and the palette banned the marble-and-gold default. The vertical 4:5 matches listing displays. Licensing for scenes you did not generate: see the design asset licensing checklist.
Iteration step: lock the scene and vary only the props per scent - cedar block and pine sprig for cedar, halved orange and cinnamon for spice.
Job 6: A Brand Board and Palette Image
The job. A designer opening a brand project for a children's swim school wants one board image that sets palette and tone.
The weak prompt.
PROMPT: fun colorful moodboard for a kids swim school brand
What goes wrong: a collage of clashing blues, cartoon fish, and pool photos in three styles. It sets no palette, because it contains every color, and no tone.
The improved prompt.
PROMPT: Brand board image for a children's swim school identity. Subject: simple shapes suggesting water - layered paper-cut waves, one sun shape, one swim ring - laid out as color swatches and shape studies. Style: flat paper-cut illustration, clean cut edges. Composition: tidy grid, even spacing, each study in its own cell. Color: five colors only - deep pool blue, bright aqua, coral, warm sand, white. Format: wide 16:9. Constraints: no text, no characters, no faces, no photographs, no gradients.
The outcome, and why. A tidy grid of wave and sun studies in five stated colors, ready to sample. The job slot reframed a mood collage as a brand board; composition imposed the grid that separates a board from a pile. Five named colors is the point: a board containing every color decides nothing. Constraints keep photography and characters out, and paper-cut gives the brand a texture it can use in print.
The Iteration Workflow
Good prompting is mostly good editing. Run every job through three passes.
Pass 1 - Generate wide. Run the full prompt three or four times without editing it. Variation shows which parts the tool holds steady and which it re-rolls. Pick the strongest on composition and subject first; color and light can be fixed later, a broken composition cannot.
Pass 2 - Critique against the checklist. Score the result in order:
| Check | Question | Slot to fix if it fails |
|---|---|---|
| Subject | Subject as named, right count and scale? | 2. Subject |
| Style | Medium as asked, flat or photo? | 3. Style and medium |
| Layout | Empty space where needed? Bad crops? | 4. Composition |
| Light and color | Light has a direction? Colors in palette? | 5. Light and color |
| Format | Right ratio and size for the placement? | 6. Format |
| Defects | Any garbled text, extra fingers, stray objects, watermarks? | 7. Constraints |
Pass 3 - Refine one variable at a time. Change exactly one slot, regenerate, and compare with pass 2. Slower per step, faster overall: you learn which words cause which effects, and a three-slot change teaches nothing. Keep the log:
| Generation | Prompt change | Result | Keep or revert |
|---|---|---|---|
| 1 | Full prompt, first run | Good composition, colors saturated | Keep |
| 2 | Slot 5 only: "muted tones" | Colors in palette, composition held | Keep |
| 3 | Slot 7 only: "no people in background" | Clean background, edge artifact | Keep, crop edge |
Row 3 carries a lesson: cropping, artifact removal, and real text are editor jobs. Prompting is for decisions an editor cannot undo.
Failure Fixes: Five Common Defects
Short on purpose: the repairs to reach for during pass 3.
Garbled text in the image
Symptom: headlines and labels come out misspelled or as letter-like shapes. Fix: remove text from the generation; add it afterward in a design tool.
PROMPT: (add to any prompt) Constraints: no text, no letters, no words, no signage, leave a plain empty area where text will be placed later.
Outcome: a clean image with space reserved for type, spelled correctly because a person set it. If a job truly needs a word in the image, inspect every letter; treat success as luck.
Hands with wrong fingers
Symptom: extra, fused, or bending fingers, worst when hands hold objects. Fix: design hands out of the shot - subject from behind, from overhead, or cropped at the wrist.
PROMPT: (revise the subject line) Subject seen from behind at a three-quarter angle, hands out of frame. Constraints: no visible hands, no fingers.
Outcome: the same scene, defect removed by composition, not luck. Where a hand is the point, budget extra generations and finish in an editor.
Off-brand colors
Symptom: a well-made result in colors your brand does not use, often saturated blue and purple. Fix: name the palette in slot 5, with color names or hex values, and ban wrong colors in slot 7.
PROMPT: (revise slots 5 and 7) Color: palette limited to warm cream, rust red, and charcoal. Constraints: no blue, no purple, no neon colors, nothing saturated outside the palette.
Outcome: generations land inside the brand range. Paste the same palette line into every project prompt.
The generic stock look
Symptom: smiling models, perfect offices, floating 3D shapes, a scene that could sell anything. Fix: add specific, worn detail to the subject and setting slots, and ban the clichés by name.
PROMPT: (revise slots 2 and 7) Subject: a working pottery studio bench, clay dust on the surface, tools laid out mid-task, one chipped mug. Constraints: no smiling at camera, no symmetry, no floating objects, no lens flare.
Outcome: a scene with a point of view. Dust, wear, and a task in progress read as real because stock imagery removes exactly those things.
Wrong ratio or format
Symptom: a vertical subject cropped by a square frame, or cut at the edges. Fix: set the format slot before generating, in whatever form your tool accepts - a ratio setting, size words, or a parameter - and match the framing to it.
PROMPT: (revise slots 4 and 6) Composition: full subject inside the frame with margin on all sides, nothing touching the edges. Format: vertical 4:5, tall orientation.
Outcome: the whole subject survives, with margin to crop. Margin is cheap at generation time, impossible afterward.
A Short Note on Commercial Use
Prompting skill and usage rights are separate questions; this tutorial teaches the first. A perfectly prompted image can still carry terms that limit its use; rules differ by tool and plan. Before a generated image goes into client work or an ad, read the rights side in the companion guide, AI-generated images and commercial use, which covers platform terms, copyright, and disclosure. Two habits help there too: keep your prompt log, and keep logos redrawn by a designer (Job 1).
FAQ
Why does the same prompt give a different image every time?
Generation includes randomness, so each run re-rolls the details your prompt left open. The parts that change between runs are the parts the prompt did not pin down. Add specificity to the slot that owns the changing element, or pick the best of several runs.
Should I write prompts as sentences or as comma-separated phrases?
Comma-separated phrases suit design jobs: each phrase maps to one slot, so editing one variable means editing one phrase, and a style block copies exactly. Sentences tangle slots together and are hard to debug.
Can AI image tools make a finished logo?
They make strong concepts and weak finished logos. A finished logo needs clean vector paths, exact letterforms, and controlled spacing; image tools output raster pictures with unreliable text. Explore with the Job 1 method, then have the winner redrawn as vector artwork.
How do I keep a set of images in one style?
Reuse one style block word for word across the set, as Job 4 shows, and paste one palette line into slot 5 project-wide. Change only the subject slot. A reference image helps if your tool offers one, but the shared words carry most of the consistency.
What is a negative prompt, and where does it go?
It is the constraint slot: things the image must not contain, such as no text, no watermark, no extra people. Some tools give constraints their own field; others take them in the main prompt. Spend constraints on defects you have actually seen; a short, specific list beats a long copied one.
What to Read Next
- AI-generated images and commercial use - platform terms, copyright, and disclosure.
- Design asset licensing checklist - the pre-shipment review for every asset type.
- How to create a brand style guide - where to record palettes and style blocks.
- Accessible color systems - testing prompt palettes for contrast.
- AI design prompting tutorials hub - the pillar page for this series.
Maintained by the Lucky Graphics editorial team. Prompts are shown in full so you can copy them; outcomes describe what each prompt is written to produce.