Image generation tips: how I get the picture I actually asked for

I make every image on this blog with Google’s Nano Banana models. When an image misses, it’s almost never the model. It’s what my prompt left out. These are the tips that made the biggest difference, three side-by-side examples, and the free skill I now use so I don’t have to remember all of it.

A clean pencil drawing of a cherry tree with blossoms bound to its branches, standing beside a venue doorway, with small figures for scale.
02BThe keeperFrom the second example below

The model fills every gap with a guess

Most of us write image prompts the way we’d brief a friend: one line, and the rest is obvious. To the model, nothing is obvious. Anything you don’t say, it guesses: the colours, the layout, what the picture is about, where the words go. Sometimes the guess is lovely. It’s rarely what you had in mind.

So the tips are mostly about closing those gaps.

Seven tips that changed my results

  1. Say what’s in the picture, what it’s doing, and where.

    Plain nouns for the subject, a verb for the action, and a setting. Then stop: anything you don’t mention, the model may invent, and anything extra you mention will show up.

  2. Name your colours, and say where each one goes.

    “Warm colours” gets you whatever the model thinks is warm. Named colours, ideally with hex codes, get you your palette. Say which colour is the background and which is the accent.

  3. Plan the space before the picture.

    Give the aspect ratio and say where you need empty space. If the image is a blog cover, say the top third stays clear for the title. Otherwise the model puts something there.

  4. Put exact words in double quotes, and say where they go.

    “The words "House rules" in hand-lettered capitals, bottom left.” Models are much better at text than they were, but only when you spell it out.

  5. Describe what you want, not what you don’t.

    A long list of “no” rarely works. Describe the thing you want instead, and keep any “don’t” to one short sentence at the end.

  6. Watch your genre words.

    Some words carry baggage. Sketch, concept, proposal, diagram and blueprint all invite arrows and labels, because that’s what those images usually have. If you want none, ask for a finished illustration.

  7. Say what it’s for.

    A cover, a slide, a client proposal. The model leaves room for what comes next, and picks a style that fits.

Three examples

A blog cover

The brief: a cute picture-book cover for a post about giving AI a page of “house rules”, cream background, forest green and tomato red, the words “House rules”. Same model, same size, first result each time.

One-line prompt result: a busy picture-book scene with a robot reading a house rules sheet, model houses, house posters, and a large 'House rules' title across the top.
Full prompt result: a small character in red dungarees pins a 'House rules' sheet with ticked lines to a cream wall, next to a blank easel, with empty space across the top.

01AOne lineScored 3 of my 6 checks

Prompt32 words

Cover image for my blog post about giving AI a page of house rules before it designs. Cute picture-book style, cream background, forest green and tomato red, with the words "House rules".

01BFull promptScored 5 or 6 of my 6 checks

Prompt180 words

A cute children's picture-book illustration on a plain cream background (#F6F1E7). A tiny, round designer character with simple dot eyes, wearing a tomato red (#E4572E) jumper, stands on a little step stool and presses the last pin into a single large sheet of paper on the wall. Render exactly one line of text at the top of the sheet, spelled exactly "House rules", in a bold rounded hand-lettered sans-serif in forest green (#1F4D3A). Below the title, the sheet has five short ruled lines, each with a small tick mark, and no other readable words. To the right stands an easel holding a blank artboard, with a small pot of pencils at its foot. 16:9 wide composition: the sheet and character left of centre, the easel right of centre, and generous empty cream space across the top third. Thick rounded forest green outlines, chunky soft shapes, flat colour, no gradients, flat even light with small soft shadows. Tomato red only on the jumper and the pins. No other text, no logos, no robots, no glowing AI effects, no real product interfaces.

The one-line prompt is charming. It also took “house” literally (little houses everywhere), added a robot, brought in grey, blue and brown, and put the lettering right where my title needs to go.

The full prompt, which covers tips 1 to 4 and 7, is plainer, and it’s what I asked for: rules pinned to the wall, a blank easel waiting, only the colours I named, and empty space across the top. It wasn’t perfect: dungarees instead of a jumper, four ticked lines instead of five. Against my six checks it scored five or six. The one-liner scored three.

A client sketch

The brief: a sketch of a cherry blossom arrangement for a wedding proposal. This time I asked an AI assistant to write the prompt for me, once on its own and once with the skill below.

Assistant's own prompt: a photo of a sketch on a desk, with pencils, a ruler, a coffee cup and handwritten labels, some of them repeated.
Prompt written with the skill: a clean drawing of a cherry tree with blossoms bound to the branches, beside the venue doorway, with people for scale.

02AThe assistant on its ownA photo of a sketch, with props and repeated labels

02BThe assistant with the skillA clean drawing I’d show a client

On its own, it wrote a decent paragraph and got a photo of a sketch on a desk, with pencils, a coffee cup, labels everywhere and two of them written twice. It didn’t know which details this model needs. With the skill, I got a clean drawing I’d show a client: one controlled colour, how the flowers attach to the branches, the venue doorway, and people for scale.

And tip 6 came from this same job. When I asked for a final version as “a florist’s concept sketch” with “no text” at the end, it came back with arrows and handwritten labels anyway. “Concept sketch” outweighed my one-line ban. Asking for “a clean, unannotated watercolour presentation illustration” fixed it.

A sketch, then a photo

Style is one line of the prompt, and it changes the whole picture. Here is one idea twice: a pyramid of bowling pins dressed as a Christmas tree, first as a sketch for the proposal, then as a photo-real version standing in the space.

Pencil sketch of a Christmas tree built from rows of bowling pins on a stepped wooden frame, with berries and leaves, and a family drawn beside it for scale.
Photo-real version of the same bowling-pin Christmas tree, dressed with greenery and cotton, standing in a busy shopping arcade while a small child looks up at it.

03ASketchFor the proposal: structure, scale, materials

03BPhoto-realThe same tree, as it could look in place

The checklist I use before I send a prompt

  1. SubjectWhat’s in the picture
  2. ActionWhat it’s doing, or how it’s posed
  3. SettingWhere it is, and what else is there (and nothing more)
  4. FramingAspect ratio, where things sit, where the empty space goes
  5. StyleIllustration, photo, watercolour; line and texture
  6. ColoursNamed, with hex codes, and where each one goes
  7. LightSoft daylight, golden hour, flat studio light
  8. WordsExact text in double quotes, lettering style, position
  9. UseCover, slide, proposal

The skill I use so I don’t have to remember all this

Writing all of that for every image is a lot of careful work, so I wrote a Claude skill that does it with me. It’s called Image Prompt Engineer, and it’s free. It started as a Nano Banana helper; the public version also writes prompts for OpenAI’s GPT Image, the model inside ChatGPT.

The prompt comes in two parts: a short plain-language brief on top (the scene, the mood, who stands where, what must not go wrong), and the JSON detail below (appearance, exact text, materials, camera, light). Prose alone got the relationships right but lost detail; JSON alone kept the detail but misread positions. For simple images, you can ask for prose only. Either way, every decision is visible, so nothing gets left out.

If you have a Gemini or OpenAI API key, there’s an optional script that lets Claude Code generate the image, check it against the spec and fix only what’s wrong. You don’t need it: copy and paste is the default, and no setup is needed.

Get it

Free · MIT licence github.com/stevensmith-design/image-prompt-engineer

In Claude Code, clone the repo and copy the skill folder into your skills. In the Claude apps, download the .skill file from the latest release and upload it as a skill. It covers Nano Banana 2.1, Pro, 2 and 2 Lite, and GPT Image 2.5 Sunburst, 2.5 Flare and 2, checked against Google’s and OpenAI’s docs on 26 October 2026.

Read nextDesigned, not generated — the same idea for whole pages.