PROMPT CRAFT

Five prompts, written as objects.

A macro portrait, a product, an interior, a second shot of the same character, and a short clip. Each one complete, each one with the single line to change first.

THE SHORT VERSION

Five finished objects you can paste today, and the one rule about key names that decides whether they work.

•  Most models read a JSON prompt as text rather than parsing it. There the key names are yours to invent, and structure helps because it stops you forgetting things, not because the model knows the schema.

•  Ideogram 4.0 is the exception. It genuinely parses a documented caption, so there the keys are fixed and anything else is dropped. Yesterday's issue covered that one.

•  Every prompt below is complete. Paste it, change the one line named underneath, run it again.

•  Honest caveat: these are starting points I use, not settings that guarantee a result. The value is that every run after the first changes one thing instead of everything.

People keep asking for the prompts themselves rather than the method, so here are five I actually reuse, written out in full. They are ordinary objects: a title, what it is for, the subject, the setting, the camera, the look, the output.

One thing to settle first, because it decides how much the key names matter. Nearly every model reads JSON as text. It is not parsing a schema, it is reading a tidy document, so you can call a key whatever you like and it still lands. Structure earns its place by stopping you from forgetting the light or the crop, not by unlocking anything.

Ideogram 4.0 is the one exception worth knowing, because it does parse a documented caption and ignores keys outside it. If you are working there, use the shape from yesterday. Everywhere else, use these.

For when you want texture rather than a face: pores, moisture, a single held expression. The detail lives in the subject block, and the camera block is what stops it looking like a phone photo.

{
 "title": "Ultra-realistic close-up portrait with water droplets",
 "intent": "Extreme macro emphasising skin texture, moisture and emotion",
 "subject": {
  "focus_area": "eye, nose and lips",
  "expression": "intense, contemplative, held",
  "skin_texture": "visible pores, fine hair, sitting droplets",
  "eyes": "greenish blue, reflective highlight from the key light",
  "hair": "dark, wet strands falling across the cheek",
  "lips": "slightly parted, wet, no gloss"
 },
 "environment": {
  "setting": "dark, close, no visible room",
  "lighting": "one soft directional source, high contrast on wet skin",
  "background": "out of focus, neutral, one stop down"
 },
 "camera": {
  "lens": "100mm macro",
  "aperture": "f/4",
  "focus": "the near eye and the droplets on the lash line",
  "white_balance": "neutral daylight"
 },
 "style": {"look": "hyperreal, cinematic", "post": "sharp micro detail, controlled speculars"},
 "output": {"aspect_ratio": "9:16", "quality": "maximum"}
}

Change first: the aperture. Going from f/4 to f/2 collapses everything but the lash line, and that single move decides whether this reads as a portrait or as a study of water.

The one I reuse most. It works because it names the surface and the reflection separately, which is where product shots usually fall apart.

{
 "title": "Single object, editorial still life",
 "intent": "Show the object honestly, no styling clutter",
 "subject": {
  "object": "an amber glass bottle, unlabelled",
  "material": "thick glass, matte cap, no branding",
  "condition": "clean, one fingerprint left deliberately"
 },
 "surface": {"material": "raw plaster", "colour": "warm off white", "texture": "slightly uneven"},
 "environment": {
  "lighting": "one large soft source from camera left, no fill",
  "shadow": "long, soft edged, falling right",
  "reflection": "a single specular strip down the left of the glass"
 },
 "camera": {"lens": "85mm", "aperture": "f/5.6", "height": "just above the object", "focus": "the front edge of the label area"},
 "style": {"look": "editorial still life", "colour": "muted, no saturation boost"},
 "output": {"aspect_ratio": "4:5", "quality": "maximum"}
}

Change first: the shadow direction. Moving it from right to toward camera changes the whole mood without touching the object, and it is the cheapest way to get three usable variants from one prompt.

Interiors come apart when the light is described as a mood rather than a source. Name the window, the time and what it lands on, and the room builds itself around that.

{
 "title": "Quiet interior, late afternoon",
 "intent": "A lived-in room, not a showroom",
 "space": {
  "room": "a small kitchen, plaster walls, wooden counter",
  "contents": ["a kettle", "two mugs, one used", "a folded cloth"],
  "walls": "empty, one nail hole"
 },
 "light": {
  "source": "one west window, out of frame left",
  "time": "an hour before sunset",
  "behaviour": "a long bar of light across the counter, everything else falling away"
 },
 "camera": {"lens": "35mm", "aperture": "f/4", "height": "standing eye level", "framing": "counter across the lower third"},
 "style": {"look": "documentary, unstaged", "colour": "warm highlights, cool shadows, no teal"},
 "output": {"aspect_ratio": "3:2", "quality": "maximum"}
}

Change first: the time. Move it to mid morning and the same room becomes a different photograph, because the light behaviour line does the work rather than the adjectives.

This is the one that saves a series. Everything under constant is copied unchanged from the first shot, and only the variable block moves. Keep the constant block in a file and paste it every time.

{
 "title": "Second shot, same character",
 "intent": "A different frame that still reads as the same series",
 "constant": {
  "person": "woman, late thirties, short dark hair pushed back, no makeup",
  "wardrobe": "grey cotton shirt, sleeves rolled twice",
  "light": "one window left, no fill, overcast",
  "palette": ["warm grey", "off white", "one muted green"],
  "lens": "50mm at f/2.8"
 },
 "variable": {
  "framing": "waist up, turned three quarters away",
  "action": "reaching for something outside the frame",
  "expression": "unaware of the camera"
 },
 "output": {"aspect_ratio": "4:5", "quality": "maximum"}
}

Change first: nothing in the constant block, ever, or the series stops matching. Every new shot edits only the three lines under variable.

Video prompts fail by asking for two things at once. One camera move, one thing happening, and a first and last frame that are both describable.

{
 "title": "Six second clip, single move",
 "intent": "One motion, no cut, usable as an insert",
 "scene": {"place": "the same kitchen", "time": "late afternoon", "sound": "none"},
 "first_frame": "the counter empty, light across it",
 "action": "a hand enters from the right and sets down a mug, then leaves",
 "last_frame": "the mug alone on the counter, light unchanged",
 "camera": {"move": "slow push in, 10 percent", "lens": "35mm", "height": "counter level"},
 "constraints": ["no cut", "no second subject", "no camera shake"],
 "output": {"duration_seconds": 6, "aspect_ratio": "16:9", "quality": "maximum"}
}

Change first: the constraints list. It is doing more work than the action line, because most bad clips come from the model adding a cut or a second person you never asked for.

Reading models. Most of what you use reads the object as text, so name keys for yourself. Call it look or style or vibe, it makes no difference, and a key you understand a month later is worth more than a clever one.

Parsing models. Ideogram 4.0 reads a documented caption and drops the rest, so there the names are not yours to choose. That is the exception, not the rule, and it is the only place where an invented key silently costs you something.

Paste this with a picture you already like, and it returns an object in the same shape as the five above.

Below is a description of a picture I want to be able to make repeatedly. Write it as one JSON object with a title, an intent, a subject block, an environment block naming the light source and its direction, a camera block with lens, aperture, height and focus point, a style block, and an output block with aspect ratio. Use only detail I gave you. Where I was vague, do not guess: put that item in a notes list outside the object and say what you would need from me. Then name the single line I should change first to get a useful second version.

The notes list is the whole point. A model that fills your gaps quietly hands you its taste, and you will not notice until the third run.

Paste the second prompt, swap the bottle for something on your desk, and run it three times changing only the shadow direction. Three usable variants in four minutes is the whole argument for writing prompts this way.

THE NEXT WORKING NOTE

The format sheet is nearly done: one prompt written four ways side by side, a sentence, a line of tags, a set of blocks kept constant, and the object, with a note on which held under repetition.

Want it when it ships? Reply with send me the format sheet and I will get it to you.

A QUESTION FOR YOU

Which of the five did you actually run?

Reply with the number and what you changed. I will publish the sharpest variants next to the originals, credited if you want.

If this was useful, forward it to someone who retypes the same prompt every week.

MAKE SURE THIS REACHES YOU

Mailbox providers decide what counts as wanted mail, and they learn from what you do. If this landed anywhere other than your main inbox, three things fix it permanently.

Reply to this email. One word is enough. A reply is the clearest signal you can send that this belongs in your inbox, and it takes three seconds.

Move it to your main tab. If it arrived under Promotions or Updates, drag it across. On a phone, open the menu at the top right and choose Move.

Add the sender to your contacts. Tap the name at the top of this email and choose Add to contacts. That keeps it out of the spam folder for good.

WHAT THIS IS

Luxe Prompting is a working journal about prompting and generative work, written for the people making images, video and music with these tools rather than writing about them. One issue at a time, no announcements, no hype, and nothing I have not run myself first.

Partners cover the cost of writing it. I only take the ones my own readers would plausibly use, and I say so plainly at the top rather than dressing it as an issue. If you ever want to know how something here was checked, reply and ask. I keep the sources for every issue, and I would rather show them than be believed.