A good prompt tells the model exactly what to do and what to hand back. The prompt tools handle the structure for each target; your part is the substance. These habits get you a prompt that works on the first try.

Give the tools a useful idea

The prompt tools can only include what you tell them. A useful idea answers three questions:

  1. What is the job? "Write a cold email", "a product shot of a perfume bottle", "a drone shot over a coastline".
  2. Who or what is it for? The reader, the platform, the brand or the use: "for a podcast host", "for an Instagram ad", "for a beginner".
  3. What must it include or avoid? Length, facts, words, colors, things to leave out: "under 120 words", "the bottle is frosted glass", "no people".

"Email to a podcast host" gives a generic prompt. "Cold email to a podcast host about being a guest, friendly, under 120 words, mention one recent episode" gives a prompt you can send.

Where the tools need a fact you did not give, they leave a placeholder in square brackets, such as [PRODUCT NAME]. Fill those in before you paste the prompt.

Chat prompts

For ChatGPT, Claude, Gemini, DeepSeek and Grok, a strong prompt has:

  • the task in one sentence,
  • the context the assistant cannot guess: your situation, your audience, the material to work from,
  • the requirements: tone, length, what to include and what to avoid,
  • the output format: a list, a table, an email with a subject line, code only.

Paste the material the assistant should use (notes, a draft, an error message) into the target tool after the prompt, not into the idea box, unless the prompt itself needs it. Use Add an example when the format matters, and Try for when you switch assistants.

Image prompts

Describe the picture the way you would brief a photographer or illustrator:

  1. Subject: who or what, with the details that matter. "A ceramic ramen bowl with a soft egg and spring onions", not "food".
  2. Setting: where it is and what is around it.
  3. Light and time: "overcast morning light", "neon signs at night".
  4. Style or medium: photo, watercolor, 3D render, flat illustration.
  5. Framing: close-up, overhead, wide shot from a low angle.

Keep it to one scene. For words inside the image, put the exact text in quotes and choose a target that handles lettering well, such as GPT Image. Choose the aspect ratio and style in Options rather than typing them in.

Video prompts

A video prompt needs what moves and how the camera moves:

  • Action: "the cat pounces at the laser dot and slides on the floorboards".
  • Camera: pick a move in Options (static, dolly in, orbit, handheld, drone), or describe one.
  • Sound, for targets that make audio, such as Veo: "rain on the window, a kettle starting to whistle".

Shorter clips are easier to get right. In the studio, video is priced per second, so test a short clip before you make a long one.

Editing a photo in the studio

When you upload a photo, say what changes and what stays: "Replace the background with a beach at sunset. Keep the person, clothes and pose the same." A short, specific instruction beats a full scene description.

Leave these out

  • Settings that have their own control: aspect ratio, duration, number of images.
  • Long strings of quality words such as "masterpiece, 8k, ultra detailed". Current models gain little from them.
  • Names of real people you do not have permission to depict, and anything the content rules do not allow.

When the prompt misses

Change one thing at a time. If the prompt is too long, press Shorter; if it is vague, add the missing fact to your idea and generate again. If an image or video result is close but not right, change one detail in the prompt and run it again, so you can tell what helped.