The practice of writing and systematically refining prompts to get reliable, repeatable results from a generative model — concrete language, deliberate structure, and one change at a time.
Despite the name, most effective prompt engineering is editorial rather than technical: concrete nouns over abstract adjectives, one idea per clause, the most important elements stated first, and ruthless deletion of words that do not change the image. 'Golden-hour sidewalk café, linen shirt, reading a menu' beats a paragraph of vibes.
Persona workflows add one structural rule: never describe identity. With a trained model the face is already fixed — appearance words in the prompt ('young woman, brown eyes...') at best do nothing and at worst fight the trained identity and cause drift. The prompt's territory is the situation: action, setting, wardrobe, light, mood.
The common failure modes are keyword soup (twenty comma-separated adjectives pulling in different directions), internal contradictions ('candid documentary style, polished studio lighting'), and fighting the preset — prompting a beach scene into a studio-portrait Forge Style. Refine like an experiment: change one thing per iteration, reuse the seed, and keep what works. The first photoshoot guide shows the workflow end to end.
Related terms
The text instruction describing what a generation should depict — scene, wardrobe, mood, composition. With a trained persona, prompts describe the situation only; the identity comes from the model.
A secondary prompt listing what a generation should avoid — unwanted objects, styles or artifacts. The model steers away from the listed concepts instead of toward them.
Generating an image purely from a written description. The model synthesizes composition, subject, lighting and style from the prompt alone, with no input image.