How do you prompt AI for costume design when you only have a vague creative vision?

Hand the vague vision to a costume designer sub-agent as a mood ('feral, monastic, modern decay'), ask for four options per character, then escalate the winner through a fixed descriptor stack — silhouette + era, fabric + texture, construction detail, lighting + finish, plus a negative list — and lock that stack as a verbatim block reused on every shot.

Start by giving the costume problem to a dedicated costume designer sub-agent inside the invideo agent (invideo is an agentic video tool where you spin up named crew sub-agents that share the project's context). Don't try to write a perfect description — give the agent the feel. One documented production did exactly this for a vampire character: "I did not have a clear description of the sort of costume for Sylvia, who is our female vampire. But I always knew the sort of feel I want from her costume. So agent 1 was able to give me multiple options in the same zoom," says Hridaye, invideo's creative director. Mood-in, options-out is the whole unlock when specs are vague.

Ask for four options per character in a single round, not one image at a time. Documented productions generate four variations of every key asset (character sheets, costumes, environments) and pick the strongest before any video runs — a pattern that scales: 11 reference images covered 4 characters and 1 prop across a 3-minute episode at roughly $9.78 per locked character. Image generation is cheap; optionality is the point.

Once one option clicks, escalate it into a concrete descriptor stack in this order: (1) silhouette + era ('long Victorian-cut coat, floor-length'), (2) fabric + texture ('heavy wool, worn leather trim, hand-stitched'), (3) construction detail ('asymmetric closure, raw hem, no visible zippers'), (4) lighting + finish ('matte surfaces, candlelit warm falloff'), (5) negative list ('no synthetic sheen, no modern hardware, no plastic buttons'). This subject → silhouette → material → construction → lighting → mood formula is the structure most working AI fashion prompts converge on, and it gives the model the physical anchors it needs to render drape and weight instead of guessing.

If a returned option feels unexpectedly bold, keep it. The working heuristic from documented productions: "If you feel like it's too off, then it means we should lock it in." Surprise is signal, not noise — vague briefs often produce the most interesting answers on the first pass.

Lock the winner as a verbatim descriptor block and reuse it word-for-word on every shot. This is how character consistency holds without fine-tuning: in a 70-second short film with 2 characters, the same person appeared in every scene — no LoRA — because the costume descriptor lived inside the invideo agent's persistent context and was attached to every generation. Pair the locked text block with a 4-angle character sheet (front, side, profile, back) generated at high resolution, and the costume becomes a reproducible asset rather than a re-rolled guess.

For close-ups, build a separate descriptor for the construction details the wide won't reveal — stitching, buttons, fabric weave, fastenings — and store it as its own anchor. Continuous sequences where the costume evolves (a coat picks up dust, a character adds a trinket city by city) need a fresh descriptor and sheet per beat, not one master sheet stretched across the arc.

When continuity breaks later, fix the descriptor block and the sheet, not the shot. Ask the invideo agent to inspect the character sheet for the error; it will identify the exact panel, correct it, store the updated version in context, and inherit the fix into every subsequent generation — surgical, not slot-machine re-rolls.

Watch some of these to see what works for you:

I did not have a clear description of the sort of costume for Sylvia, who is our female vampire. But I always knew the sort of feel I want from her costume. So agent 1 was able to give me multiple options in the same zoom.

— Hridaye, invideo's creative director