How to Deliver an AI Project a Client Will Pay For (Pre-Production First)

Professional AI work for clients is not about better prompts. It is about better systems. The creators who get paid again are the ones a client can trust with a brief, and trust comes from pre production, not from the model.

This is the pre production workflow we use on client projects, commercials and narrative work. It takes you from scattered ideas to structured, cinematic scenes with the same character, the same location and the same lighting across every shot. I built a full board on camera so you can follow it.

One place for everything

We run pre production in Milanote. It is an infinite canvas where images, tables, notes, links, scripts, video and music all sit together. Columns stack your images, tables hold your prompts and descriptions, comments let a client leave notes on a specific frame, and colour tags mark a column green for signed off or red for needs changing. A YouTube reference and a Spotify track are both playable inside the board. A member's ad that hit 4.6 million views lives on one of ours as inspiration.

If you are a solo creator this keeps you honest. If you run a team it is the difference between a project and a mess. I use it for our social videos and our Meta ads too.

Step one: cast the character

Every Hollywood show holds auditions. You should too. A character carries a tremendous amount of weight in a story or a commercial, and the first face the model gives you is rarely the right one.

Build a table with three columns: character one, two and three. Rows for description, age and key emotional features. For the description, find a face on Pinterest or Cosmos that is close to what is in your head, upload it to ChatGPT with a paragraph of what you want, and ask for a prompt that describes the character in detail as a head and shoulders shot on a white studio background with soft light. Ask it to summarise the features into six bullets. That table is your casting sheet.

Generate all three. Then look at them next to each other. On our board the first woman read as confident and self possessed. The second looked spiritual and gaunt. The third looked like she had stories. I would rather sit down to dinner with her, so she got the part. That comparison is the point of the exercise, and you cannot do it with one character.

The outfit matters as much as the face. My first full body shot came back looking like a business executive when I wanted a Russian grandmother, so I re prompted for a knee length floral skirt and black tights, then used the full body image to replace the collared shirt in every close up. Get the clothing right once, before you generate anything else.

Build the character sheet

From the hero image, generate the angles: side profile, back of the head, a close up of the eyes, a full body shot for height and outfit, and the hands. Ours wears a wedding ring and an engagement ring, and that detail survived into every scene later because it was on the sheet.

Then collage those images into one file in Photoshop. Hero portrait big on one side, the profiles and details around it, eyes blown up so the face reads. Export it. That one image is what you upload with every generation from now on, and the model pulls whichever angle it needs. Duplicate the column with Command D and repeat for the next character.

A note on models as of September 2026: I use GPT Image 2 for the faces because the skin texture holds, and Nano Banana Pro for locations, lighting and dropping the character into a scene.

Step two: scout the location before you commit

Same table, three locations. Find references on Cosmos, upload one to ChatGPT and ask for a prompt that recreates the scene as a Soviet era apartment full of period artefacts with the same lighting as the reference. Generate all three. Tell it no people in the shots.

Then compare them for one thing: what can a character do here? Our first apartment was empty and told me nothing. The second had drawings and a telephone. The third had the cupboards open, jars of pickles on the table, pots and pans on display, a bare bulb on a wire and a stereo in the corner. That room says these are open people with nothing to hide and cheap easy living. It is the one with a story in it, so it won.

I generate everything at 2K. Some video models will not take 4K anyway, and 2K is quicker.

Turn one hero shot into a whole set

Upload the winning room to ChatGPT and ask for seven prompts that shoot it from different angles: top down on the stove, close up of the pickles, the cupboards on the left, a low angle from under the radio, the view from outside looking in through the window. Then run each prompt in Nano Banana Pro with the hero image in the prompt bar.

Because the hero shot is always in the bar, every angle inherits its lighting and its detail. You end up with a location that behaves like a real set, where you can put the camera anywhere.

Step three: lock the lighting

If you are telling a longer story in one location, you need to know it works at 6am, 2pm and 11pm. So make a lighting table and change the time of day on the hero shot until you get one you love. Ours was a hazy sunrise.

Then the trick that saves hours: upload the new lighting to ChatGPT and ask it to extract everything about the light, shadows and time of day into one reusable prompt that starts with "adjust the lighting in this image to". Paste that same prompt onto every other angle of the room. Same scene, same light, all seven shots. Repeat for night. Check the small things. My first night version lost the hanging bulb, and a client will notice.

Step four: put the character in the room

Upload a room angle plus the character sheet. The sheet never changes. Then say plainly what you want: she is sat in the chair facing camera, a cigarette in one hand, the other on her lap, she looks tired. Or an over the shoulder shot with her out of focus walking away from us. Or the camera focused on the pickles with her soft behind them.

Run it four times and judge like a director. One result changed the lighting, so it is gone. One got the proportions wrong. One kept the corner of the table exactly where it was and had the ring on her finger. That is the shot. By the end of the board she is looking out of the window, cooking, picking pickles from the jar, and seen from the street through the glass, all with continuity.

Why this is the skill

Everything above happens before a single video generation, and it is the reason the video works. Clients do not pay for prompts. They pay for a character they signed off, in a room they chose, lit the way they agreed, delivered the same way every time. Do this and the generation step becomes the easy part.

Common questions

What tools do I need for this workflow?

Milanote for the board, ChatGPT or Gemini for writing prompts, Photoshop for the character sheet collage, and an image model. We use Nano Banana Pro on Higgsfield because the generations are unlimited on their plans, with GPT Image 2 for faces.

How many characters should I design before choosing?

Three. Two is not enough of a comparison and five is procrastination.

How do I keep the lighting the same across shots?

Extract a reusable lighting prompt from one frame you love and apply it to every other angle. Keep the hero shot in the prompt bar for every generation of that location.

Can I share the board with a client?

Yes, and that is the best use of it. They comment on the exact frame, you tag the column green when it is approved, and there is no back and forth over email.

Does this work for a commercial as well as a short film?

It works for anything with a character and a place. Commercials, narrative, social series. The board just gets more columns.

Want the board template and the video side of this workflow? Join the community.