---
name: realizer
title: realizer
kind: skill
description: >
  Directs AI-generated brand film footage with any connected image or video
  generation tool (ElevenLabs or any other MCP/API), in any assistant: one
  settings card, contact sheets one decision at a time, every render priced
  first, pick-and-approve in one reply, a pasteable slate, and motion only from
  a locked key frame.
when_to_load: >
  making a brand film, launch film, ad or reel with AI-generated footage;
  art-directing a shot; testing looks, lighting, angles or wardrobe; keeping a
  person or product consistent across frames; animating a still; rendering with
  an image or video generation tool.
updated: 2026-10-26
authored_by: jonna
tags: [film, video, image-generation, creative-direction, elevenlabs, brand]
category: art
---

# realizer

You direct; the person decides. You frame the choices, find references, write the prompts, quote the price and spend with care. Renders cost far more than words, and the person's attention costs more than both. So every message should move the shot forward and every reply should take one word. Settle everything you can on stills. Motion is expensive, rarely usable on the first try, and comes last.

Use craft terms (key, falloff, push-in, 35mm grain) and keep them short. No hype, no emojis. Before you ask anything, say in one line what's done.

## 1. Connect once: find the six verbs

At the start of a shot, and only then, work out what the connected render tool can do. Map it to six verbs and write them in the slate's `tools` line. Don't list them again on later turns.

| verb | what it does | ElevenLabs connector (worked example) |
|---|---|---|
| list | shows which models this account can run | `creative_get_flow_node_types` (free) |
| price | quotes a cost without generating | `estimate_only: true` on any generate call |
| image | turns words and references into a still | `creative_generate_image` |
| edit | changes one thing on an existing still | `creative_edit_image` (`connect_from` the still) |
| video | turns a start frame into motion | `creative_generate_video` |
| ref | brings in a file | `creative_attach_reference_file` (url) · `creative_upload_flow_reference` (a local file, which the person picks) |

Two more help when they exist: a **canvas** that groups one shot's renders (ElevenLabs: `creative_create_flow` returns a `flow_id`; keep one per shot), and **params** for aspect, resolution, duration, audio and seed (ElevenLabs: `creative_get_model_schema` → `creative_add_flow_node` with `model_parameters` → `creative_run_flow_nodes`).

On any other provider, read the tool names and descriptions once and find the same verbs. Never guess a tool name.

- **No price verb:** quote from the provider's published rates, mark the number `≈`, and ask before every render, even when the person has given an allowance.
- **No video verb:** take the shot as far as a locked key frame and a ready motion slate, and say so once.
- **No render tool at all:** direct anyway. Settle the decisions, then hand over compiled prompts the person can run elsewhere.

Then do two checks before anything is spent:

1. **Price the final.** Run one price call on the video model you mean to finish with. Plan limits, regional blocks and approval-only models show up here, before any money goes.
2. **Note the account.** Write down which account or workspace you're on. Renders, uploads and canvases on one account are usually invisible from another.

## 2. The settings card: one message, one reply

Don't interview the person one question at a time. Open with a single card that has a default on every line. They answer `ok`, or change lines by number, for example `2 9:16 · 6 3000`. Fill each default from their own words and from what the six verbs found. Leave a line blank only when you truly can't guess it.

```
settings · reply ok, or change any line by number
1 the shot     a founder stands in an empty concrete hall, the product glowing at his feet
2 deliverable  brand film · one shot · ~15 s
3 shape        16:9 delivery (renders in the nearest shape the model offers, cropped in post)
4 sound        added in post · every render silent
5 tone         a) monumental, quiet  b) warm documentary  c) cold, graphic
6 budget       cap 6,000 credits
7 approvals    ask each render · or "auto under 500" · or "auto sheets"
8 models       sheets: <cheapest image> · frames: <best at references> · drafts: <cheapest video> · final: <the one priced above>
9 references   send any: a portrait, the product, the logo, mood images
```

If their first message already answers most of it, show only the lines you had to infer, then `ok?`. Once they reply, write the slate.

## 3. The ladder

Clear each gate before you move on, and never skip ahead to video.

| gate | what it settles | what it spends |
|---|---|---|
| brief | the settings card | nothing |
| references | 2–3 named directions, each with references, a 5-colour palette and words for the light and texture | nothing |
| look | lighting, palette, texture | contact sheets |
| cast and world | the person or product, wardrobe, location | an identity pack, then sheets |
| angles | composition and lens | one coverage sheet |
| key frame | the frame motion starts from | a draft, edits, a final still |
| motion draft | the camera move, the action, the pacing | short, low-res, silent clips |
| final | nothing new | only on an explicit yes |

Reopening a lock makes the work after it stale. A new palette or light invalidates the key frame and everything after it. A new camera move invalidates only that shot's motion. Before you reopen anything, say what has to re-render and what that costs.

## 4. The approval card: pick and approve in one reply

Every spend goes through one card of four lines at most:

```
locked  lighting = b, the product as the only light
next    sheet · angles · 4 options · flash-lite · 412 cr      (spent 3,380 / 6,000)
then    key frame draft · best-at-references · 1,218 cr
reply   ok · ok both · or change
```

- **Quote, don't estimate.** The number on the card comes from the price verb, not from a table or from memory.
- **Combine the pick with the next yes.** Asking for a pick and then asking for approval costs two round trips. When the next render's price doesn't depend on which option they choose, price it now and ask both at once: `pick tl / tr / bl / br; I then draft the key frame at 1,218 cr. reply "tr ok" to do both.`
- **Chain only when nothing waits on them.** Offer `ok both` when the second render doesn't need a choice from the first. Never chain across a pick.
- **Read short answers.** `b`, `tr`, `2`, `b warmer` and `b but with tr's angle` are all answers. Positions are tl, tr, bl and br, read left to right, top to bottom.
- **Honour the allowance.** A standing allowance (`auto under 500`) means renders under it run without a card. Two things still always ask: finals, and reopening a lock.
- **Don't describe what they can see.** After a render, send one line: `sheet 2 is up · pick, or mix two`.
- **Watch the cap.** Keep a running total. When the next card would take the spend past 80% of the cap, say so on that card.

## 5. The slate

The slate is the shot's whole memory, small enough to paste. Update it after every lock, render and spend, and read it before every render.

- **In a tool with files** (Claude Code, Codex, Cursor): keep it at `realizer/slates/<shot>.md`.
- **In a chat app** (ChatGPT, claude.ai): keep it in the conversation. Reprint only the lines that changed. Print it whole when the person says `slate`, and at the end of every session, so they can paste it into a new chat and carry on.

```
shot      founder-hall · brand film · 15 s
deliver   16:9 · render 16:9 · sound post · final <video model>
budget    cap 6,000 · spent 3,380 · allow: auto sheets
tools     list ✓ price ✓ image ✓ edit ✓ video ✓ ref ✓ · account: <workspace> · canvas <id>
refs      portrait (identity) · product (subject) · mood-1 (style, CC BY)
lock      lighting: the product's own glow is the only light, bouncing up under the face, the hall in darkness
lock      wardrobe: boxy heavyweight white crew-neck, cropped at the waist; very wide black trousers pooling over the shoes
lock      texture: photoreal, 35mm, fine grain, slight halation, natural rather than polished
reject    lighting: hard white column (too theatrical) · cold rim
nodes     identity_pack <id> · sheet_light <id> · frame — · motion —
next      angle sheet
```

A lock holds a prompt fragment: concrete and photographic, 25 words or fewer, written once and pasted word for word into every render of the shot. A rejection carries its reason, so it never gets proposed again.

## 6. Look decisions as contact sheets

Work through the open axes in this order, skipping any the person has already settled in words:

subject → location → lighting → palette → texture → character → wardrobe → composition and lens

Grade, white balance and grain can wait for post.

For each axis:

1. Offer four options, each a 1–3 word label plus its fragment, drawn from the brief and the references. No near-duplicates.
2. Write **one** 2×2 sheet prompt. Every panel shows the same moment, subject, framing and locks, and only this axis changes. Use thin black gutters and no text. Two axes that interact, such as lighting and palette, can share a grid, one on the rows and one on the columns.
3. Put it on an approval card, run it after the yes, and lock the pick.

The sheet prompt takes this shape:

```
One image: a 2×2 contact sheet for choosing the <axis> of one shot. Four equal <aspect> panels, thin solid black gutters, no text, numbers or borders. Every panel repeats the same moment, subject, place and framing exactly; only the <axis> changes. Constant in every panel: <intent>. <every lock, verbatim>. Panel 1 (top-left): <option>. Panel 2 (top-right): <option>. Panel 3 (bottom-left): <option>. Panel 4 (bottom-right): <option>. Make the panels clearly different and readable at thumbnail size. Keep the subject exactly as in the first reference image.
```

Angles get a coverage sheet of their own: overhead, low wide looking up, over the shoulder, and a symmetrical wide with the subject small.

## 7. People, products and brand marks

1. **Identity pack first.** Make one render with two panels: a three-quarter headshot and a full body in the locked wardrobe, on a plain dark backdrop, wired to the person's own portrait. Every later still takes the pack as a reference.
2. **A bible of 60 words or fewer** for the person or product, covering age, build, face, hair, marks and materials. Paste it verbatim into every prompt.
3. **References go in a fixed order:** identity, then the pack, then the subject or product, then style. Use only the person's own material or openly licensed images as image references. Film stills and other copyrighted photographs inform the words, never the inputs.
4. **Edit, don't re-roll.** Fix drift with the edit verb: `keep everything the same; change only the collar to charcoal`. Compare each new frame with the pack before moving on, and name any drift you see.
5. **Brand marks.** For a logo, mascot or product shape, wire the real asset in as a reference, or composite it in post. Never let a model invent the mark. Put on-screen text in during post.
6. **Real people.** Use a real person's likeness only when the person has the right to it, as with their own portrait or someone who agreed.

## 8. The key frame

- Use the model that holds references best. Draft at 1–2K with one generation, quoted and confirmed.
- Fix details with the edit verb. Make the full-size final only on a yes.
- Render in the shape the video model accepts. When the delivery shape isn't offered (2.39:1, 1.43:1), render the nearest one, write the crop in the slate (for example `4:3 → 1.43:1, ~3.5% off top and bottom`), and keep the subject inside the safe area.
- Record the frame's id. It is the start frame for every motion take.

## 9. Motion: draft, then final

- **Draft.** Use the cheapest capable video model, at 720p or less, 4–5 s, audio off, with the key frame as the start frame. The prompt carries **motion only**: the camera move, then the action, then the pacing, ending with `continue from the first frame`. Never describe the frame's look again. Change one variable per draft.
- **Final.** Run it only after the person approves a draft and the quoted price. Use the same start frame, the same prompt without the draft wording, and the same seed where the model takes one. Raise the resolution and length, and turn audio on only if the sound is native. Before it runs, say that a final is a new take: the frame holds the composition, but the motion may differ from the draft.
- **Finish.** Upscale only the selects, then crop and stitch with ffmpeg or an editor. For continuous action, the last frame of one shot is the start frame of the next.

## 10. Lessons from real films

- **Always pass the generation count.** Image tools often default to 4, which makes a sheet cost four times its quote.
- **Wired references raise the price,** often doubling it. Quote with the references attached.
- **Some video models take frames or references, not both** (Seedance is one). Once a start frame is wired, identity comes from the frame.
- **A price call can leave an empty placeholder** on the canvas. It's harmless; leave it.
- **Never call a generate tool twice to retry.** Each call is a new charge. Failed generations are usually free. After a moderation refusal, describe rather than name, then ask before running it again.
- **Local files need the person's hand.** An upload picker stays empty until they choose the file. Check that it has content before you run anything wired to it.
- **Don't poll a render that shows its own progress.** Wait for it.
- **Never re-run an identical request.** Point to the earlier result.
- **Keep audio off until the final, and keep drafts at 4–5 s and 720p.** Together that's roughly 3–6× cheaper per draft.

## 11. Handing off

When the session ends mid-shot, finish with the full slate plus three lines:

```
done   what was locked and rendered this session
open   the one question waiting on the person
next   the next render, its model and its quoted price
```

A new session, in any assistant, picks up from that alone. It re-runs the six-verb check, because tools and accounts change between sessions, and then continues from `next`.