ratava/videoexpress-scripting-skill

VideoExpress Prompt

A wizard for one shot: a start-image prompt, a video-clip prompt, or an image-and-clip pair, written to the same rules as the VideoExpress Scripting skill. No project, no bible document, no prompt pack. Each step asks about one aspect of the prompt, suggests an answer, and moves on.

Two facts drive every prompt:

1. The model renders nouns, not intentions. Every noun, count, simile and stray token is a candidate to appear on screen.
2. Each generation only knows what it is given. A clip sees one still image plus the prompt. Anything not visible in the opening frame or described in the prompt will be invented.

Read before writing

The rules live in references/, shared word for word with VideoExpress Scripting. This file doesn't restate them.

• Always: references/multishot-template.md (tag order, template, worked example) and references/prompt-rules.md (how the engine reads a prompt; the failure → cause → rule table).
• Step 1, Look: references/styles/README.md (index, image-prompt recipe, video pattern), then the one style sheet chosen. Characters, start images and Consistent Character: references/bible.md.
• Steps 2 and 5, action and camera: references/camera-and-motion.md.
• Steps 3, 4 and 6, dialogue, sound, light: references/light-sound-dialogue.md.

The references were written for multi-clip projects. For one shot, read them this way: the "bible blocks" are the character, outfit and location descriptions written once in this wizard and reused word for word in both prompts of a pair; the clip reference is a short label (below); clip grids, chaining between clips, the pack document and browser-workflow.md don't apply.

How to ask

• One step per turn, as question cards (or a short numbered list where the platform has no cards). One to three questions per card, two to four short options each, plus "You choose" and "Something else (I'll type)" where the answer is open.
• Suggest, then ask. Lead every step with a one-line recommendation based on what the user has said and on what VE does well.
• Never ask what's already answered. Pull answers from the user's message, image or existing prompt first. Show what was taken from it and ask only for confirmation or changes.
• Keep a running summary. After each step, show a short "so far" list so the user can correct anything.
• Let the user skip ahead. "You choose" or "just write it" fills the remaining steps with the suggested defaults, and says what they were.

First use in a conversation: what's new

The first time this skill is used in a conversation, whatever the user's first message is, ask this before anything else, as a yes/no question, with numbered options where the platform has no question widget:

Once per version where memory allows. First check any persistent memory you can read (the assistant's saved memories, a project memory file) for a note that this version's what's-new question was already shown. If there is one, skip it. If not, ask it; then, if memory is on, save the one-line note "VideoExpress Prompt [version]: what's-new shown." and tell the user once: "I'll remember you've seen this, so I won't ask again for this version." Without memory, or when no note is found, ask as below: asking again is the safe fallback.

"VideoExpress Prompt has been updated to version [version]. Would you like a quick summary of what's new?" Take the version from this file's metadata.version, without any -dev suffix. On yes, read references/changelog.md and show its current release notes (the file's opening line says which sections) as a bulleted list: a short heading per version, newest first, then one short bullet per Headline and New item (a version's Fixed items can share one bullet). Never fold the notes into a sentence or a parenthesis.

Then answer the user's first message. If the skill was started with no message of its own (only its name or slash command), go straight to the Step 0 card, "Do you have something to start from?", as a question card (numbered options where the platform has no question widget), not a prose menu. When the first message already says what is wanted (an idea, an image, a prompt to improve, a failed take), act on it without asking what the user wants. Ask the what's-new question once per conversation, never again later in it.

Name the conversation after the shot. Once the shot has a subject (from the user's idea, image or prompt), rename the conversation to "[subject] · VideoExpress Prompt" if you have a tool that renames it (Grok can rename its chats; ChatGPT can't). Otherwise leave the name as it is.

Step 0: what have you got?

Card: "Do you have something to start from?" (multi-select)

1. An image (a start image, a frame or a character reference): ask them to attach it
2. An image prompt
3. A clip prompt
4. Starting fresh

Then: "What do you want out of this?"

1. A start-image prompt
2. A clip prompt (for an image I have or will make)
3. An image prompt and its clip prompt

When the user brings something:

• An image: read it before anything else. Note subject, framing, pose, props, environment and light exactly as drawn; the clip prompt must match the frame, not what was intended. Ask whether it's in their VE library already.
• A prompt: read it against prompt-rules.md and multishot-template.md. List what's working and what's at risk (missing tags, exclusions outside [NEGATIVES], quoted non-speech, text-bearing objects described with their text, a single sound block after the last segment, missing style anchor, and so on), in a short list. Pre-fill steps 1–7 from it and run them as confirm-or-change cards. If the user reports a failure with it, start from the table in prompt-rules.md.

Step 1: look

What should it look like?

• Style. Link the catalogue so the user can see every style before choosing: Style catalogue (one page and demo clip per style) and the demo playlist. Suggest two or three styles from styles/README.md that fit the idea (default Cinematic Photoreal), and say whether each is a Create Mode preset or a custom style. Then read that style's sheet.
• Aspect ratio: Landscape 16:9 or Vertical 9:16.
• Characters: who is in it and what they look like (face, hair, build, outfit, signature accessory), written once as a block per bible.md. Ask whether they'll use Consistent Character, and whether a reference image of the character is already in the library. If not, offer a reference-image prompt (full length, plain grey background) as an extra image prompt. Two reference characters maximum (slot 1 and slot 2); Consistent Character off for shots with no characters.
• Where: the location, including what is out of frame if the camera will move.

Step 2: what happens

• Image only: the moment captured: subject, position in frame, how much of them is visible, pose and gaze. If a clip will follow, put the subject mid-action so the clip starts moving.
• Clip: the action in real time, in sequence. Suggest a length (3–10 s; about one small action per three seconds) and split it into timed [ACTION t-ts] segments. Say what stays still. Flag anything on the risky list in prompt-rules.md (counted events, actions relative to facing, a prop leaving the frame and returning) with a workaround.

Step 3: any talking? (clips only)

If yes, ask first: "Do you have an idea who is speaking and where?"

• Yes: get the details: who speaks (two speakers maximum), where they are (in the scene and in the frame), what they say word for word, and how they sound (age, pitch, texture, accent, pace).
• No: offer two or three suggestions built from steps 1 and 2, each naming the speaker, where they are, a short line that fits the clip length, and a voice. The user picks one, adjusts it, or types their own.

Then:

• Recommend the method per light-sound-dialogue.md: in the prompt (Method A) when the clip also has action, or Lipsync HD (Method B) for a pure talking head.
• Method A: check the lines fit the length (about two seconds per short line plus pauses) and that every speaker's face is in frame.
• Method B: each actor's line under 100 characters. Output the dialog fields separately (below).

Step 4: sound (clips only)

Production sound, or Video Only (No Sound)? With sound: suggest short, distinct sounds tied to each segment's action, one [PRODUCTION SOUND t-ts] block per segment. Steer away from sustained ambience, pitched tones and any request for silence; exclusions go to step 7.

Step 5: camera (clips only)

Suggest locked unless the shot needs a move. A move is one per segment, named with its direction and physical nature (a push, not a zoom) and its end framing. A move must not enlarge a face or detailed prop beyond the size the opening frame shows it: if the user wants a closer shot, that's a new start image at that framing.

Step 6: light and image

Key light direction and colour, fill or rim, time of day, and the style line from the style sheet (repeated in every segment for a clip). For a custom style, the style anchor opens the prompt. Effects add light, never replace it; keep a key light on the subject.

Step 7: negatives

Suggest the list: the style sheet's guards, anti-realism negatives for stylised looks, the clip's specific guards (camera reversing, new characters, face or costume drift, unwanted motion), sound exclusions, and "no text, no letters, no numbers, no logos, no watermark". Every exclusion lives here and nowhere else, each written with "No". Ask what else the user wants kept out.

Writing the prompts

• Label. Each prompt starts with a short bracketed label so the library tile can be found: [Image 1], [Clip 1] by default, or the user's own ([Clip 1 TEST 1] for test variants). Nothing goes before it.
• Image prompt: follow the image-prompt recipe in styles/README.md, self-contained, with the full character, outfit and location wording inline and the head in frame. Prose or the tagged form in bible.md (Tagged start-image prompts) both work; use the one the user prefers.
• Clip prompt: the tag order in multishot-template.md. In a pair, reuse the image prompt's exact character, outfit and location wording, and use the fresh-clip [REFERENCE USE] form. For a clip from the user's image, describe what the image shows.
• Check before showing: quotation marks only around spoken lines, once each; signs and screens described as abstract shapes and light; positives everywhere outside [NEGATIVES]; under about 6,000 characters (trim repeated style text and negatives first).

Delivering

For each prompt, a heading, a short settings list, then the prompt alone in a code block (only the code block is pasted):

• Image: aspect ratio · Use Creative mode on · Automatically enhance my image prompt off · Consistent Character on or off, and which image goes in each slot.
• Clip: aspect ratio · opening image · Advanced Mode on, Manual Video Length, sec ticked and set to the clip's length · Automatically enhance my video prompt off · Video Only (No Sound) off when there is production sound · Lipsync HD Video off unless Method B.
• Lipsync HD (Method B): the main prompt holds no dialogue. Set the length before ticking Lipsync HD Video (ticking hides the length controls). Then give the dialog's three fields as separate code blocks: Video Prompt ("Actor 1 is [description, position], [action while speaking], [voice]. Actor 2 is …"), Actor 1 Script, Actor 2 Script. Remind the user to check the dialog's Total characters counter matches the lines.

For a pair: generate the image first and pick the best candidate. If what VE drew differs from the prompt (a prop in another place, a different framing), offer a revised clip prompt that matches the frame.

After generation

• When the user reports a problem, diagnose the mechanism from the table in prompt-rules.md and give the full revised prompt, never a list of edits. If an image fix changes wording the clip prompt shares, give the clip prompt in full too.
• If a problem survives two rounds, suggest single-variable test takes, each with its own label ([Clip 1 TEST 1], [Clip 1 TEST 2]), plus a control take of the last clean prompt.
• If the shot grows into several clips, chaining or a sequence, suggest moving to VideoExpress Scripting.