v2.5.6: Straight into production at script sign-off

v2.5.6: Straight into production at script sign-off

A small update to VideoExpress Scripting 2.5. VideoExpress Prompt is unchanged at 1.1.0.

Straight into production. Until now nothing marked the point where production starts in VideoExpress: after the script was signed off, the session went on to the Bible and pack without handing you a prompt to start from, and the browser workflow only began when you asked for it. Now, as soon as the script is signed off, the skill gives you the first paste-ready prompt in pack order (usually the first Consistent Character reference, otherwise Image 1A), in full, with its settings: aspect ratio, Use Creative mode, Automatically enhance my image prompt off, Consistent Character on or off. That happens whoever operates VideoExpress. If the assistant is driving VideoExpress, it also asks whether to open the two VideoExpress tabs now, and on yes sets up the Creation and Review tabs, the project and the Create Video From Prompt dialog.

What's new as a list. The what's-new summary is shown as a bulleted list, a heading per version and a bullet per item, never folded into a sentence. This was already on main after 2.5.5 (in both skills); 2.5.6 is the first release that carries it.

Where it changed. SKILL.md, references/project-intake.md, references/browser-workflow.md and references/changelog.md in both builds, the scripting manifests and CHANGELOG.md.

Upgrading. Marketplace installs pick this up on their next sync or update. If you use the zip, remove the old skill and upload the new one. In Grok, ask Grok to reinstall the skill from the same folder link, or let its regular update checks pick it up.

Release packages. videoexpress-scripting-claude.zip and videoexpress-scripting-chatgpt.zip, and for VideoExpress Prompt 1.1.0 (with the bulleted what's-new list) videoexpress-prompt-claude.zip and videoexpress-prompt-chatgpt.zip, each with the skill folder at the zip root.

Both builds (Claude, ChatGPT) are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.5.5: Clip review, video and sound checks; VideoExpress Prompt 1.1

v2.5.5: Clip review, video and sound checks; VideoExpress Prompt 1.1

An update to VideoExpress Scripting 2.5, with VideoExpress Prompt 1.1.0.

VideoExpress Prompt 1.1.0

What's new. The first time the prompt skill is used in a conversation, it asks whether you'd like a summary of what's new, from its own changelog. Where memory is on, it asks once per version instead of in every conversation; without memory it asks on first use in each conversation.

Conversations named after the shot. Once the shot has a subject, the skill renames the conversation "[subject] · VideoExpress Prompt" where it has a tool for that (Claude Desktop's session title, Grok). ChatGPT can't rename chats, so the name is left as it is there.

Chat or Code (Claude build). After the what's-new question, the skill checks whether it can reach files on your computer. In Claude's Chat it tells you once that it can't open or save files there: upload images to the chat and copy the prompts from it, or use the Code tab in Claude Desktop to work with your files. In Code it says nothing.

Starting with just the skill's name. Called with only its name or slash command, the skill goes straight to the "Do you have something to start from?" card instead of a list of options in text. A first message that already says what's wanted is acted on without asking.

VideoExpress Scripting 2.5.5

Clip review (frames and sound). When you ask for a review of your takes, or hand take picks to the assistant, each finished take is checked inside VideoExpress without downloading or playing it. A page script fetches the take from the library and draws twelve frames from start to end and a spectrogram of its soundtrack on one sheet, with per-second level and high-band share. The review reports against the prompt beat by beat: the opening frame, each action beat at its planned time, the camera move and its direction, the ending frame and whether the next clip can chain from it, and drift in characters, outfits, props and location. For sound it reports whether there is a track at all, overall level, steady hiss or hum, a music bed, and whether planned sounds and speech land at their times. It says its limits when reporting: it can't tell which words are spoken or check lip-sync, names a sound's source only as a likely guess, and doesn't see motion between the sampled frames. Part of the browser workflow; the what's-new summary leads with it.

Conversations named after the project. Once a new project has a working title, from your idea or the brief, the skill renames the conversation "[title] · VideoExpress" where it has a tool for that (Claude Desktop's session title, Grok). ChatGPT can't rename chats, so the name is left as it is there.

Chat or Code (Claude build). After the what's-new questions, the skill checks whether it can reach files on your computer. In Claude's Chat it tells you once that it can't open or save files there: upload scripts or images to the chat and download the pack when it's ready, or use the Code tab in Claude Desktop for a full project. In Code it says nothing.

Private takes. Some takes come back from the media list with no public file link, and the review stopped with "take not found or still rendering". It now checks the take's status separately and fetches the same file through the library's download address.

Starting with just the skill's name. Called with no message of its own (only its name or slash command), the skill treats it as a new project and goes straight to the "How do you want to start?" card instead of a list of options in text. A first message that already says what's wanted is acted on without asking.

What's new once per version. Where memory is on, the what's-new questions are asked once per version instead of in every conversation: the skill checks memory it can read for a note that this version's questions were shown, skips them if so, and saves that note (and says so) after asking. Without memory it asks on first use in each conversation, as before.

What's new. The summary now covers 2.5.5, 2.5.1 and 2.5 together, and only asks a headline item's question when it has one (Clip review is built in, so there's nothing to install).

Where it changed. Scripting: SKILL.md, references/browser-workflow.md and references/changelog.md in both builds. Prompt: SKILL.md and a new references/changelog.md in both builds. Both skills' manifests and CHANGELOG.md.

Upgrading. Marketplace installs pick this up on their next sync or update. If you use the zip, remove the old skill and upload the new one. In Grok, ask Grok to reinstall the skill from the same folder link, or let its regular update checks pick it up.

Release packages. videoexpress-scripting-claude.zip and videoexpress-scripting-chatgpt.zip, and for VideoExpress Prompt videoexpress-prompt-claude.zip and videoexpress-prompt-chatgpt.zip, each with the skill folder at the zip root.

Both builds (Claude, ChatGPT) are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.5.1: What's new is asked on first use

v2.5.1: What's new is asked on first use

A small update to VideoExpress Scripting 2.5. VideoExpress Prompt is unchanged at 1.0.0.

What's new is asked on first use. In 2.5.0 the skill asked whether you'd like a summary of what's new, and offered the new VideoExpress Prompt companion skill, only at the start of a new project. A session that opened with anything else, such as a single prompt or a question about the skill, never saw them. The questions now come the first time the skill is used in a conversation, whatever the first message is, once per conversation, and the skill then answers the message you sent. The summary covers 2.5 and 2.5.1 together, so updating from 2.4 still shows everything that changed in 2.5.

Changelog wording. The Consistent Character entry now says plainly that VideoExpress hasn't changed: Consistent Character has only ever affected Create Image, and it's the skill's guidance that was corrected.

Also on main since 2.5.0. VideoExpress Prompt's SKILL.md no longer says a clip sees reference images (a text correction with no version change; the v2.5.0 prompt zips were replaced with the corrected files).

Where it changed. SKILL.md and references/changelog.md in both builds, the scripting manifests and CHANGELOG.md.

Upgrading. Marketplace installs pick this up on their next sync or update. If you use the zip, remove the old skill and upload the new one. In Grok, ask Grok to reinstall the skill from the same folder link, or let its regular update checks pick it up.

Release packages. videoexpress-scripting-claude.zip and videoexpress-scripting-chatgpt.zip, each with the skill folder at the zip root. For VideoExpress Prompt, use the zips on the v2.5.0 release.

Both builds (Claude, ChatGPT) are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.5.0: VideoExpress Prompt 1.0, Grok support, what's new on update, and Consistent Character corrected

v2.5.0: VideoExpress Prompt 1.0, Grok support, what's new on update, and a Consistent Character correction

VideoExpress Scripting 2.5.0 and the first release of its companion skill, VideoExpress Prompt 1.0.0. New rules for seated action, Lipsync HD gestures and misread words, a correction to how Consistent Character is used, a narrower numbers rule, one new style, and Grok as a third platform.

New companion skill: VideoExpress Prompt 1.0.0. A step-by-step wizard for one prompt: a start image, a clip, or an image and its clip, or an improved version of a prompt you already have. It asks about one aspect of the shot at a time (look and style, what happens, dialogue, sound, camera, light and image, negatives) and applies the same prompt rules, multishot format, style sheets and lip-sync methods as VideoExpress Scripting. Use it for a single shot; use VideoExpress Scripting for a whole multi-clip video. It installs from the same marketplace (install guide); builds are in claude-prompt/ and chatgpt-prompt/.

Grok support. Grok uses the platform-neutral ChatGPT build. In the Grok app (free trial and paid plans only), ask Grok to install the skill from its GitHub folder link and answer yes to regular update checks; the local Grok Build CLI can use npx skills add instead (install guide). Give Grok the folder link, not the repo's home page, which also holds the Claude builds.

What's new on update. At the start of a new project the skill asks whether you'd like a short summary of what's new in the installed version, then offers headline additions, this time VideoExpress Prompt with a link to install it (skipped if it's already installed). The notes live in references/changelog.md, and every release is listed in the new CHANGELOG.md, which the plugin manifests now link to.

Corrected: Consistent Character in clips. VideoExpress hasn't changed: Consistent Character has only ever affected Create Image, and Create Video ignores its tick and reference slots. Earlier versions of the skill wrongly treated it as affecting clips.

• A video prompt's [REFERENCE USE] now names only the start image, the exact opening frame. The template, worked example and bible no longer tell the model to use supplied reference images in a clip.
• The browser workflow no longer sets Consistent Character or its slots before Create Video.
• Turning Consistent Character off for shots with no characters is stated as a start-image setting.
• Removed: the claim that face references turn clips realistic, which was never tested. The style anchor rule is unchanged.

References and pose. A reference in either slot can carry its pose and props into the image, though it doesn't always; a pose stated in the prompt overrides it. State the pose and how each prop is held.

Sitting and walking to a mark. VE doesn't model collision or which way furniture faces: a character sat down through a café table, and a chair turned toward the water was still drawn facing its table, so he sat on it sideways.

• Use a seat with no front (a crate on its end, a stool, a low wall, a bollard) with clear space around it and nothing beside it on the path. It sat correctly in six of six takes.
• A seat or mark the character walks to must be clearly nearer the camera than them in the start image, with visible ground between, and the walk short (1–2 s). When it was level with or behind them, they walked past it in every take.

A gesture under a Lipsync HD line. The Lipsync dialog's Video Prompt does steer one directed gesture, described in frame terms: which arm by frame side, how high, the target third of the frame, when it starts, a hold while they speak, and the return to the opening pose. Points at six frame zones each landed. Multi-step action under a line is untested; give it its own clip.

Tagged start-image prompts. Start images can use [SUBJECT], [FRAMING AND POSITION], [BACKGROUND], [LIGHT AND STYLE] and [NEGATIVES] tags as well as prose. In testing the tagged form held subject, props, framing, background, light and negatives as well as prose, with no tag text in the image. Keep one form throughout a pack.

Misread words. When VE mispronounces a word in a spoken line, respell it inside the quotes ("Skrip-ting") and keep a plain-English note on the sound in [ACTION], outside them. IPA or a note outside the quotes alone didn't change it; never put IPA inside the quotes.

Numbers. The blanket "digits only in labels" rule is replaced. In testing, digits in ages, durations, distances and counts produced no numerals in the frame, in the start image or the clips. The rule now covers text-bearing objects only: describe signs, screens, labels and holograms as abstract shapes and light, and end every [NEGATIVES] with "no text, no letters, no numbers". Counted events ("blinks three times") are still out.

Chaining inside VE. Clips chain from VideoExpress's Save Last Frame, with no trimming or retiming outside VE. Plan the framing so a clip ends where the next one begins; where a chain risks detail drift, the skill flags it once and you decide.

New style: Graffiti (graffiti.md). Aerosol spray-paint toon with caricatured characters on a street-mural ground. Demo clip to follow. That makes 45 Creative mode styles.

Where it changed. SKILL.md, bible.md, camera-and-motion.md, light-sound-dialogue.md, multishot-template.md, prompt-rules.md, browser-workflow.md, project-intake.md, styles/README.md, styles/anime-1990s-cel.md, the new styles/graffiti.md and changelog.md, plus the manifests, READMEs and CHANGELOG.md.

Upgrading. Marketplace installs pick this up on their next sync; install VideoExpress Prompt from the same marketplace. If you use the zip, remove the old skill and upload the new one. Prompt packs are worth updating before their next generation: take any "use the supplied reference image of …" wording out of video prompts' [REFERENCE USE], so it names only the start image.

Release packages. videoexpress-scripting-claude.zip, videoexpress-scripting-chatgpt.zip, videoexpress-prompt-claude.zip and videoexpress-prompt-chatgpt.zip, each with the skill folder at the zip root.

Both builds (Claude, ChatGPT) of each skill are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.4.0: Positives only, no requested silence, details that hold their size, and a new anime style

v2.4.0: Positives only, no requested silence, details that hold their size, and a new anime style

How VE reads exclusions and silence, a framing rule for fine detail, lip-sync in other languages, one new style, and a tighter browser workflow.

Exclusions work only in [NEGATIVES]. VE reads every other section as things to include, even with "no", "not" or "never" in front: "no music, no hum" inside [PRODUCTION SOUND] adds music and hum, and "never reversing" in [CAMERA] names reversing. VideoExpress's developers give the same guidance.

• Write positives everywhere else: what should exist, what stays still, how the camera holds ("then holds on that framing").
• Every exclusion, sound included, goes in [NEGATIVES], each written with "No" ("No music, no hum"). In testing, a bare list there ("music, hum, rumble") added the listed sounds.
• The rule, the templates and the worked example are updated; the failure table has a new row for an excluded sound, motion or object that appears anyway.
• Style sheet anchors are unchanged: their tested wording stays as it is.

Never ask for silence. A segment or clip whose sound is "silence" renders as a low rumble and crackle, because VE's audio has a noise floor that shows whenever it has nothing to generate. This replaces the old "complete silence before, between and after them" and "silent segments say so" advice.

• Every segment carries sound tied to its own action; keep pauses under a second and give any quiet stretch a soft action sound (cloth shifting, a boot settling on stone).
• Fit the manual length to the action and speech, so there is little empty time.
• New failure-table row for rumble and crackle under a pause.

A detail holds only at the size VE last drew it. A face or prop that is small in the opening frame is redrawn from the text when a push-in or a cut to a closer framing enlarges it, and fine detail (a machine eye, a scar, a weapon's mechanism) drifts.

• Start each clip from a start image at its closest framing and keep one framing per clip; a closer shot is a new clip from its own start image.
• Chaining works at any distance while the camera doesn't enlarge the face or a prop, so a long talking scene can chain close-up into close-up.
• The full-body ending rule is relaxed: end at the framing the next clip needs.
• The hard-cut dialogue tip now notes that a cut to a closer framing can drift a detailed face.
• New failure-table row for face detail drifting during a push-in or after a cut.

Lip-sync in other languages. Write the line in its own language and script (for example Japanese in kanji and kana), quoted once in [ACTION], with "speaking in [language] as a native [language] speaker" in the voice description. Captions are added in the edit.

New style: Anime: Shōnen Cyberpunk (anime-shonen-cyberpunk.md). Shōnen line and action with cybernetic character design: thick ink outlines, hard two-tone cel shading, metal drawn as flat cel colour, a cool palette with one hot accent, and the fast-beat-then-long-hold cadence. Tested in production. Demo clip to follow. That makes 44 Creative mode styles.

Reference images made in VE. For a detailed character, the staged reference route now iterates on VE's own best head shot and cleans its artefacts before it goes into slot 1. Another image tool is offered only as a last resort, when VE consistently fails after iteration.

Browser workflow.

• Always two tabs (Creation and Review); no longer asked.
• Every review question says what a correct image or take shows, from its prompt.
• Ask which take to keep straight after submitting, instead of first asking whether the renders have finished.
• Refresh the library once a tile reaches 50%: the grid stops updating partway through a render, and the take has often already finished.
• Known VE behaviours written in as steps: one image candidate per pass with Consistent Character off and two with it on, the Manual Video Length checkbox, the orientation toggle, clearing reference slots after unticking Consistent Character, the missing "Your video will appear" message, naming a new project on its first save, and finding saved images by their [Image X] prefix.

Where it changed. prompt-rules.md, multishot-template.md, light-sound-dialogue.md, camera-and-motion.md, bible.md, browser-workflow.md, styles/README.md, the new styles/anime-shonen-cyberpunk.md, and the rules in SKILL.md.

Upgrading. Marketplace installs pick this up on their next sync. If you use the zip, remove the old skill and upload the new one. Prompt packs are worth updating before their next generation: move any "no …" phrases out of [PRODUCTION SOUND] and [CAMERA] into [NEGATIVES], and replace requested silence with a soft action sound.

Both builds (Claude, ChatGPT) are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.3.0: One sound block per segment, character sheets in stages, and props that hold their shape

v2.3.0: One sound block per segment, character sheets in stages, and props that hold their shape

An update to how production sound is written, and a set of character and prop techniques proved in production. No new styles.

Every segment gets its own [PRODUCTION SOUND] block. The template used to put one [PRODUCTION SOUND] block at the end of the prompt, after the last [ACTION] / [CAMERA] / [LIGHT AND IMAGE] trio. In production, VE binds that block to the last segment only and makes up sound for the earlier segments, often as music. Each segment is now a group of four, [ACTION 0-4s] → [CAMERA] → [LIGHT AND IMAGE] → [PRODUCTION SOUND 0-4s], with the sound block carrying its segment's time label.

• A segment meant to be silent says so in its own block instead of being left without one.
• Each block fills its own segment and closes with what is heard in that segment.
• A dialogue line is still quoted once, in [ACTION]; the sound block of the segment where it's spoken refers to it.
• In testing, one block per segment gave five music-free takes of a clip whose earlier versions had music. The unresolved climax-music case in the Sound section now points to this as the first thing to check.

Detailed character reference sheets, built in stages. VE's image model loses features buried in the middle of a long prompt; a detailed face dropped its defining features in prompts of about four thousand characters and held them under about fifteen hundred. For a highly detailed character, the bible now gives a staged route:

• A head shot first, with a short prompt that puts the face first.
• If VE can't draw the design from words, generate the head shot in another image tool, import it and put it in Consistent Character slot 1; the reference carries the design and VE redraws it in its own style.
• The full body next, with the head shot in slot 1 and a short prompt about body, outfit and props.
• A composite sheet (full body, large head shot, key props; one sheet per prop state, such as sheathed and armed) in slot 1, which leaves slot 2 free for a second character. In testing it gave exactly one figure with no layout bleed in every take.
• A reference's stance can carry into start images, so a start image that needs a turn puts the character mid-action.
• The long-prompt advice for style now notes this exception.

Props that hold their shape.

• Hybrid props collapse into their stronger half: a "sword-pistol" was drawn as a handgun with a blade attached, in VE and in other tools. Describe the parts and their layout as geometry, keep the stronger noun out, and simplify the design if it keeps reverting.
• A held prop that leaves the frame, through a push-in or a swing, is redrawn from the text and can come back a different object. Weapons stay fully in frame through their action, with the camera locked and the strike ending inside the frame.
• The failure table has four new rows: prop redraw, hybrid collapse, features lost in long image prompts, and a troubleshooting suggestion for metal spreading onto a body part named in the same sentence.

Where it changed. multishot-template.md (section order, per-segment groups, the sound section and all three templates), light-sound-dialogue.md, prompt-rules.md, camera-and-motion.md, bible.md, styles/README.md and the sound rule in SKILL.md.

Upgrading. Marketplace installs pick this up on their next sync. If you use the zip, remove the old skill and upload the new one. Prompt packs written with a single sound block at the end are worth updating before their next generation: split the block by segment.

Both builds (Claude, ChatGPT) are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.2.1: Clip references in every prompt, and install from the plugin marketplace

v2.2.1: Clip references in every prompt, and install from the plugin marketplace

A small update to 2.2. No new styles, and the production method is unchanged apart from how prompts are labelled.

Every prompt starts with its pack reference. When working from a prompt pack, each prompt now opens with the reference the pack declares for it, in square brackets: [Clip 12] for a video prompt, [Image 12A] for a start image, [Clip 12 TEST 1] for a test variant. This replaces the optional caption label ("Harbour Run clip four, the landing."). Library tiles show the first words of a prompt, so the reference is what matches each tile and take to the pack, and how takes are reported ("Clip 12, take 3").

• The pack declares a reference for every start image and clip, and keeps them stable once generation starts. Alternates get their own reference (Clip 12B), never a reused one.
• The digits rule allows digits in the reference as well as the [ACTION] time labels.
• For custom styles, the style anchor goes straight after the reference, before [REFERENCE USE].
• The templates, worked example, browser workflow and styles index are updated to match, and the failure table has a new row for takes that can't be matched to their clip.

Install from the plugin marketplace. The repo is now a plugin marketplace for Claude Desktop and ChatGPT. Add ratava/videoexpress-scripting-skill once and install VideoExpress Scripting from it; in Claude Desktop, updates arrive automatically while Sync automatically is on. To support this, each build's skill folder moved into a plugin root: claude/skills/videoexpress-scripting/ and chatgpt/skills/videoexpress-scripting/, with plugin manifests alongside.

Docs. New wiki walkthroughs with screenshots: Install in Claude Desktop and Install in ChatGPT, plus a Requirements and Recommendations page. The ChatGPT guide was briefly titled "ChatGPT Codex"; the skill is for ChatGPT, and the old link now points to the renamed page.

Upgrading. Marketplace installs pick this up on their next sync. If you use the zip, remove the old skill and upload the new one; if you copy the folder from the repo, note the new path above.

Both builds (Claude, ChatGPT) are identical in substance. Scope remains VideoExpress 3.5 and later.

v2.2.0

v2.2.0: 43 Creative mode styles, demo clips for every style, and a wiki

32 new tested Creative mode styles. The style library grows from 11 to 43. Every new style was tested in VideoExpress 3.5 with its own start image and a ten-second clip. Each test clip shows the style name as a title for three seconds, then fades it out and performs one action with one camera move. Every style passed, and each has its own sheet in references/styles/. A sheet gives the style's medium phrase (the opening words of every image prompt), its video style anchor, the stability line that closes every [LIGHT AND IMAGE], motion guidance, guards, and the known failures from the test take.

• Painted and drawn: Charcoal and Chalk, Ink and Wash, Gouache Picture Book, Oil Painting, Acrylic, Pure Watercolour (no line), Pencil Sketch, Pencil and Watercolour Picture Book ("Little Golden Book").
• Print and craft: Linocut, Risograph, Engraving, Chalkboard, Blueprint, Embroidery, Tissue-Paper Collage, Mosaic, Stained Glass, Origami.
• Japanese painting: Nihonga, Suibokuga (sumi-e, with kasure and nijimi named).
• Vintage animation: Rubber-Hose Cartoon (1930s black-and-white cel).
• Hybrids (two media in one frame): Pencil on Gouache, Charcoal on Watercolour, Clay Puppet on a Painted Set, Needle-Felt on Cut Paper, Pixel Sprite on Painted Scenery.
• Anime: Modern Digital TV, Shōjo, 1990s Cel, Shōnen Action, Cyberpunk (mid-1990s cel film look), Mecha.

The ten Create Mode presets and Cinematic Photoreal were tested the same way, so all 43 styles now have demo clips.

Demo clips and wiki. The styles index (references/styles/README.md) now links every style to its demo on YouTube, plus two master reels: the existing presets reel and the v2.2 new styles reel. The new project wiki has:

• A How-To home page with flow diagrams for install, the start methods (concept, script, existing images), intake, the prompt pack, the three operating modes (manual, automated with Claude or an agent driving VideoExpress, shared), fixing problems, and finishing.
• A Style catalogue with one page per style: demo clip, thumbnail and full style sheet.

Restructured skill. SKILL.md is now a short router (about 1,700 words, down from 7,700). It keeps the two driving facts, the workflow and the rules that apply to every prompt, plus a table saying which reference to read for each step. The detail moved, word for word, into references that load only when needed: bible.md, camera-and-motion.md, light-sound-dialogue.md and prompt-rules.md. create-modes.md is replaced by the styles/ folder: an index with the shared image recipe and video pattern, one sheet per style, and a _TEMPLATE.md for adding new ones. With this layout, adding styles costs nothing at read time: the assistant reads the index and the one sheet it needs.

What the tests taught (now in the sheets and the index):

• Mixed media: name the medium for each layer, state the separation as a rule ("strictly two media: … only, … only"), and rule out each cross-over. All five hybrids held with this wording.
• Pencil and watercolour: the picture-book look comes from pencil line and watercolour across the whole scene, not a pencil character over a painted background. The sheet is corrected. Pencil Sketch is corrected to full coloured pencil rather than a light tint.
• Blueprint: extrude, don't rotate. Asking a line drawing to turn made it spin out of control. Drawing it up from a plan held.
• Tile and glass styles: state "every tile / lead line stays exactly where it is" in the identity block, not only in the stability line.
• Titles in images: put the title sentence first in the image prompt and repeat it in the closing guards, or long character blocks can push it out.
• Look-alike characters: describe the whole design so a cartoon character can't drift toward a famous studio mascot.

Other changes. SKILL.md now records its version in the frontmatter (metadata.version: "2.2"), following the Agent Skills spec. The main README links the wiki and lists all 43 styles.

Upgrading. Replace the whole skill folder rather than copying files over it. references/create-modes.md no longer exists, and an old copy left behind would compete with the new styles/ folder. In Claude, remove the old skill and upload the new zip under Settings → Capabilities → Skills.

Both builds (Claude, ChatGPT/Codex) are identical in substance, and their 43 style sheets match byte for byte. Scope remains VideoExpress 3.5 and later.

v2.1.0: Lip-synced dialogue in the multishot prompt

v2.1.0: Lip-synced dialogue in the multishot prompt, and how the engine reads a prompt

New dialogue method (SKILL.md 6b, Method A). Quoted lines written inside the multishot prompt render as on-camera lip-synced speech, with Consistent Character on and Lipsync HD Video unticked. It works alongside actions, hard cuts and camera moves in a normal 3–10 second generation, so it is now the preferred method whenever a clip has anything to do besides talk. Lipsync HD (Method B) remains the choice for pure talking heads where its automatic timing matters.

• No automatic time calculation: design each dialogue clip for a set length (3–10 s or the 5 s default), set the Manual Video Length to match, and budget about two seconds per short line plus a beat per pause. Too long and the engine pads with invented business; too short and the last line or action is cut.
• Quote each line exactly once, in [ACTION], with explicit mouth movement and a voice description; [PRODUCTION SOUND] refers to it without repeating the words. VE's dialogue-length check ("Your dialogues are too long") counts everything in quotation marks.
• Nothing in the mouth and nothing edible in either hand at any chain point before a dialogue clip. A sip or a held pastry in the previous clip renders as chewing under the lines, and rewording does not fix it; the frame does.
• Every speaker needs a visible face; a line given to a head-out-of-frame character is dropped. Faceless speakers go in [PRODUCTION SOUND] as an off-camera voice.
• Voices bleed like adjectives between two speakers and toward the setting's language. Describe each voice specifically and repeat it every clip.
• Prompt box limit of about 6,000 characters.

New section 4a, "How the engine reads a prompt". Quotation marks mean speech; two speakers per clip maximum; positive anchors ("what stays still") over prohibition lists; continuity assumed until an action changes it; camera locked by default, every move physical and with a destination; one take unless a cut is declared; chronological present tense; sound tied to its action and closed positively; lettering never promised; why "Automatically enhance my video prompt" stays off; and where this skill's long tagged format deliberately departs from one-clip house style.

Updated. 6a Sound (positive closing line instead of music prohibitions). Failure table: nine new rows covering dialogue, quotation marks, auto-enhance, chewing, dropped lines, accent drift and push-in versus zoom. Multishot template: [NEGATIVES] as a short tail, new Quotation marks section, dialogue note in [PRODUCTION SOUND]. Project intake: dialogue method per clip and the two-speaker limit. Browser workflow: Lipsync HD and auto-enhance setting rows. README.

Both builds (Claude, ChatGPT/Codex) updated identically in substance. Scope remains VideoExpress 3.5 and later.

v2.0.0: VE 3.5 production learnings and guided project intake

v2.0: VE 3.5 production learnings and guided project intake

• New: guided project intake (references/project-intake.md). A question system that starts from a concept or a script, works through a decision checklist, runs a feasibility check, offers optional development tests, and produces a signed-off Core Bible.
• New guidance:
  • Consistent Character reference slots and their side effects.
  • Custom Creative mode styles and the style anchor.
  • Staging against the frame's real geometry, and cause-and-effect beat order.
    Transient-only production sound.
  • Lipsync with one or two actors.
  • Point-of-view and HUD shots.
  • A method for isolating stubborn failures.
• Browser workflow rewritten:
  • Five-take batches with position tracking.
  • A question-driven render cycle.
  • The Lipsync dialog procedure.
  • Recovery after reloads, and project saving.
• Scope: VideoExpress 3.5 and later by PaulPonna.com only.