Turn a scene idea into a production-ready prompt.
Movie Scene Director is a reusable instruction-and-skill package for planning detailed, single-avatar scenes in ChatGPT or Claude. Describe the scene and your creative intent; the agent helps shape a copy-ready HEYGEN PROMPT - ONE AVATAR. You review it, paste it into the appropriate HeyGen field, and generate the video yourself.
It helps make the setting, performance, dialogue, camera, timing, sound and continuity explicit—not just add cinematic adjectives. The intended benefit is a clearer brief and a more repeatable revision process; reduced production time, cost or generation attempts has not been measured.
Who it is for
- Solo creators, storytellers, personal-brand creators and independent video producers already using or planning to use HeyGen.
- Freelance video producers and small creative or content agencies preparing repeatable briefs for clients.
- Independent filmmakers, film students, educators, course creators, and brand or marketing creators exploring one-avatar scenes.
Client approval, generation, editing, delivery and factual review remain human tasks. These audience recommendations reflect the product scope, not validated demand, interviews or buyer trials.
A deliberate one-avatar scope
The supplied workflow uses exactly one selected visible avatar throughout a video; an insert may show none. It is not suited to multi-actor dialogue, automatic production, frame-perfect cinematography, guaranteed realistic action, or a complete film-production system.
A brief with the decisions that shape a scene
The instructions gather the scene purpose, audience, context, genre, duration, language, selected avatar, exact words, location, intended emotion, camera preferences, sound and required details. Each requirement is classified as a documented control, generated request, manual step or unresolved item, with its destination, evidence, acceptance condition and repair considered. The agent reuses supplied facts, labels proposals, and asks essential questions when a contradiction would undermine the result. Unknown routes, unsupported guarantees and essential missing assets keep the work in draft rather than finalizing it. A scene can be suspenseful, curious, empathetic, comic or quietly observational; it does not need a social-media hook or rapid cuts.
Story, avatar and performance
- Scene intention: what the viewer should feel or understand, what is revealed, and the emotional start and finish.
- Avatar continuity: one selected visible avatar, actual appearance, outfit and assigned voice, without an unrequested replacement.
- Performance: posture, gaze, gestures, facial expression, reactions and visual motion kept distinct from vocal delivery.
- Environment: location, background geometry, materials, props, atmosphere and visible states.
Camera, timing and sound
- Camera: shot size, physical position, vertical angle, movement, focus, eyeline and screen direction.
- Timing: desired start and end, exact seconds per view, and preparation, speech or action, reaction and settling within each view.
- Dialogue: literal words kept separate from concise visual MOTION and any supported VOICE DIRECTION control, with supplied or estimated speech duration.
- Lighting and color: motivated sources, readability, direction, palette and continuity.
- Sound and ending: intelligibility, ambience, timed effects, acoustic distance, music or silence, final image, cut points and sound bridges.
For example, a four-second wide view, a six-second medium view for a complete line and reaction, then a five-second close-up can be planned as a 15-second sequence. The instructions explain what occupies each interval and why it serves the scene. Requested seconds are planning targets, not guaranteed generated timings. Source duration, used interval, requested final duration and measured result remain distinct.
Four distinct routes, four different handoffs
- Avatar Shots: separate, self-contained source-clip prompts and assembly notes. The documented 5-, 10- or 15-second choices apply to this route only, and only when offered for the selected source/look.
- Single Scene Cinematic: one scene-specific cinematic prompt for the confirmed route and source duration. A v0.5 example uses a 12-second source with 5 + 7 second views; do not apply the Avatar Shots 5/10/15 list to this route. Account controls and eligible choices still require creator confirmation.
- Video Agent Chat: a project-level prompt with global identity/world/style and a structured plan for each scene, including visuals, VO/script, duration, performance and sound. Review the generated project plan before treating it as final.
- Presenter/Studio: standard speaking-avatar fields, with exact literal words in SCRIPT, concise visual performance in MOTION, and VOICE DIRECTION only when that control is actually available and confirmed. Do not paste a full cinematic prompt into these fields; camera, props and background may need another route or manual work.
Engine, look, source options, language support and other account-specific controls start unknown until the creator confirms them. A language choice is not proof of engine support: unlisted cinematic languages need a reviewed pilot or an approved alternate route. Keep supplied words and locked timings intact while resolving the route. Use only controls, eligible avatars and looks available in your own account.
Scene Brief Studio 1.2.0: plan, copy, review
The included offline HTML editor has 81 fields across eight sections, with editable camera-view rows:
Across the sections, the studio organizes the brief, route and source, requirements and evidence, avatar/language/world, camera views and timing, performance/dialogue/sound, assets and feasibility, plus prompt review and conversation. Review declared and observed controls separately from generated requests and manual steps; identify essential unknowns, unsupported guarantees and available, missing or optional assets. Inspect starting values and adjusted settings, lock view durations, check cumulative timing and conflicts, and set an attempt budget. The selected visible avatar remains one-avatar scope; creative and delivery settings are adjustable.
The studio assembles a message for the agent. Copy it into ChatGPT or Claude, paste the response back into the artifact, add feedback, then copy the next revision request. Copyable text remains a draft until the creator reviews the current field text and known requirements and explicitly marks it reviewed. Any edit to the brief, views or final field text invalidates that declaration; a declaration is not a quality certificate. A separate final-prompt field copies only the matching generation block for the selected destination.
It supports editable JSON exports and imports, Markdown request exports, explicitly recorded local conversation history and opt-in browser storage. Imports retain creative data but reset route/compatibility and final-review confirmations, so they must be renewed. The artifact does not send messages, call a model, upload media or connect to HeyGen. Browser or host previews may restrict rendering, clipboard, downloads or storage; a written brief entered directly in chat is the fallback.
A manual workflow, with the creator in control
- Describe the scene, its purpose, constraints, selected avatar and exact dialogue in ChatGPT or Claude, directly or with a Scene Brief Studio request. Name the intended route and actual controls where known.
- Classify requirements, resolve essential unknowns and assets, and review the matching plan and field text. Preserve exact words and locked timings; confirm supported language/voice controls and the intended destination.
- Declare a creator-set retry budget, inspect the cost displayed for the actual plan, and manually paste the reviewed text into the appropriate HeyGen field before initiating generation.
- Review the result yourself. Report a specific defect to the agent for repair instructions, then make changes in the website or editor yourself.
The agent prepares instructions; the creator reviews and operates HeyGen; HeyGen generates the video. There is no universal credit price, retry savings promise or guarantee that trimming can repair missing content. The agent does not connect to apps, access an account, submit prompts, inspect every generated frame, edit, mix audio or render.
What the complete package includes
- Portable agent instructions for a Project/chat route; full skill source and relative resources.
- One registered skill design with 18 workflow modules—not 18 separate AI models or employees. They cover intake, dramatic design, feasibility, requirements and assets, dialogue timing, characters, production design, blocking and coverage, cinematography, performance, sound, route/field mapping, prompt compilation, editing directions, review, revisions, timing, the creator artifact, and platform adaptations.
- A dedicated, correctly rooted native Claude Skill installer ZIP.
- Scene Brief Studio 1.2.0 standalone HTML artifact with 81 fields across eight sections; an 11-page setup, use and troubleshooting PDF; and a detailed Markdown user guide.
- Six fictional plan/prompt examples, including route-specific Avatar Shots, Single Scene Cinematic, Video Agent Chat and Presenter/Studio examples.
- Optional Python standard-library helpers for local planning checks and artifact generation, plus research material, evaluation records, release metadata and integrity hashes.
The complete archive contains 78 entries. Example controls and assets are hypothetical prerequisites, not account or footage verification. Research is background material; proposed or historical designs do not override the current one-avatar, manual-only instructions.
Requirements, privacy and boundaries
Install using the included inner native Claude Skill installer, if your Claude account or organization has native Skills upload and the required code-execution and file-creation capability. Otherwise use the portable Project/chat instructions. Platform mechanisms differ; installation is not guaranteed to be identical across hosts.
The portable route requires access to ChatGPT or Claude, the supplied instructions and reference material, and a suitable HeyGen workflow. Scene Brief Studio is optional; a written brief works directly in chat. Python is optional and needed only for local helper scripts. The complete workflow is not offline: ChatGPT or Claude receives what you share, and HeyGen receives the prompt you paste. No integration does not mean those services never process your content.
- No standalone AI model, hosted generation service, video generation, HeyGen or other app integration, login or account access, API-key use, automated submission, or bundled ChatGPT, Claude or HeyGen subscriptions or generation credits.
- One selected visible avatar throughout the supplied workflow; inserts may show no avatar. No avatar creation, cloning, or bundled avatar, footage or sound assets. You select an eligible avatar/look and assigned voice, and provide authorized references or other required assets as needed.
- No automated editing, audio mixing, rendering, footage retrieval or continuous monitoring. Identity, exact speech, lip-sync, frame-perfect shots, cuts, professional acting and complex physical action are requests—not guarantees.
- No guaranteed retention, conversion, audience emotion, educational outcome or advertising performance; no measured savings in credits, money, time or attempts. Better planning does not eliminate retries or generation cost.
- No casting management, budgets, scheduling, crew coordination, distribution system, team approvals, shared production database or automated client workflow.
You remain responsible for supplying requirements and exact words; confirming actual controls, language, voice and assets; reviewing prompts and the displayed plan/cost; setting an attempt budget; initiating generation; checking results and any manual finishing. Factual content and client or brand approvals need their own review. The product contains no licensed media and establishes no commercial licence, support arrangement or future-update terms.
Validation and known testing gaps
Version 0.5.0 is a documented, locally checked prompt-planning product. The supplied records report these package checks; they have not been independently verified by MiReVerse:
- 79 passed, 0 failed: local planning and renderer checks for arithmetic, route-specific duration handling, requirement/asset blockers, literal fields, migrations and text output.
- 84 passed, 0 failed: checks of the shipped artifact JavaScript in a lightweight DOM harness, including form events, draft/final-copy state, reviewed text, import migration and review invalidation.
- 12 inspected checks passed, 0 failed: checks on three fresh agent text responses: a confirmed 12-second Single Scene with 5 + 7 views, an impossible standard-route guarantee, and an unknown-engine Slovak brief.
Archive and installer structure/integrity are reported as checked; all 11 PDF pages received limited visual review. These results support local planning, routing and declarations only and do not establish account controls or end-to-end video quality. Native Claude execution, ChatGPT Custom GPT setup, live artifact browser behavior and permissions, authenticated HeyGen controls, and generated footage/audio/acting/lip-sync have not been verified. No footage pilot or audience study was performed. Agent response checks are not native-host, customer-trial or HeyGen generation tests; a lightweight DOM harness and limited document-render review are not real-browser verification.
Example: a confirmed 12-second single scene
A creator has confirmed a 12-second Single Scene Cinematic source and plans two perspectives of five and seven seconds with one selected avatar and exact dialogue. That route-specific plan is not the Avatar Shots 5/10/15 source-duration list. If the actual account route, language, engine, look or required asset is still unknown—or a requirement depends on an unsupported guarantee—the package keeps the work in draft rather than asserting readiness.
Once the creator confirms the matching route and reviews its current field text, they paste the prompt and generate manually. If a view is too short, words change, or an asset is absent, the creator identifies the defect and asks for a feasible repair; trimming is not presented as a way to invent missing content. This illustrates planning and handoff, not a verified HeyGen result.
Unspecified terms: The supplied description does not establish support arrangements, future updates or licence terms. These should not be assumed.