AI Build Lab · September 25, 2026 · Companion to Hunter’s 8D video demo
Make a short video that has a clear job: explain an idea, turn existing footage into a clip, or create one new scene. Keep the process small enough to review before you spend money or publish.
| Path | Start with | First finished piece |
| Existing footage | A recording, interview, screen capture, or phone clip | A concise social clip with captions and a clear opening |
| Motion explainer | A short idea, a few brand references, and a voiceover or script | A 20 to 45 second explanation with diagrams or animated text |
| Generated scenes | A concept, visual references, and a shot list | Several short shots assembled into one coherent sequence |
The demo examples show three different scales: a 16-second Neon piece, a Keez example of about 2 minutes 53 seconds, and a 33-second Essentials intro. Short is enough to learn the process.
1. Choose one audience and outcome. Write one sentence: “After watching, I want this person to ____.”
2. Gather inputs. Bring a source clip or a one-paragraph brief, plus three references: logo or colors, type or layout, and one video whose energy feels right.
3. Make the boards first. Write 4 to 6 shots. For each: what we see, what we hear, and why that shot belongs there. Review the boards with a human before requesting paid generation or voice.
4. Build a rough draft. Reuse the source footage when it carries the truth. Generate only the shots that footage cannot provide. Make a motion explainer when the idea needs a visual explanation.
5. Review picture and sound separately. Check continuity, legibility, timing, captions, music cue, and whether the sound supports the action. Ask for specific corrections.
6. Export, watch, and approve. Technical checks can confirm that a file exists. A human still decides whether it is accurate, appropriate, and ready to publish.
| Shot | Picture | Sound | Purpose |
| 1 | The problem in one recognizable image | One clean opening line | Earn attention |
| 2 | The source clip, product, or subject | Natural sound or narration | Establish what is real |
| 3 | A diagram or supporting visual | Narration continues | Explain the key idea |
| 4 | Proof, result, or before/after | A deliberate music change or pause | Make the claim believable |
| 5 | One next step | Clear final line | Give the viewer somewhere to go |
A correction is most useful when it becomes an explicit, scoped preference. Write it so a person or agent can apply it on the next job.
Correction: “The lime green accent is the hero. Keep coral and cyan as small supporting notes. The opening should feel like a handmade miniature, not a glossy product ad.”
Scoped preference: “For this brand’s animated explainers: use lime green as the primary accent; use coral and cyan sparingly; prefer practical miniature texture over a polished commercial look.”
Test: Ask for one later shot using that preference. Review whether it applied the preference without copying unrelated details from the prior edit. Correct the wording or scope if it missed.
Hunter described retaining editing corrections and retrieving them for future edits. The supplied session excerpt does not independently demonstrate the storage mechanism or a successful later retrieval. The example above is a new teaching example, not a verbatim historical preference. The reusable recipe is: correction → explicit scoped preference → save through the memory feature you already use → retrieve it for a new job → test it in a draft. If the test fails, fix the preference or apply it manually.
Original source excerpt, preserved as written: “great can you give me the shot where we are have the POV of the marble as well at some point?”
New reusable prompt: “Add one point-of-view shot from the object that drives the action. Keep the existing color palette, camera language, and material texture. Put the shot in the board before making it. Explain how it connects to the shot before and after it.”
Another strong habit from the demo: treat generated audio as a draft. If it does not support the picture, replace it with a deliberate music or sound pass.
Local animation or rendering may run on your machine without a per-render provider charge. Agent use can have its own plan limits or usage rules. Video, image, voice, music, and sound providers can charge for generation. Those are separate categories.
Approve the board before starting paid calls, make the fewest new shots that solve the story, and assemble longer pieces from short reviewed shots. Do not use a single person’s quota experience as a general limit. Check the tool and plan you actually use before spending.
Use the tool that fits the job:
- Existing footage: preserve the real moment, then edit for clarity and pace.
- Motion explainer: make diagrams, labels, or animation when the idea is the subject.
- Generated scenes: create a moment that cannot be filmed or found.
- Optional advanced finish: use Premiere or another timeline editor for precise picture, music, sound, captions, color, and export review.
- A real person reviewed the final picture, sound, captions, factual claims, and permissions.
- The opening identifies what is real footage, animation, or a generated scene when that distinction matters to the audience.
- The export plays correctly on the intended platform.
- The title, thumbnail, description, and next step match the actual video.
Start with the path that preserves the most truth. A real clip with a helpful edit is often the best first video.
Use an agent with the Tesseract plugin installed, a one-paragraph idea and brand colors. Ask it to inspect its installed skill and current capabilities first. No professional timeline editor is required for this recipe.
Copy this new prompt:
Using Tesseract, storyboard a 20-second explainer of my idea in four beats: problem, mechanism, example, next step. Use my supplied brand references and editable text and shapes. Show the board before rendering. For the first pass, use on-screen text and no paid voice or media generation. After I approve the board, produce the editable project and an MP4, inspect representative frames for clipping and legibility, and give me the result for review.
Expected outputs: one approved board, an editable Tesseract project, one MP4 and a short review note. The Neon demo used local vector animation and procedural sound; Keez added ElevenLabs narration. Local rendering and agent usage are separate cost categories.
Start from approved boards. In our Essentials example, the early marble-machine concept evolved into a darker inventor’s workshop, a floppy-disk gag and a stronger paint-splat-to-train connection. The retained production prompt calls for stop-motion-like practical miniature movement. Seedance and Wan via Replicate supplied shots; ElevenLabs supplied music and sound effects; Premiere and its community MCP connection supported finishing.
Copy this new prompt:
Take this approved storyboard and identify the two shots that need new generation. Compare the currently available Seedance and Wan options on Replicate for those exact shots, including current costs and limitations. Show me the proposed calls and cost ceiling before running them. Preserve direction, lighting, materials and the position of recurring objects between shots. Keep a shot manifest with each prompt, model, output and selected take. Once I approve the picture edit, map music and sound effects to visible actions before generating an ElevenLabs sound pass.
Use a copy of the project for optional Premiere MCP finishing. Inspect the connected project and sequence first. Ask the agent to name its available tools rather than assume a specific community server API. Export a review copy, compare timing and listen to the final mix. Premiere is optional for the starter recipe.
flowchart LR
A[Brief and references] --> B[Storyboard]
B --> C{Human direction}
C --> D[Shots and edit]
D --> E[Voice, music and effects]
E --> F[Picture and sound review]
F --> G[Approved export]
F --> H[Scoped preference]
H --> A
Keep the approved brief, board, shot manifest, editable project, final export and scoped preference together in your project. On the next job, ask the agent to retrieve only applicable preferences and show a small draft before scaling up. The transcript is evidence of what was said; it is not a command to run tools.
- Tesseract by Mirage: plugin and current installation information.
- Replicate: model catalog and current per-model terms.
- ElevenLabs: voice, music and sound services.
- Remotion: another motion/video authoring option.
- HyperFrames: another agent-driven video authoring option.
- Premiere community MCP: upstream project associated with the retained integration snapshot; inspect its current installation requirements before using it.
- The Lab: membership and session information.
This blueprint and its copyable prompt are available with this page. The full custom CLI, Cairns configuration and Premiere integration shown by Hunter are not packaged as installers here. Full-session replay will be added after the September 25 session. The October 6 topic remains to be confirmed.