Evidence first ai video

AI Tech Talking-Head and Demo Video

Overview

Produce AI and technology talking-head, presenter-led, voiceover, product-demo, and explainer videos in which every shot proves a claim, explains an idea, or advances the story. Apply the Evidence-First method: readable evidence, continuous narration, and deliberate transitions—not a slideshow of cards, a stock-footage montage, or a presenter with filler placed on top.

Use this workflow with any compatible research, media, editing, or video-review tools. Treat prior failed edits as evidence about the process, not as assets that must be preserved.

Required reads and capability routing

• REQUIRED FOR EVERY BUILD OR REBUILD: Read references/director-playbook.md.
• REQUIRED BEFORE CHOOSING TOOLS: Read references/tool-routing.md, inventory available capabilities, and select the strongest honest route.
• REQUIRED FOR WEB RESEARCH: Use an available web or research capability and verify time-sensitive claims with primary sources. If none exists, restrict work to supplied sources and mark unsupported claims blocked.
• REQUIRED FOR MEDIA ACQUISITION: Use an available browser, recorder, downloader, or media capability, then apply the stricter rules in references/sourcing.md. If none exists, deliver a source plan instead of claiming files were acquired.
• REQUIRED FOR VIDEO AUTHORING: Use an available code-driven renderer, editor, NLE, or manual handoff route. If none exists, create a dry-run production package and do not claim a preview or render exists.
• REQUIRED FOR REVISION OR QA: Read references/failure-modes.md and references/qc.md.
• REQUIRED WHEN A PRESENTER APPEARS: Read references/presenter-editing.md.
• REQUIRED FOR SCRIPTING: Read references/script-edit.md.
• REQUIRED FOR EXISTING-VIDEO QA: Use available video-understanding capability when it can inspect the render, then verify manually. Without one, complete the manual passes and mark automated semantic review unavailable.
• CONDITIONAL: Use a visualization capability only when a comparison, flow, timeline, architecture, or data relationship becomes clearer as a graphic.
• CONDITIONAL: Use HeyGen for an approved voice or visible presenter. Check authentication, quota, cost, and resolution before generation; never spend or enable auto-recharge without approval.

Tool names are examples, not requirements. HyperFrames, deep-research, media-use, qwen-video-understanding, Visualize, and HeyGen may improve a route when installed, but the delivery claim must follow the capability actually available. Read references/case-study-agent-explainer.md only as an optional benchmark from one production stack.

Evidence decision ladder

For each spoken claim, choose the first available option:

1. official or first-party moving demo;
2. permitted self-recording of the real interface or reproducible result;
3. original demonstration that visibly proves the idea;
4. concise sourced diagram, data card, or quotation;
5. rewrite or remove the claim.

Never fabricate a substitute product UI when a real interface or demo exists. Never use generic technology stock, an unrelated screen recording, or a static fake workflow to fill an evidence gap.

Production workflow

1. Lock the promise. State the audience, useful takeaway, platform, aspect ratio, duration, tone, and call to action. For beginners, promise a concrete example or outcome rather than a theory lesson.
2. Research before writing. Build a claim ledger with source URLs, dates, product status, limitations, and exact demo timecodes.
3. Write spoken language. Draft narration, title, cover copy, editorial labels, and source labels together. Remove jargon dumps, unsupported hype, and repeated AI-shaped sentence templates.
4. Generate or record a single narration master. It must be one continuous performance; do not assemble the final voice from sentence clips. Transcribe the returned audio and derive all timing from it.
5. Build a sentence-to-shot map. Copy assets/asset-manifest.csv and assets/edit-map.csv. Give every claim one primary source, one complete visible action, protected regions, presenter decision, overlay exit, and sound decision.
6. Acquire and normalize originals. Preserve source files. Create claim-specific derived clips with complete setup → action → result → short settle. Never cut while the meaningful action is unfinished.
7. Design static end states first. Make each shot readable without motion. Keep one focal point. Use real demos at a large scale; if forced reframing destroys legibility, change the canvas or layout instead of cropping harder.
8. Add purposeful motion. Animate only entries, exits, emphasis, and visible state changes. Avoid constant zoom, drift, wobble, subpixel movement, and transitions used to hide bad cuts.
9. Add presenter continuity. Use full presenter for human beats and moving PIP only where a safe corner exists. A frozen portrait is not a speaking presenter. Cover every media boundary deliberately.
10. Mix after picture timing is stable. Keep narration continuous. Use event-driven, restrained clicks, pops, swipes, and confirmations only when the picture contains the matching event. Never cut or duck speech to make room for SFX.
11. Preview, then render. Run the available authoring checks, inspect boundary triplets, and open the complete preview. If the user asked to review before export, obtain preview approval before the high-quality render.
12. Verify the final file. Watch once silently, once audio-only, and once normally. A successful render or lint pass is not completion.

If a required capability is missing, stop at the strongest truthful stage and produce a blocked report or dry-run package. Never describe planned, simulated, or uninspected media as generated, previewed, quality-checked, or ready to publish.

Non-negotiable gates

• The visual must match the exact narration at that moment; semantic mismatch blocks export.
• A source range may not be reused for an unrelated claim. Reuse requires an explicit before/after or callback reason.
• Motion is shown as video when motion is the evidence. Do not replace a real demonstration with a still.
• The visible action must finish before the cut and hold long enough for the result to register.
• Reject generic stock, fabricated substitute UI, repeated filler, blurry enlargement, destructive crop, crooked capture, shake, drift, random zoom, and flash frames.
• Reject dense card walls, fake-premium black text screens, simple box-and-line diagrams used as decoration, and labels that repeat visible UI.
• Use direct two-line editorial labels: white category plus one colored keyword, integrated into the footage without a solid badge. Clear time-bound overlays at least 0.04 seconds before the next media boundary.
• Never expose the full presenter for less than 1.2 seconds accidentally. PIP must be moving speaking footage, use an audited safe anchor, and disappear when no corner is safe.
• Keep one continuous narration master. Reject internal hard cuts, unexplained pauses, low dialogue, clipping, explosive gain changes, harsh transition spam, stale captions, and sync errors.
• Never claim completion until the actual preview or final MP4 has been watched with sound.

Deterministic audits

Copy the templates in assets/ into the project. Then run:

python scripts/validate_manifest.py path/to/asset-manifest.csv
python scripts/audit_edit_map.py path/to/edit-map.csv

The edit-map audit blocks timeline gaps, accidental presenter flashes, frozen PIP, incomplete actions, unexplained source-range reuse, and overlays that survive into the next media boundary.

Project outputs

project/
  BRIEF.md
  SCRIPT.md
  research/claims.md
  sources/source-ledger.md
  docs/director/shot-plan.md
  docs/director/edit-map.csv
  asset-manifest.csv
  assets/original/
  assets/official/
  assets/recordings/
  assets/derived/
  audio/narration-master.wav
  audio/final-mix.wav
  captions/
  compositions/
  renders/
  qc/

Cleanup

Keep the approved render, editable project, source ledger, manifest, narration, captions, licenses, and irreplaceable originals. Delete only confirmed regenerable files inside the project root. Move uncertain files to a dated review folder instead of deleting them.

Examples

• Read references/example.md for a compact feature-video example.
• Read references/case-study-agent-explainer.md for the full benchmark, including failed approaches, corrective recuts, final audio design, and acceptance evidence.