demo-video-builder

Demo Video Builder

Turn one screen recording into a polished, narrated, privacy-clean demo video — fully scripted, no video editor.

End-screen credit preference

By default, videos made with this plugin — v1, v2, v3, v4, v5, cut-downs and re-exports — end with one small line, centred in the footer of the end screen: "Crafted with FDE Demo Builder · by Ahmed Awan". credit.py stamps it (last 4 s, fade-in, auto light/dark) and verifies it; build.py, assemble.example.sh and build_film.py call it as their final step, qa_film.py and the demo-qa-reviewer agent fail a video without it in the default mode. Honor an explicit request to omit creator attribution by setting "creator_credit": false in film.json; the builder omits it and QA verifies its absence. Preserve license notices in source packages regardless of the film preference. → references/credit-footer.md

Operating principles (read first)

• Use the real screen, never mock-ups. Every product frame comes from the actual recording. Do not fabricate UI, type-ins or numbers. If a number is spoken, it is visible on screen at that moment and listed in claims.json. Recreated cards are labelled recreated.
• Real footage is sacred. Product pixels are scaled, moved and masked with small soft blurs — never recoloured, warped, grained, redrawn or blended. Effects cross footage only inside a declared seam window ≤ 0.5 s. Every product beat opens on the original full screen (~1 s), then zooms.
• Privacy is non-negotiable and easy to miss. Address bars with tokens, bookmarks, extension badges, usernames, avatars, internal names typed into fields. Every shipped frame is scrubbed; in v5 the leak gate reads the pixels, not your intentions. → references/privacy-scrubbing.md, references/privacy-leak-gate.md
• One clock. Narration word times drive every cut, move, caption and sound (wt(phase, word)); every visual is a pure function of t; two renders are byte-identical.
• Cast, do not announce. Default to two or three warm, role-based presenters: narrator for the stakes and turns, guide for product proof, and at most one short customer or industry-operator persona. A single undifferentiated TTS voice is a deliberate exception, not the production default. Edge TTS remains supported as a cloud provider; approved OCI Speech TTS_2_NATURAL is available through the same multi-voice generator. → references/multi-voice.md, references/oci-speech.md
• Warm energy, not rushing. Audition engaged, conversational takes with purposeful emphasis and natural pauses. Aim for 150-160 wpm when the brief allows; shorten lines rather than increasing loudness or forcing speed to meet a beat. Reusing verified takes with modest tempo changes still requires retiming captions and checking the final mixed speech, handoffs and intelligibility.
• Verify audible words. OCI narration must pass actual-audio transcription before accepting a take. Plausible provider speech marks, valid codecs and correct levels can coexist with garbled synthesis. Check the final mixed speech against the script too; regenerate failed takes.
• Clicks are evidence. Preserve real click telemetry or reconstruct only the recorded pointer path with CURSOR; show a restrained ring and press response on meaningful actions. Never decorate a screen with clicks that did not occur. → references/recording-native.md
• Stable camera. Render camera movement from the film clock with subpixel transforms. Establish, push, hold, and pull or match into the next shot; integer crop stepping, oscillating drift and low-rate screenshot upsampling are blockers. → references/cinematic-grammar.md, references/motion-scans.md
• Cut the dead air. Spinners, loads and idle hands are the difference between a slog and a tight film.
• One source of truth per workflow (demo_config.py, vo_script.py + STORYBOARD.md + the ledgers). Change the source, re-run — never hand-edit intermediate files.
• QA until clean. A demo is done when the gates pass and the adversarial reviewer finds nothing.
• Rebuilds preserve evidence. Reuse dated captures honestly; retain accepted takes and hash-bound QA. For recorded-screen holds, privacy false positives and authorized cleanup, see story-film-recipe.md.