Reel Animated Explainer

Turns a vertical talking-head video (1080x1920) into a polished reel. Two modes, picked from the video itself:

Step 0: detect, then ask once

1. 1Detect the mode yourself: ffprobe the video, grab 4 to 6 frames and look. Mean brightness below 12 across the top rows means split panel; otherwise full screen. Confirm in one line ("Full-screen video, so I'll use the full-screen mode. OK?") instead of asking from scratch.
2. 2Ask the rest in one message (skip anything already answered, never assume a series name or episode number): - Series name and episode number/topic for the on-screen label and caption. - Transcript: exact text, ideally with rough timestamps (an SRT is perfect). It is used word for word, and it saves about 30 minutes of timing work on mixed-language speech. - Caption language: as spoken (e.g. Hinglish in Roman script) or translated. - Last reel's theme, so this one differs, and any colours to avoid. - Extras: logo files for the technologies shown (never redraw a logo), sound effects, music, screenshots. - Deliverables: Instagram caption and hashtags? File size limit?
3. 3After the answers, work through without further check-ins unless something blocks you.

Hard rules (both modes)

1. 1Transcripts are sacred: only split into lines, never reword.
2. 2Never leave a single word alone on a line in titles, cards or captions: one line, or balanced lines with at least 2 words each.
3. 3Rotate themes and respect banned colours.
4. 4Brand icons: use the official logo file as is; until it arrives, a plain text badge.
5. 5No fabricated screenshots of real posts; use real screenshots or clearly designed quote cards.
6. 6Preview first: a grid of frames from every scene before the full render, and frames checked from the final export.
7. 7Background removal only in full-screen mode, with a good matte (see the guide). In split-panel mode, keep the creator's real background.

Shared steps

• Word or line timing: use the transcript's timestamps when given. Otherwise transcribe (whisper.cpp or faster-whisper, short windows with no_context), and for mixed-language speech anchor on clearly heard words and spread the rest by voiced time (scripts/full_screen/align_words.py). Tell the user which lines are uncertain.
• Plan scenes as a timeline table (t0, t1, scene, visuals), each starting within 0.5 s of when the speaker starts that point.
• Render: a master at CRF 18, then a two-pass H.264 copy at about 1750k video / 128k audio, which keeps a ~2 min reel under 30 MB.
• Deliver: the video, a one-line summary of what changed, uncertain caption lines, and, if asked, an Instagram caption: "Episode N: How <Topic> works", one hook line, one comment prompt, 5 to 6 hashtags.

Before delivering

• Mode matches the video; theme differs from the last reel; no banned colours
• No black gaps (split) or matte flicker and halos (full screen)
• Every scene starts when the speaker starts that point
• Captions exact, readable, never over the face, no single-word lines
• No text overflows, or hides behind the speaker when it must be read
• Frames checked from the final exported file