Video Podcast Maker Nano

Video Podcast Maker Nano

A 7-step pipeline for personal use: research ∥ materials → script → TTS → audio checkpoint → video → preview checkpoint → 4K render ∥ publish kit. No bundled scripts, no hardcoded backends, no templates. The skill owns the logic and runs autonomously by default; the TTS backend and video tool are chosen per video (see Tool selection).

What makes this work regardless of tool choice — the three invariants:

1. Checkpoints are never skipped — but who checks is policy. Three checkpoints exist: script (after Step 2), audio (after Step 3), preview (before render). Default checker is the agent itself (self-verification, defined per checkpoint); a project's AGENTS.md/CLAUDE.md may upgrade any checkpoint to a human gate — see Oversight policy.
2. Audio is the master clock. The final video's duration must match the narration audio within ±0.5s (ffprobe both). Visuals are cut to the audio, never the reverse.
3. A script change invalidates everything downstream. Edit the script → re-run the script checkpoint, then TTS → audio checkpoint → visuals → preview checkpoint → render. Never hand-patch timings.

Oversight policy

Default mode is autonomous: the pipeline runs end to end, with the agent performing every checkpoint's self-verification. A working project's AGENTS.md/CLAUDE.md can override per checkpoint with one line each:

• Checkpoint 1 (script): human — halt after Step 2 until the user approves the script. Worth it: a late script change costs a full re-run.
• Checkpoint 2 (audio): human — the user listens to the full audio before visuals. Worth it when TTS misreadings are costly to catch later.
• Checkpoint 3 (preview): human — the user reviews the draft before render. Worth it for style-sensitive channels.

Unmentioned checkpoints stay agent-verified. Project policy files start from AGENTS.template.md (next to this SKILL.md): copy it into the video project root as AGENTS.md (plus a CLAUDE.md copy for Claude Code) and fill in the tool bindings. The skill never uploads or publishes anything anywhere — the publish kit is files on disk; publishing is always a human act outside this pipeline.

Tool selection (per video)

Decide the TTS backend and the video tool once, before Step 3, in this priority order:

1. Project bindings win. A filled AGENTS.md/CLAUDE.md in the working project (see Oversight policy) IS the user's standing specification — use it, no scanning, no second-guessing. If the user names a tool in-session, that overrides the file. If the policy file is missing or still contains template placeholders (/absolute/path/to/...), resolve the bindings via rules 2–3, then write the resolved values back into the project policy file so the next run starts bound.
2. Auto-detect installed skills. Scan the session's available skills for anything that can do the job — video: skills wrapping an authoring tool (Remotion, HyperFrames, CapCut, ...); TTS: any skill wrapping a TTS engine. Exclude end-to-end pipeline skills (e.g. other video-podcast-maker variants, if installed): they are pipelines like this one, not backends — invoking them here would nest workflows and double the checkpoints. One fit → use it; several → pick the best match for the project's language and output needs. Record the choice in research.md; surface it in the final summary. Never block waiting for a tool confirmation.
3. Nothing found. Fall back to plain CLIs the user already has (ffmpeg + any TTS CLI) and say so — never install a new tool unprompted. If a TTS CLI cannot emit subtitle timing, derive cues by splitting the script evenly across the audio duration and flag the approximation at Checkpoint 2. If a user-specified tool is missing or cannot meet the capability floor, say so and drop to rule 2 instead of improvising.

Capability floor (applies to rules 1–3): the TTS choice must produce narration audio plus subtitle timing; the video choice must export a draft video file (for Checkpoint 3 verification) and 4K. Live preview is a bonus, never a substitute for the draft export. Auxiliary jobs need no selection: research uses the built-in web search, stills/cover the video tool's still export or any image tool, duration checks ffprobe. If ffmpeg/ffprobe are absent, report it and stop — the ±0.5s invariant is non-negotiable and is not skipped to keep a run alive.

Working layout

All artifacts for one video live in videos/{name}/ ({name} = lowercase English, hyphen-separated). File names for audio/timing adapt to the chosen backend; the set is what matters:

videos/{name}/
├── research.md          # Step 1 — facts + sources
├── podcast.txt          # Step 2 — narration script
├── podcast_audio.wav    # Step 3 — narration audio   (name per backend)
├── podcast_audio.srt    # Step 3 — subtitle timing  (or the backend's equivalent)
├── assets/              # Step 1 — images/BGM + sources.md (source + license per asset)
├── video-project/       # Step 5 — whatever the video tool produces
├── final_4k.mp4         # Step 7 — 3840×2160 render
├── cover.png            # Step 7 — video cover
└── publish_info.md      # Step 7 — title / description / tags / chapters

Workflow

Entry point check (before Step 1). Look for videos/{name}/ at the project root (or the directory the user names). If it already contains artifacts from an earlier session, this is an iteration — resume from the earliest step affected (see Iterating), do not re-run Step 1. Only start at Step 1 when no artifacts exist.

Step 1 — Research ∥ collect materials

Two parallel outputs from one investigation pass (facts and assets come from the same sources):

Research → research.md. Investigate the topic (web search, papers, the user's pointers). Distill into research.md: facts, numbers, and their sources. Every number the script will claim must trace back here — a precise number without a source is fabricated; drop it or attribute it.

Materials → assets/. Collect everything the visuals and cover will consume into videos/{name}/assets/: per-section images/illustrations/screenshots, brand logos, BGM. Two sources, in priority order:

1. User-provided — files the user hands over or points at; copy into assets/, never reference them in place.
2. Auto-collect — official material first (product banner, spec card, screenshot), else free/licensed sets (unDraw SVG, Pixabay/Pexels, OpenMoji / Microsoft Fluent Emoji / Google Noto Emoji, @lobehub/icons for brand logos).

Record each asset's source URL + license in assets/sources.md at collection time (attribution-required sets must be credited in the video description at Step 7). No suitable asset exists for a section? Do not fabricate a screenshot — fall back to a text-only layout for that section (or a generic free-license illustration), record the decision in assets/sources.md, and move on; asking the user is optional in human mode. Missing assets can be added any time before Step 5.

Step 2 — Script → podcast.txt, then Checkpoint 1

Spoken text only, no markdown. Split the script into segments with [SECTION:xxx|display-label] markers (lowercase English names, e.g. [SECTION:hero|intro]) — one section per video segment. This marker convention is the portable contract between script, TTS chunking, and visual layout; any TTS/video tool can consume it. Markers are structural metadata: never spoken, never shown as subtitle text — TTS and subtitles consume the text between markers only.

Style rules (language-agnostic): see Script style.

Checkpoint 1 — script self-review (mandatory). Verify the script against every Script style rule, then reconcile its sections against assets/: list sections with no matching asset, collect or fall back per the no-asset rule above, and note what goes without. In human mode (project policy), halt instead and hand the script over — do NOT run TTS until the user explicitly approves. Audio, timings, and visual entrances all derive from the script; a late script change costs a full re-run.