storyreel

Fully automated multi-channel story pipeline: an upvoted subreddit story is retold in the channel's own words (critic-gated, publish-or-nothing), edited over real footage using OpenMontage's documentary-montage tools, and published to YouTube. Runs entirely on GitHub Actions with free services.

How a video is made

1. Story — stories.py harvests full-text self-posts from each channel's subreddits; Reddit's own votes (score + 2.5 × comments) are the quality filter, a safety list blocks stories that aren't ours to tell, and every used post id in posted.json is excluded forever.
2. Script — an LLM retells the story in the channel's voice (never verbatim, names invented, source credited in the description). A separate critic call scores hook/arc/faithfulness/own-words/pacing/landing and rejects until the draft earns a pass. Nothing that fails review is published.
3. Footage — montage.py drives OpenMontage headlessly: an LLM splits the approved narration into 6–10 visual slots; corpus_builder downloads and CLIP-indexes candidates from Pexels/Pixabay (free keys) and Archive.org/ Wikimedia/NASA (keyless); clip_search ranks the corpus per slot; clips are fitted to slot length and stitched with video_stitch.
4. Finish — edge-tts narration (per-channel brand voice), sentence-timed burned captions, ffprobe sanity checks, YouTube upload with the synthetic-media disclosure, state committed back to the repo.

OpenMontage itself is agent-first and interactive; this repo deliberately hard-codes one of its recipes so it can run unattended. The LLM chain is mpt's: NVIDIA NIM → Groq → OpenRouter free tiers, with a repo-local Ollama (llama3.1:8b, cached across runs) as the last-resort fallback.

Setup (~30 min)

1. Free keys: NVIDIA NIM (build.nvidia.com), Pexels, Pixabay, a Reddit "script" app (reddit.com/prefs/apps — recommended, it's the only source with live votes).
2. One YouTube channel per entry in channels.json. Google Cloud project → enable YouTube Data API v3 → Desktop OAuth client. Per channel, run python get_youtube_token.py, sign in as that channel, save the token as the YT_REFRESH_TOKEN_<CHANNELID> secret.
3. Add every key from .env.example as an Actions secret. Done — it runs at 08:00/20:00 UTC and commits posted.json back after each upload.

Test locally or via Run workflow with dry_run checked; the rendered mp4 is attached as an artifact instead of uploaded.

Channels

Edit channels.json — each entry is a full brand: voice, orientation (portrait/landscape), footage sources, subreddits, retell prompt, critic rubric, caption style, hashtags. Add a channel = add an entry + its YT_REFRESH_TOKEN_<ID> secret + one line in the workflow's env block.

Notes and honest limits

• First run is slow: torch CPU wheel + CLIP weights (~350 MB) + Ollama model download. All cached; later runs skip them.
• The clip corpus persists via Actions cache, so retrieval improves over time.
• OpenMontage is checked out at ref: main; pin a commit SHA after your first verified run.
• OpenMontage is AGPL-3.0 — fine to run for your own channels; don't offer this pipeline as a closed hosted service.
• YouTube may limit monetization of mass-produced content. The critic gate, real-footage edits, and per-story source credits are the mitigations, not a guarantee. Start at 1–2 uploads/day per channel.

Trend research and TikTok

Before falling back to plain top-voted stories, each run researches what's viral: viral.py searches YouTube keylessly (yt-dlp, relevance + freshness) for the channel's viral_queries, keeps videos above viral_min_views, and has the LLM read the theme — the underlying story shape, never the title or transcript. That theme is then matched against our own subreddit harvest, and the best-fitting real story is retold in our brand and voice. We recreate themes and formats, never content — no transcript is fetched, nothing is paraphrased from another creator, and the script passes the same critic gate. Used trend video ids are recorded in posted.json so a trend is only ridden once per channel. When nothing fresh matches (or research fails), selection falls back to the subreddit-votes path.

TikTok posting goes through Buffer (buffer.py, ported from mpt): Buffer's TikTok-approved app publishes publicly with caption + hashtags, which an unaudited first-party app cannot. Set BUFFER_ACCESS_TOKEN and per-channel BUFFER_CHANNEL_ID_<ID>; the rendered mp4 is hosted as a rolling GitHub release asset for Buffer to fetch (public repo required). No token = TikTok is skipped, YouTube unaffected.