Video to knowledge
SkillDocs & knowledgeUse by2kb to turn a Bilibili/YouTube URL or a downloaded local audio/video attachment into a transcript, a short abstract, and long-form study notes.
Available today. Use it from your connected AI after setup.
No other account needed.
Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
Then ask your AI: use the Video to knowledge skill
What this skill tells your AI
The instructions your AI receives, as published by charlesmpc/by2kb in by2kb/integrations/hermes/skills/video-to-knowledge/SKILL.md and read by ahel’s review.
The installed Hermes plugin handles a bare Bilibili or YouTube URL automatically. For a media attachment, first save the attachment to a private temporary path, then pass that exact path as the source argument. The workflow runs the deterministic transcription stage first, then uses the Hermes host model for bounded, program-planned enrichment operations. Do not clone or inspect the by2kb repository.
For an explicit/manual request:
- Check installation with
by2kb version(0.5.3+). If missing or setup is incomplete, load the bundledskill_view("by2kb:install-by2kb"). The default installation ispipx install by2kb[asr-whisper,youtube]withby2kb init --preset agent-local. - Use cloud Doubao ASR only when the user explicitly selects it. Never ask the user to send TOS or ASR credentials through an IM conversation.
- Run
by2kb ingest <URL_OR_LOCAL_PATH> --enricher external_agent --json. - Run
by2kb enrichment next <JOB_ID> --provider <HOST_PROVIDER> --model <HOST_MODEL> --json. - With by2kb 0.5.3+,
needs_inputreturns a small operation ticket, not prompt text. Read bothsystem_promptanduser_promptusingby2kb enrichment read --request-file <REQUEST_FILE> --field <FIELD> --offset 0 --limit 1000 --json. Continue usingnext_offsetuntileoffor each field. Check every page'soperation_idagainst the ticket. Follow the complete prompts, save the bounded response as UTF-8 (title operations request JSON, others Markdown), and runby2kb enrichment submit <JOB_ID> --operation-id <ID> --output-file <PATH> --provider <HOST_PROVIDER> --model <HOST_MODEL> --json. - Repeat steps 4–5 until
nextreturns statuscompleted. - Report the three knowledge-base paths: raw transcript, short abstract, and long-form study notes.
Never print an entire request file or transcript into terminal output. Do not chain
claim and next; staged callers do not need claim. The program plans chunks and
reductions; do not invent operation IDs or summarize unseen/truncated input. If a
page is truncated, reread that offset with a smaller limit before continuing.
Inspect submit.status, not just its exit code. On operation_mismatch, query
status and next, discard the old output, and generate for the new operation; stop
after three resynchronizations. Never merely replace an ID on an old answer.
On conflicting content or other errors, check job status and report the step issue
without marking the shared job failed. Completed jobs should be reported as success.
Never place model credentials in by2kb when using this path. Preserve checkpoints
and user configuration. Older CLI inline prompts are compatibility-only; upgrade
if its tool output is truncated.
Signals
- GitHub stars
- 25
- Forks
- 2
- Last commit
- Sep 2026
Advanced
- Item type
- skill
- Key
video-to-knowledge- Source
- github.com/charlesmpc/by2kb