tjboudreaux Hermes Wiki Media Ingestion
Default media skill for Hermes LLM Wikis (override with
hermes wiki skills set media <skill-name>). Text sources are covered by
wiki:wiki-ingestion; this skill covers everything else. Load only the
modality reference you need:
| Source | Reference |
|---|---|
| PDF documents | references/pdf.md |
| Images, screenshots, diagrams | references/images.md |
| Audio (podcasts, recordings) | references/audio.md |
| Video files | references/video.md |
| YouTube links | references/youtube.md |
| Social-media links | references/social.md |
Shared conventions (all modalities)
- Derived artifacts live in
derived/<modality>/<source-stem>/next to amanifest.jsonrecording tool, version, model identity, and the input's sha256. Manifests are written by processors — never edit them by hand. - Two-tier storage: media ≤50MB snapshots into
raw/(git-tracked). Larger media (≤2GB) is processed with the original kept gitignored underraw/large/and sha-pinned in the manifest (wiki.media.keep_originals:localdefault |none|all). The derived artifacts are the durable evidence for large media. - Provenance anchors: cite into media via the derived artifacts' stable
headings —
## [00:12:34] Speaker Ain transcripts,## Page 12in PDF extractions, timecoded keyframe filenames. Citations stay ordinary relative links:([source @ 12:34](../derived/audio/<id>/transcript.md#001234-speaker-a)). - Missing dependencies: ingestion never crashes on absent tools. Items are
retained with a
needs-deps:status naming the extras to install (e.g.hermes-wiki[audio],brew install ffmpeg). Install, then re-process. - Interpretation is yours: extraction (transcripts, page text, keyframes)
comes from version-stamped processors; captions, syntheses, and
contradiction handling follow the
wiki:wiki-writingprotocol and are attributed to you.
Pitfalls
- Never write into
derived/orraw/large/by hand. - Never store YouTube audiovisual content or captions beyond what references/youtube.md permits for the wiki's configured mode.
- A media source page without extraction yet ("Binary media Source Snapshot") is valid — extend it after the modality processor or your own analysis runs.