← back
LLM Knowledge Bases: a practical guide — Ben Holmes, Warp
Takeaway
Capture ideas freely, then use repeatable enrichment skills and controlled metadata to turn scattered notes into navigable knowledge.
Summary
- A practical knowledge-base workflow starts with raw Markdown notes, then adds tags, source links, backlinks, topic wikis, and visualizations.
- Voice dictation lowers the friction of capturing research and reflections; the speaker demonstrates local tools such as Handy and VoiceInk.
- An enrich-note skill records enrichment timestamps, looks up sources, and finds related notes through file searches.
- A shared tag reference discourages agents from inventing inconsistent categories, while Markdown viewers such as Hubble or Obsidian keep notes inspectable.
knowledge-baseagent-skillsmarkdown
Original description
Talking runs about 200 words a minute and typing does not, which is why Ben Holmes starts a knowledge base with voice dictation rather than with organization. The pipeline only works if there is enough raw material, so the capture step is deliberately sloppy: ramble into your phone after a podcast or a meeting, skip the formatting, and let agents impose structure afterward. He points at Handy and Voice Ink as local, on device options, so none of it needs a subscription. Structure then arrives in passes. An enrich note skill stamps each file with a timestamp so later runs know what has been processed, assigns tags from a fixed reference list instead of inventing new ones, since as he puts it Claude loves to get creative, researches the source on the web, and adds backlinks found by key term search. On top of that pile, a wiki generated from a Karpathy gist produces browsable entries for the people, concepts, organizations, and sources buried in your own notes. Then the whole thing runs on a schedule in a cloud sandbox: the Obsidian headless CLI syncs the markdown down, the agent enriches whatever is not yet stamped, and it syncs back, so you wake up to a refreshed wiki he describes as the daily paper except it is yours. The closing trick is a graph view of every note, built in HTML and Tailwind purely by asking an agent for it, which turns out to be a decent map of where your own thinking has gaps. Speaker info: https://x.com/bholmesdev https://bholmes.dev / bholmesdev Timestamps: 0:00 - The disorganized notes folder problem 1:30 - Where we are going: raw notes to a browsable wiki 2:46 - Getting raw thoughts down is the hard part 3:25 - Why dictation beats typing 4:02 - Local voice tools that need no subscription 5:17 - Be scrappy, you need volume 5:53 - What an enriched note looks like 6:32 - The enrich note skill 7:11 - Tagging from a fixed list, not inventing new ones 7:49 - Running the skill 9:07 - Following your own rabbit holes 9:45 - Generating a wiki over the notes 11:40 - The Karpathy gist behind the idea 12:54 - Wikis for work: people, meetings, sources 13:31 - Making it run without you 14:10 - Local automations against cloud ones 15:29 - Syncing markdown into a sandbox and back 16:45 - Waking up to a fresh wiki 18:01 - Visualizing the whole notebook 18:39 - A graph view built by asking for it 19:17 - Finding the gaps in your own thinking 20:30 - Tools mentioned