Architecture

Three layers

graph LR
    Claude[".claude/<br/>Agents + Skills"] --> Python["distillary/<br/>Python utilities"]
    Python --> Brain["brain/<br/>Obsidian vault"]
    Brain --> Quartz["Published site<br/>+ agent.json"]

Layer	What	Contains
.claude/	The brain	14 agents + 9 skills + CLAUDE.md
distillary/	The hands	10 Python utility files
brain/	The output	Obsidian vault with all knowledge

Agents

14 agents — 10 haiku (bulk) + 4 opus (reasoning)

Source processing

Agent	Model	Job
extract	haiku	Text → atomic claims with source_ref
dedupe	haiku	Merge duplicate claims
entities	haiku	Find people, concepts, companies
entity-link	haiku	Add [[wikilinks]] to bodies
link	haiku	Find tensions, patterns, evidence
group	opus	Cluster claims into parents
pyramid	opus	Build hierarchy to root

Brain-level

Agent	Model	Job
doctor	haiku	Fix orphans, discover ghosts
combine	haiku	Merge multiple sources
concept-mapper	opus	Find same-concept pairs
compare	opus	Write cross-source synthesis
bridge-builder	haiku	Create unified entity notes
annotate	haiku	Help write personal notes
explore	opus	Suggest what to investigate

Skills

Skills chain agents into workflows. Say what you want, the skill orchestrates the agents.

Skill	You say	What happens
distillary-add-source	”Add this book”	Full pipeline + auto-bridge
distillary-decompose	”Extract claims”	Core extraction pipeline
distillary-combine	”Combine these”	Concept mapping + bridges
distillary-doctor	”Fix my brain”	Doctor + suggestions
distillary-publish	”Publish”	Quartz + agent.json
distillary-retrieval	”Query this brain”	Shareable navigation skill
distillary-use-brain	”Explore”	6 navigation strategies
obsidian-bases	”Create a base”	Analytical database views
quartz-rendering	”Write docs”	What Quartz can render

Brain structure

brain/
  sources/{slug}/           Per-source content
    _source.md              Metadata (title, author, type, URL)
    claims/
      atoms/                Layer 0 — individual propositions
      structure/            Layer 1 — argument parents
      clusters/             Layer 2 — chapter groups
      [root note]           Layer 3 — thesis
    entities/
      concepts/             Technical terms
      people/               Authors, figures
      companies/            Organizations
  shared/                   Cross-source
    concepts/               Bridge entities
    analytics/              Comparison + stats
  personal/                 Your voice
    annotations/            Reactions to claims
    notes/                  Freeform thinking

Design decisions

Why these choices

• Agents are the pipeline, Python is the plumbing. Agents think. Python does file I/O.
• Haiku for bulk, opus for reasoning. 16 parallel haiku = minutes not hours.
• Post-processing catches mistakes. fix_ghost_links, _wire_parent_links handle format inconsistencies mechanically.
• One vault, not many. Sources accumulate. Bridges grow.
• Backlinks are the search engine. Entity pages + backlinks answer questions without keyword search.

Acknowledgments

Published with Quartz by Jacky Zhao — renders wikilinks, graph view, backlinks, Mermaid, and callouts natively.