A creative practice vault where human writing and agentic collaboration sustain each other.
An Obsidian vault built as a shared environment for a human writer and an AI collaborator (Claude Code). The human writes freely — journals, drafts, captures — without structural obligations. The agent maintains a linked atomic layer, surfaces connections, and produces argued reflections. The vault's hyperlinked quality grows from collaboration, not manual linking.
The architecture is a single-pool zettelkasten traversed by multiple frames (perspectives). Ideas exist once and are seen differently through different lenses.
The agent operates through repeatable skills and conversational collaboration:
- reflect — the agent's daily reading practice. Reads the journal, engages with the thinking, follows threads into the vault, produces the agent's own developing journal.
- atomize — extracts concepts, claims, and methods from journals and inbox into single-concept atomic notes with typed relations.
- tend — enriches the atomic layer: discovers connections, promotes maturity, harvests concepts from project drafts, proposes new frames.
- frame-read — reads the atomic layer through a named perspective and produces an argued reflection.
- trace — generates Obsidian canvas files mapping atoms, connections, and provenance paths visually.
Skills learn from each other through structured memory entries
with atoms_touched metadata. Each skill reads the most recent
output from every other skill before it runs.
Human writes here: Agent never modifies these
_journal/ (daily notes)
_inbox/ (external captures, feedback, references)
_projects/*/notes/ (brainstorming per division)
_projects/*/drafts/ (composed writing)
Agent maintains here: Human browses
_atoms/ (single-concept notes, wikilinked)
_memory/ (skill traces, collaboration logs)
_reflection/ (agent's journal)
_frames/ (proposed) (frame proposals from tend)
Shared:
HOME.md (human zones + agent-refreshable zones)
_frames/ (active) (user-created or user-activated)
_projects/*/brief-*.md (user-written, agent helps draft)
The vault's vision and architecture are documented in _system/:
VAULT_VISION.md— pillars and principlesVAULT_DESIGN.md— full architecture referenceAGENT_PROTOCOL.md— what the agent reads on every invocation_skills/— skill definitions_design/— design critiques, comparisons, and experiment seeds
The design documents in _system/_design/experiments/ include five
experiment seeds (THE_TABLE, THE_MEMBRANE, THE_SEANCE, THE_COMPOST,
THE_DRIFT) and design notes on architectural decisions. These are
living documents — some aspirational, some ready to test.
- Open this folder as an Obsidian vault.
- Install three community plugins (Settings → Community Plugins):
- Dataview — powers dashboard queries on HOME.md
- Templater — auto-applies templates for journal and inbox
- Homepage — opens HOME.md on startup
- In Dataview settings, enable JavaScript Queries and Inline JavaScript Queries.
- Plugin settings, CSS snippets, and folder templates are pre-configured.
The vault's skills work without external integrations. These are optional enhancements that extend what the agent can do during skill runs.
Reddit access (for reflect to follow Reddit links in journals):
claude mcp add --transport stdio reddit-mcp-buddy -s user -- npx -y reddit-mcp-buddy
Provides two capabilities:
get_post_details— fetches a Reddit post and its top-level comments. Used when a journal entry links to a Reddit post and asks about its content or the shape of the discussion.user_analysis— looks up a user's recent comment history. Used when the journal asks "which commenter is me?" and the human's Reddit username is available.
If you're reading this vault with an AI assistant, start with:
_system/VAULT_VISION.md— what this vault believes_system/VAULT_DESIGN.md— how it's built_system/AGENT_PROTOCOL.md— how the agent operates_system/_design/DESIGN_NOTE_TANA.md— what makes this vault different from structured note-taking tools_system/_design/DESIGN_CRITIQUE.md— known shortcomings and aspirations