The guide
CLAUDE.md tells the agent how to behave. agents.md describes what the agent can do. Neither one tells the agent what you’ve been thinking about.
The guide fills that gap. It tells Enzyme which entities to focus on and what to ignore. Ten to thirty lines. Every tag, folder, and wikilink you list becomes an entity Enzyme generates catalysts for. Everything else stays indexed but doesn’t get its own questions.
#seedling #evergreen #question #unresolved folder:Readwise/Books folder:Readwise/Articles folder:journal/daily [[open questions]] [[§ Design principles]] excludedTags: - status/done - template
Works with whatever system you already have. PARA folders become entities. Zettelkasten note-type tags become entities. MOCs and evergreen note titles become entities when they're wikilinked. Whatever vault structure you've settled on — enzyme reads the primitives it already uses.
#architecture #product-decisions #user-feedback folder:docs/ADRs folder:meeting-notes folder:.claude/memory [[migration]] [[v2 API]] excludedTags: - draft - wontfix
Your repo already has opinions — docs/, .claude/memory/, meeting notes from the last six months.
Enzyme reads those folders as entities. The agent's own accumulated
memory files and
decision records
become searchable by theme, not just by filename.
Wikilinks to projects let enzyme track when references start and stop.
#high-signal #returned-to #considered-and-rejected folder:saves folder:explorations folder:feedback
Tags here are behavioral — applied by the agent from interaction patterns, not by the user.
#returned-to means the user came back unprompted.
#considered-and-rejected means it was shown and passed over.
These become entities with temporal weight.
The organizational primitives represent
taste, not categories.
On first init, the agent scans your workspace, presents what it found, and writes one from your confirmation. If catalysts feel too broad, tighten the guide and run enzyme refresh. However, Enzyme also works without a guide.