Route, Don't Organize

The moment you ask "where should this note go?" the system has already failed. That question is friction, friction kills capture, and a knowledge system that does not capture has nothing to compound on.

The principle

The user expresses something — an insight, a decision, a reference, a person, a half-formed thought. The system identifies what kind of thing it is and routes it to the right place, with the right shape, under the right links.

The user never picks a folder. Never picks a tag. Never decides between "is this a note or a doc." That mental load belongs to the system.

Why it matters

Every routing decision a human makes is a tax on capture. Compound it across a year and you get the predictable failure mode of every personal knowledge system: a graveyard of orphaned notes filed in folders no one revisits, with categories that drifted three reorganizations ago.

Routing rules don't need to be sophisticated. They need to be consistent and invisible. A small set of unambiguous rules applied to every input beats an elegant taxonomy applied unevenly.

How it works

Three pieces:

1. Type detection — the system recognises the kind of input (entity, decision, source, idea, note-on-existing-thing) from natural language. "I just talked to Sam about the renewal" is a meeting log on an existing entity. "Let me write down a frame for evaluating retainer deals" is a new concept.
2. Canonical shape — each type has a fixed file template and folder. Insights go to /insights/, entities to /entities/{type}/{slug}.md, sources to /sources/. The shape is the same every time.
3. Link-on-write — when the system files something, it also resolves and creates the wikilinks back to whatever entities, concepts, or sources the input mentioned.

The user only ever talks. The filing happens behind the scenes.

In Claude primitives

• CLAUDE.md carries the routing table — "if user mentions a person, file to entities/people/; if user mentions a decision with rationale, file to decisions/; never ask where to put something."
• Skills encode each routing destination as a callable handler (save-insight, log-meeting, update-entity).
• Subagents can do the type-detection step in parallel when input is ambiguous — one agent classifies, another routes.
• Hooks (UserPromptSubmit) can pre-process raw input before Claude sees it, stripping noise and tagging type.

Where I've seen it

The pattern holds in my personal orchestrator, in a team KMS's ingest pipeline, and in an app-scoped second-brain KMS. In every case, the move from "user files" to "system routes" was the single biggest unlock for capture rate.

The contrarian read

You lose the satisfaction of a tidy filing system. Files end up where the rules say, not where you would have put them. That is the whole point — your aesthetic preferences are a worse organising principle than a deterministic ruleset applied a thousand times.