A knowledge workspace that navigates itself. A Claude skill that turns a pile of files, documents, and scattered notes into a structured workspace both humans and agents can use: a hub-and-spoke knowledge base with a tiny front-door index, documented practices and procedures, and rituals that keep it from going stale. Built token-first, so an agent answers any question by loading the smallest possible slice of your workspace instead of drowning its context in it.
Built by @the.jacked.dev. Free. MIT. Markdown-first and tool-agnostic (plain folders work; Obsidian users get wikilink niceties for free).
Humans and agents fail at messy workspaces differently. Humans can't find things; agents can, but pay for every token they load doing it — and a "helpful" 40k-token context dump crowds out the actual task. The same design fixes both: hub-and-spoke with progressive disclosure. A one-screen index points to one-screen hub notes; hubs point to deep notes; cost correlates with relevance. One fact lives in one place, and everything else links to it.
The part most systems skip: a workspace nobody updates is worse than none, because it serves stale answers confidently. So the rituals ship with the structure.
Five phases:
- Survey — inventories what exists, classifies it (projects / areas / reusable knowledge / reference), collects the real questions the workspace must answer, and notes which existing conventions people actually follow.
- Architect — designs the taxonomy, frontmatter schema, naming, and templates, and writes the workspace's own conventions doc: the constitution a cold-started agent could work from tomorrow. You approve the map before anything moves.
- Migrate — builds the skeleton, then moves content one unit at a time: dedupe toward one-fact-one-place, archive (never delete), most-asked-about unit first as the exemplar.
- Wire in agents — entry contracts in every repo ("read this hub first, log decisions there"), a reading protocol with per-layer token budgets, and write-back duties so freshness becomes a side effect of doing work.
- Ritualize & verify — installs capture / wrap-up / drift-audit rituals, then runs the cold-start test: can a fresh agent answer your real questions from the index + one hub + a leaf or two? Structural gaps get fixed in structure, never with a fatter index.
On big corpora it orchestrates subagents: parallel inventory readers, per-unit migration agents working from the conventions doc, and a fresh-eyes auditor who cold-start-tests the result. Architecture itself is never delegated.
Home.md the index — one line per thing, always cheap to load
Projects/<Name>/<Name>.md hub notes — one screen of current truth per unit
<Name> Decisions.md append-only decision logs (what / why / alternatives)
Areas/ Knowledge/ ongoing responsibilities; reusable patterns & procedures
Ideas/ (_inbox/) zero-friction capture, processed later
Templates/ Archive/ note skeletons; retired things (moved, never deleted)
…plus the conventions doc that makes it self-explaining, and rituals wired into your agents' entry points.
git clone https://github.com/nate-zz/Lodestone.git ~/.claude/skills/lodestoneOr add it to a single project at .claude/skills/lodestone/.
Then, in Claude Code:
/lodestone
…or just ask: "organize my notes into a knowledge base," "make this workspace agent-friendly," "help me document our procedures so agents stop asking."
SKILL.md the method (survey → architect → migrate → wire agents → ritualize)
references/
structure.md the topology — hub-and-spoke layers, taxonomy, naming, one-fact-one-place
conventions.md note anatomy — frontmatter, hubs, decision logs, procedure notes
efficiency.md token budgets, the reading protocol, the cold-start test
upkeep.md capture / wrap-up / drift-audit rituals, migration triage, archive rules
One fact, one place, and every layer earns its tokens. Everything else is a link.