This is the short version of how to grow a brain that gets more useful every month instead of slowly turning into a junk drawer. None of these habits are heavy, because most of the work is not yours to do.
Two principles run under everything here. The first is about shape: atomic pages that link. Small pages about one thing, connected to other small pages. The second is about who does the work: the AI edits, creates, and maintains; you review and decide. You are not the janitor of your brain. You are the editor who says yes or no.
The weekly rhythm
Set a recurring half-hour. Most of it is review, not labor.
- Let the AI keep the sources current. As drives and tools change, the source-sweep skill re-walks the folders the brain was mapped from, finds what changed, and updates the pages that cite those sources, with a reason for each change. Ask for it by area ("sweep the sales folders and update the pages"). You read the results.
- Ask what changed. Run the weekly-review skill: it reads the brain's real change history for the week and gives you a short, grouped summary of what was added, edited, or removed, and by whom. This is how you stay on top of a brain many people write to without reading every page.
- Capture what only lived in your head. Finish the week by telling the AI the two or three things worth keeping that never made it into a document, a decision and its why, a lesson, a commitment. Ask it to write them as pages in the right place. It drafts, you confirm.
That is the loop. The AI sweeps, summarises, and drafts. You review and decide. You will get faster at the reviewing after three or four weeks.
The monthly rhythm
Once a month, give the brain an hour of oversight.
- Look at the graph. Open the graph view in your dashboard and take in the whole shape. Connected clusters are healthy; floating islands are not. Ask the AI to walk the graph (
walk_graph) and list orphan pages and clusters that should connect but do not, then let it add the missing links. - Run a health check. Ask the AI to check the brain for broken wikilinks, duplicate entities, and stale pages, and to fix the ones it is sure about. The health-checks guide has the full routine.
- Enrich what matters. Pick the three most important pages from the month (the top project, the deal that moved, the person who now matters). Ask the AI to pull in everything it knows and make each page denser. This is where the brain stops being a filing cabinet and becomes a thinking partner.
The atomic-pages principle, with an example
Say a project kickoff happened on Tuesday. The team agreed on a Q3 timeline, picked a vendor, flagged a hiring need, and revived a shelved idea.
The wrong way to capture this is one page titled "Kickoff 2026-05-12" holding four paragraphs about four different things. It is fine for a week. After a month you cannot find the timeline decision because it is buried inside a meeting note.
The right way is a handful of small pages:
wiki/sources/2026-05-12-project-kickoff.md, the meeting itself, three lines summarising what was discussed, with wikilinks to each topic.wiki/projects/q3-timeline.md, a project page started from this meeting that will grow as the quarter unfolds.wiki/decisions/pick-vendor-2026.md, captured as a decision with its why.wiki/projects/revived-idea.md, moved out ofarchive/back intowiki/projects/.
Now when someone asks "what is the plan for Q3?" the AI lands on a single page that links back to every meeting where it was discussed. That compounds. A meeting-only note does not.
You do not do this splitting by hand. Paste the raw notes into your AI chat and ask it to split them into atomic pages, file them in the right folders, and link them together. It does the work; you confirm.
Sensitive material
Not everything belongs where everyone can read it. Personal notes, a difficult conversation, an unannounced plan: these still go in the brain, but with their visibility set.
visibility: privateon a page limits it to its author and the brain's owners and admins.visibility: confidentialscopes a page to a named audience.
Set it in the page's frontmatter, or just tell the AI ("make this page private"). Because access is enforced on every read, a private page never surfaces in an answer for someone who should not see it. When something feels sensitive, mark it rather than keeping it out of the brain.
Handling entity disambiguation
The same person shows up three ways across the brain: "Jane", "Jane D.", "Jane Doe-Smith from Acme". The fix is one canonical page (wiki/people/jane-doe.md) with the other spellings in its aliases field:
aliases: [Jane, Jane D., Jane Doe-Smith]
Ask the AI to walk the brain and re-link every mention of any alias to the canonical page. It does it in one pass; you review. The same applies to organisations and concepts. One canonical page per real-world entity, everything else points to it.
Archive versus delete
If something is finished but still interesting, move it to archive/. Old projects, decisions already executed, deals that closed. The archive is searchable and the AI still reads it for context, but it does not clutter the active wiki.
If something is wrong (a duplicate, a misfiled page, a draft that went nowhere), delete it. Nothing is truly lost: every page keeps its History, a save you regret can be reversed with Undo, and deleted pages can be restored. Do not hoard.
Rule of thumb: if you are not sure, archive. You can always pull a page back from archive/ to wiki/.
Power follows the person
You do not set up different maintenance for different roles. A skill runs inside the session of whoever invokes it, so it can only ever reach what that person is allowed to reach. When an owner runs source-sweep it touches the whole brain; when a member runs the same skill it touches only their departments. The scope is set automatically by who is asking, not by a separate tool per role.
So treat the AI as the colleague who keeps your filing system honest, at exactly your level of access. Every week or so, ask one of:
- "Look at the brain. What is missing that should be there?"
- "Which pages look like they should be split?"
- "Which people should have richer pages?"
- "Where are the dead ends in the graph?"
You decide what matters. The AI does the work.