Source: raw/5_Hacks_to_Instantly_Level_Up_Your_AI_OS.md
Creator: Nate Herk | AI Automation URL: https://www.youtube.com/watch?v=Ek1NBfnnTH0 Published: 2026-07-23 Duration: 25:03 Platform: YouTube
A practitioner walkthrough of how to organize an “AI operating system” — the file tree, wikis, and routing rules an agent reads from — framed around the failure it prevents. The stated problem is specific and matches this vault’s own experience: if the organization is wrong, the agent starts hallucinating, and not just in chat — the bad context propagates into the skills and automations it builds. The five items below are the ones the video names.
Key Takeaways
- 1. Routing rules matter, folder shape does not. The creator’s own tree is one large
projects/folder with everything drilled down inside it, but he is explicit that this is arbitrary: “it doesn’t matter if you have it super flat… or if you have one main one that you drill down everything into. All that matters is that you have the right routing rules in place so that you and your agents can find it.” Optimize the index, not the hierarchy. - 2. Have the AI audit itself. On a weekly or monthly cadence — or after a batch of changes — instruct the agent to open every file, re-read the routing rules, and confirm they are still accurate. He has since built this into a reusable skill rather than re-prompting it each time.
- 3. Promote repeated just-in-time fetches into scheduled automations. If you keep telling the agent “go pull that data,” that repetition is the signal: the data is not situational context, it is standing context, and it should be refreshed on a cron rather than fetched on demand.
- 4. Segment knowledge into separate wikis once a node starts growing. He runs a YouTube wiki and a meeting-transcript wiki as distinct stores, split because both were growing and expected to keep growing. The heuristic: when a distinct “node of knowledge” is on a growth path, give it its own store rather than letting one wiki absorb everything.
- 5. Backtrack and correct the moment retrieval misbehaves. The named symptoms are precise — the agent searches for five minutes for something it should have found instantly, or reports it lacks access to something you know is there. Correct it in the moment (“you actually do have access to that, make sure that never happens again”) so the correction is captured rather than re-encountered.
How this compares to the wiki’s own pattern
Four of the five map onto machinery this vault already runs, which is the useful part — it is convergent practice from an independent operator rather than a new idea.^[the mapping is this wiki’s synthesis; the video does not reference this vault or its tooling]
| Hack | This vault’s equivalent |
|---|---|
| 1. Routing rules over hierarchy | wiki/_master-index.md → topic _index.md as the navigation layer, with QMD as the primary retrieval path |
| 2. AI audits itself | The Lint operation and the bin/post-ingest lint chain (stale sources, frontmatter integrity, manifest integrity) |
| 3. Repeated fetches become automations | The /inbox-refresh pullers and the weekly release sweep |
| 4. Segment knowledge | Two separate QMD collections — the public AI wiki and the WEO-internal marketing brain — kept deliberately apart |
| 5. Backtrack on bad retrieval | The correction/lessons loop; retrieval misbehaviour is treated as evidence about the index, not just the answer |
The one that is not already systematized here is #3 as a trigger. This vault automates sources it has already decided to track, but has no rule that says “you asked for this three times, promote it to a scheduled pull.” That is a cheap heuristic worth adopting.
Try It
- Run hack #2 once, manually. Ask your agent to read every routing rule and index file in your vault and report anything that no longer matches what is on disk. If it finds drift, that is your argument for making it a skill.
- Audit your own repeat requests. Look back through a week of prompts for any data pull you asked for more than twice. Each one is a candidate for hack #3.
- Check the growth curve on your largest topic. If one folder is absorbing unrelated material because it is the default destination, apply hack #4 and split before retrieval degrades.
Open Questions
- No measurements. The video reports no before/after retrieval quality, latency, or hallucination-rate numbers for any of the five changes — they are presented as practitioner judgment, which is why this article is
confidence: medium. - The audit skill is not shown in enough detail to reproduce. He mentions building it but the prompt/skill definition is not given.
- Segmentation threshold is unstated. “Once it starts growing” is not an operational trigger; this vault’s own
_catalog.mdsplit uses ~30KB, but whether any size threshold generalizes is unknown.