In Part 2, I reorganized my Second Brain so the folder structure gave each kind of information a clearer home. That solved one problem: where information lived. It did not solve the next problem: How an AI agent, or even future me, should approach that information. A folder name alone cannot answer questions like: - Where should I look first? - Am I looking at raw source material or knowledge I have already adopted? - Could this contain sensitive information? - Does using it require approval? - Might it be out of date? That gap led me to create what I’m calling Source Manifest Lite. It sounds more technical than it is. Source Manifest Lite is a small routing map for the major source areas in my Second Brain. It is not an inventory of every note, file, or document. I don’t want another system that becomes too heavy to maintain. Instead, each source area gets a short description covering what it contains, what it is meant to be used for, its general sensitivity, any approval boundaries, the canonical place to start, and where supporting material can be found. The distinction between a canonical starting point and supporting material has been especially useful. I may have several raw captures about a subject, but only one current operating note that reflects what I have actually decided to use. The raw material still matters. It provides evidence, history, and context. But it should not automatically outrank the operating note. The same principle applies elsewhere: - Public research is not handled the same way as a sensitive source. - A recommendation is not the same thing as an approved action. - A historical note is not automatically a statement of current practice. Without a routing layer, those differences can be easy to miss, especially when an agent retrieves a passage that appears relevant but lacks the context needed to use it properly. Source Manifest Lite gives the agent a better starting point. It can indicate where to begin, what kind of material it is entering, and where it should slow down or ask for approval.

Part 1 ended with a promise: before asking an AI employee to use more of my Second Brain, I had to confront the knowledge already inside it. My vault had grown organically. Its folders reflected different phases, projects, and ways of thinking rather than one deliberate system. The tempting response was a broad cleanup: redesign the structure, move everything into its “proper” place, rename inconsistent notes, and remove whatever looked outdated. I decided not to do that. A large-scale rebuild could easily create more problems than it solved: - damaged links - stale source references - lost context - uncertainty about which note should count as the source of truth - more confusing retrieval later So I adopted a slower, safer pattern: 1. Inventory the existing structure. 2. Define the order in which folders would be reviewed. 3. Inspect one folder at a time. 4. Separate recommendations from authorized changes. For each folder, I worked through a consistent set of questions: - What purpose does this folder serve now? - Which note should count as the current source of truth? - Which notes are still working drafts? - Which items should remain source material rather than operating guidance? - What might be better moved or linked elsewhere? - Which decisions still require my judgment? That final question mattered. An AI agent could inspect the structure, identify inconsistencies, and recommend changes. But a recommendation was not authorization. Moves, renames, deletions, and broad rewrites required explicit approval. That boundary helped preserve context and prevented a cleanup exercise from quietly becoming a messy migration. The lesson was that folder structure is not just cosmetic organization. It is part of the vault’s retrieval infrastructure, part of how the system finds and surfaces useful context. Folders influence where I look, how an agent interprets context, which version appears authoritative, and whether useful knowledge can be found again without reconstructing its history.

Yesterday's post was the unnumbered prologue, the announcement that I was going to start sharing what I'm learning while building a Second Brain for AI employees. This is where the numbered series begins: Part 1 of 8. The idea started after I watched Nick appearance on Marketing Against the Grain. The Obsidian part of his AI employee setup made me look at my own vault differently. Giving an AI employee access to a folder of notes is easy. Making those notes something it can use reliably is not. When I looked more closely at my own vault, I realized I couldn't answer some basic questions consistently: - Which notes were current? - Which were unfinished drafts? - Which captured sources still needed a decision? - Which source should take precedence when two notes disagreed? - Could I reorganize the vault while preserving its links and context? - People often fill in those gaps from memory. An AI agent may interpret them differently from one task to the next. If it treats a stale note like a current decision, it can produce an answer that sounds reasonable but rests on the wrong context. For me, unreliable context can be worse than no context because it still looks legitimate. So the work stopped being about adding more notes. It became about building the layers underneath them: organization, source awareness, trust signals, validation, repeatable procedures, and a clear human control layer. Here is the eight-piece blueprint behind this series: 1. A folder-by-folder reorganization plan to remap the knowledge I already had while preserving links and catching problems early. 2. A source manifest to show what the major sources are, where they live, how safely they can be used, and which source takes precedence. 3. An ingestion log to record what entered the system, what happened to it, and where the work can resume. 4. A source review queue for captured material that still needs synthesis, adoption, archival, or a deliberate decision to ignore it. 5. Canonical confidence labels to distinguish what is currently authoritative from a working draft, an outdated note, a raw source capture, or background reference material. 6. A validation report to catch broken links, duplicate titles, provenance gaps, and other structural problems before they quietly weaken retrieval. 7. An SOP repository to turn trusted knowledge into repeatable procedures and bounded instructions instead of re-explaining the work every time. 8. A plain-English command registry so I can activate recurring workflows without having to remember their internal names.

I watched @Nick Vasilescu 's episode on Marketing Against the Grain about building a first AI employee, and one part of it stuck with me more than the rest: the Obsidian piece. A lot of people focus on giving an agent a computer, tools, and an inbox. Those are the obvious parts. But Nick also gave the agent a knowledge base, a Markdown vault with context around the business and current work, so it wasn't starting from zero every time. That's the part that made something click for me. I'd been using a Second Brain setup for a while, mostly the way people usually use one: dump notes, link ideas, find things later. Useful, but personal. After watching that episode, I started looking at it differently. The same structure that helps a human think clearly can also help an AI employee keep more context between sessions, instead of starting cold every time. That shift matters. It changes what the vault is for. It stops being just personal notes and starts becoming infrastructure an agent can read from and eventually write back to. I'm still building this piece by piece, not shipping a finished system. But I wanted to say thanks to Nick for the nudge and let this group know I'll be posting shorter field notes here more often as I work through it. Partly to build the habit of writing things down as I go. Partly because I'd rather get feedback early than polish something in private for months. If this rhythm holds up, I'll eventually bring it over to my own channels too. But I want to work it out here first with people who are building similar things. If you've already gone down this road, tell me where it got messy. I'd rather hear it from you now than find it myself in three weeks. This is Chris from the Digital Field of Dreams, signing off.

After resolving error, I am following demo on this video. https://youtu.be/Y7FVj4njob0?si=ztCXUAGa5oI4zxWq&t=2046 34:06 mark. Choosing model part. It says you could navigate by pressing arrow up and down but it doesn't work at all. How to choose option? First part, it worked-choosing first choice. But model part...I can't select by pressing arrow button. Up down. So I am stuck at this point. If I press enter key then it shows backend terminal choosing option which I don't know what to choose there either. ------------ This is my screen recording. https://www.loom.com/share/a22d3b996eb642c198c3b96306ea6c26 I found out I had to click bottom window which is terminal. So I could choose chatGPT 5.5 Then there was no more procedure like you showed as demo. I keep trying... This is last attempt. https://www.loom.com/share/0f5bf82b4f00431ab0e0709c6e46825b I can choose model again...how do you select? Choosing web search & extract etc...that part. Is it because I use free account so it doesn't show? -------------- I am non technical person. It would be very helpful to screen record from signing into your fresh Orgo computer then....how to set up Hermes and Obsidian. I setup Obsidian on my home mac and connected to GitHub-synced now. I want Orgo Obsidian to be synced with GitHub Obsidian. Anyway, first thing is how to make Hermes to work properly. I saw many videos on how to set Hermes on VPS. Thank you!