Knowledge Workflow

Principles

“Improving one’s understanding of a field, or communicating this understanding successfully to others.” — Terence Tao

This second brain is not a storage system. It is a long-term practice for improving cognition: seeing clearly, reasoning from evidence, expressing useful ideas, and acting with better judgment.

• Cognition compounds. Tools, sources, and models matter only when they improve understanding. Prefer sustained, honest progress over the appearance of productivity.
• Capture is a commitment, not an achievement. Inbox items are possibilities, not possessions. Information becomes knowledge only when it is examined, connected, tested, or expressed.
• Input must earn an output. Reading and collection are inputs; writing, teaching, building, deciding, and practice are the tests of understanding. Knowledge should change what I can explain or do.
• Reason from first principles. Treat conventions, popular claims, and model output as hypotheses. Start from evidence, clarify assumptions, and rebuild the conclusion.
• Keep notes revisable. Every note is a current best view. Evidence, practice, and thoughtful disagreement should improve it rather than threaten it.
• Think in decades. Follow the discipline of patient accumulation: small, consistent advances that become meaningful through time, review, and compounding.

Purpose

Turn selected inputs into durable understanding, useful expression, and better practice.

Structure

00_Inbox/Manual/   Personal captures, conversations, and ideas
00_Inbox/Auto/     Curated machine-generated research digests
01_Sources/        External material worth citing or revisiting
02_Topics/         Questions, evidence, and the current map of a subject
03_Notes/          Personal insights, models, decisions, and essays
99_Deprecated/     Superseded, low-value, or inactive material kept for history

Publishing Model

The internal directory structure serves the knowledge workflow. Public URLs serve readers and remain independent of where a note lives in the Vault.

Internal location	Default	Public route
index.md	Published homepage	/
00_Inbox/	Never publish	—
01_Sources/	Publish only selected, annotated sources	/sources/<slug>
02_Topics/	Publish mature topic maps	/topics/<slug>
03_Notes/	Publish mature notes and essays	/notes/<slug>
99_Deprecated/	Never publish	—

Publishing is explicit. A public Note uses stable, lowercase, kebab-case routing metadata:

---
title: AI Compute Architecture Trends
description: A framework for understanding device, edge, personal, and cloud AI compute.
tags:
  - ai
  - compute-architecture
publish: true
permalink: /notes/ai-compute-architecture-trends
---

The publishing service creates a public projection before Quartz builds the site. It excludes private Markdown and unreferenced files, maps public notes to their stable routes, rewrites links between published notes, and copies only attachments referenced by published content. A public note must not link to a private or missing note.

Note Templates

The YAML block at the very top of a Markdown file becomes Obsidian Properties. Keep publish as a checkbox/boolean, tags as a list, and public routes in lowercase kebab-case. Drafts do not need a permalink until they are ready to publish.

Unfinished or Private Note

Use this for working notes in 03_Notes/, or adapt it for unfinished Topics and Sources. It remains in Git and backups but never reaches the public site.

---
title: Working Title
description: What this note is trying to understand.
tags:
  - draft
status: draft
publish: false
---

Published Note

Store in 03_Notes/. The public URL becomes /notes/<slug>.

---
title: My Note
description: A concise description for readers and search.
tags:
  - ai
  - architecture
status: evergreen
publish: true
permalink: /notes/my-note
---

Published Topic

Store in 02_Topics/. A Topic is a maintained map of questions, evidence, and related Notes—not merely a long article.

---
title: LLM Memory
description: A map of research and evolving understanding about LLM memory.
tags:
  - llm
  - memory
status: active
publish: true
permalink: /topics/llm-memory
---

Published Source

Store in 01_Sources/. Preserve the original URL and distinguish the author’s claims from personal annotations.

---
title: Source Title
description: Annotated reading notes and why this source matters.
tags:
  - source
source_url: https://example.com/original-source
author: Author or Organization
accessed: 2026-08-10
status: reviewed
publish: true
permalink: /sources/source-title
---

Images and Attachments

Default rule: keep Markdown and a small number of essential, optimized images in the Vault; link to PDFs at their original source by default; archive PDFs in MinIO only when preservation is necessary; link to videos or store them outside Git.

Asset	Default location	Guidance
Markdown	Vault and Git	Versioned knowledge and publishing source
Essential diagram or screenshot	Vault	Prefer WebP; target under 500 KB and keep under 1 MB
Large or original-resolution image	Future MinIO	Compress first; keep outside Git if it remains large
External paper or report	Original landing page	Prefer DOI, arXiv, publisher, or organization page over a raw PDF URL
PDF that must be preserved	Future private MinIO bucket	Keep its source URL, checksum, access date, and rights metadata in the Source note
Video, audio, dataset, or model	External platform or MinIO	Never commit large binary media to the Vault

Git retains old versions of binary files even after deletion. A large attachment therefore expands the Vault clone, Gitea instance backup, and encrypted repository bundles for as long as that history is retained. Storage capacity alone is not a reason to put binaries in Git.

For an ordinary external paper, link to its stable landing page:

[Attention Is All You Need](https://arxiv.org/abs/1706.03762)

For an essential local image, embed it from anywhere inside the Vault:

![[ai-memory-architecture.png]]

Standard Markdown is also supported:

![AI memory architecture](../assets/ai-memory-architecture.png)

The public projection copies only files referenced by published content into /assets/. A build fails safely when an attachment is missing, outside the Vault, or has the same filename as a different referenced file. Use descriptive, globally unique, lowercase kebab-case filenames, such as llm-memory-hierarchical-architecture-2026.webp. Unreferenced attachments remain private.

Do not hotlink third-party images that are essential to an argument: they may disappear, reject cross-site loading, load slowly, or have unclear reuse rights. Preserve an authorized and optimized copy locally for a small image, or in MinIO for a large image. Nonessential images should be omitted.

MinIO is deferred until attachments exceed roughly 1 GB in total, individual assets regularly exceed 5–10 MB, or PDF/image growth reaches hundreds of megabytes per year. When introduced, use separate public and private buckets, expose stable asset URLs, keep online objects on the 2080 Ti local disk, and back them up to /mnt/nas10tb. MinIO is online storage, not the backup itself.

Homepage Maintenance

The Vault root index.md is the manually curated public homepage. It must remain published and does not need a permalink:

---
title: Welcome
description: Notes, ideas, and thoughtful conversations.
tags:
  - index
publish: true
---

Link only to content that is already publishable:

# Welcome
 
Welcome to Alan's Digital Garden.
 
## Explore
 
- [[02_Topics/llm memory|LLM Memory]]
- [[03_Notes/my note|My Note]]

Keep internal Obsidian paths in index.md; the public projection rewrites them to stable /topics/..., /notes/..., or /sources/... routes. Because the homepage is public, a link to a private or missing note stops the build and preserves the previous live release. Add only a small set of useful entry points rather than every new note.

Obsidian Git Routine

Current Mac settings pull once when the Vault opens (autoPullOnBoot: true) and pull again before Push (pullBeforePush: true). Timed auto-pull is disabled (autoPullInterval: 0), which avoids unexpected file changes while writing. Automatic Commit and Push are also disabled, so publishing remains an explicit action.

Reader Discovery

Use a deliberate hierarchy rather than exposing the internal file tree:

1. Homepage: manually curated starting points and featured work.
2. Topics: human-maintained maps that explain how related notes fit together.
3. Search and tags: direct retrieval across titles, text, and controlled vocabulary.
4. Graph and backlinks: secondary exploration of meaningful note-to-note relationships.

AI-assisted public navigation is deferred until the published corpus is large enough to justify semantic retrieval, operational cost, and citation safeguards.

Flow

flowchart TD
    A["00_Inbox<br/>Manual and Auto"] --> B["Review<br/>triage and deduplicate"]
    B -->|"useful external material"| C["01_Sources<br/>traceable evidence"]
    B -->|"weak, duplicate, or inactive"| Z["99_Deprecated"]
    C --> D["02_Topics<br/>questions and current understanding"]
    D --> E["03_Notes<br/>insights, models, decisions, and essays"]
    E -->|"ready to share: publish true + stable route"| P["Public Projection<br/>notes, topics, sources"]
    P --> F["Public Garden"]
    F --> G["Comments, conversations,<br/>and practical feedback"]
    G --> A
    E -->|"superseded"| Z
    C -->|"superseded"| Z

Capture is not completion. Review the Inbox, retain traceable evidence, develop it through Topics and Notes, and use feedback to revise the current view. Move material to 99_Deprecated only when it is no longer active or has been superseded.

Operating Rules

• The machine may suggest, summarize, deduplicate, and propose links; it must not publish or establish facts on its own.
• Keep machine intake bounded: a small daily digest per active topic, not an archive of every link.
• Keep every important claim traceable to a source, experience, or explicitly marked hypothesis.
• Give every published Topic, Note, or Source a stable permalink in its matching public route.
• Treat the homepage as a curated guide, not an automatic dump of recently changed files.
• For every five retained sources, produce at least one output: a Note update, synthesis, decision, or article outline.
• Review each active topic weekly. Pause intake when its Inbox exceeds 20 unprocessed items.
• For health topics, record evidence level and uncertainty. AI output is research support, never medical advice.

Revision Practice

For important notes, maintain confidence and last_reviewed in frontmatter. When a discussion changes your view, add a brief dated revision note explaining what changed and why.