Personal Second Brains and Workflows
Personal vault architectures, individual developer routines, voice and mobile capture workflows, and multi-month operational retrospectives.

An engineering case study detailing the architecture of Kompl, an open-source compounding LLM Wiki. Explores replacing query-time vector RAG with a compile-at-ingest pipeline, utilizing local NLP and deterministic rules to constrain LLM writes, tracking conflicting source claims in dedicated Disputed sections, and exposing the compiled wiki to external AI coding agents via an integrated MCP server.

A case study outlining a voice-first ingestion pipeline for the LLM Wiki pattern. Details how to utilize mobile audio recordings, Whisper API transcriptions, and local LLM processing to transcribe spoken thoughts and automatically compile them into a structured Obsidian wiki.
An LLM wiki changed how I work (Casey Newton on Platformer)platformer.newsA case study by tech journalist Casey Newton detailing his experience building a personal "beat wiki" inspired by Andrej Karpathy. Explores how using an LLM to process and synthesize source materials into a structured, self-updating Obsidian-compatible Markdown wiki serves as a valuable research assistant, streamlining beat reporting and productivity.
Build your AI Brain (Leo Alexandru)leoalexandru.substack.comA comprehensive production case study implementing a customized Karpathy LLM Wiki using the Claude Code CLI. Outlines a five-layer folder taxonomy, a flat permanent note topology, and a twelve-part modular agent skill ecosystem (
.claude/skills/) to drive automated multi-stream file compilation (/wiki-compile), structural link linting (/wiki-lint), and contextual content generation.
Part 3 of a 4-part series analyzing the read layer of an agentic wiki, explaining why feeding structured, pre-compiled markdown contexts back into context windows outperforms multi-round tool loops, and detailing index-first query routing.
Engineers use markdown wikis as lightweight memory for Codex, Claude Code, and Hermes (AI Primer)ai-primer.comA collection of case studies on how developers use local folders, cron jobs, and MCP servers to serve as agent memory.

An engineering case study introducing pre-KOMPL, a personal bookmark catalog designed to precede LLM Wiki compilation loops. Explores applying classical library science principles (Cutter's catalog objects, FRBR/BIBFRAME entity modeling, MARC access points, and Ranganathan faceted classification) to personal saves across browsers, X, and notes, combining local SQLite BM25 indexing with agent skills for faceted retrieval and dead-link tracking.
From Chat Memory to Knowledge Graph: My LLM Wiki Experiment (Chris Crosland-Taylor on Medium)medium.comA personal case study detailing the shift from ephemeral chat histories and vector RAG toward a persistent LLM Wiki. Explores ingesting raw notes into structured Markdown files, leveraging agentic skills for continuous compilation, and integrating a Neo4j property graph backend to support multi-hop relational queries.
From Karpathy's LLM Wiki to a Working Second Brain: My Implementation w/ Amazon Quick Desktopdoit.comDima Kramskoy's step-by-step production review of a local flat-file LLM wiki automated via Amazon Quick Desktop and served locally using MkDocs. Details a four-tier ontology schema (concepts, entities, projects, sources) and implements an elegant visual diff-approval ingestion pipeline that treats the agent as a local workspace code contributor while preserving absolute human state control.
- From Second Brain To LLM Wiki: How I Built a Compounding Knowledge Base with AI (Ian Forster on IMF Software Blog)blog.imfsoftware.com
A case study describing a transition from a traditional Logseq PARA second brain to an AI-maintained compounding wiki, detailing the three-layer layout (sources, wiki, schema) and three core compiler operations (ingest, query, lint).
From Second Brains To Compounding Intelligent Knowledge Bases (Robin Good on TRUST-able)robingood.substack.comAn essay examining the transition from personal document indexing to compiled sense-making engines. Evaluates how Tali Režun's open-source Curator application implements Karpathy's LLM Wiki pattern to build private, compounding knowledge architectures.

A practical case study detailing a capture pipeline that converts YouTube video streams into structured Markdown source evidence. Explores Slack-triggered automation, yt-dlp audio extraction, local speech-to-text transcription, and publishing formatted transcripts into an immutable raw/ directory for downstream agent synthesis.

A personal case study detailing an operational autopilot (AFCS) built over Karpathy's LLM knowledge base pattern to run four businesses. Highlights the critical compound loop of filing answers back into Markdown context, using strict quality review gates between drafts and permanent memory to prevent compounding noise, and keeping context sovereign so AI tools remain interchangeable.
HUGE update on obsidian-second-brain: The Architect (Eugeniu Ghelbur on The AI Operator)theaioperator.ioA case study detailing Ghelbur's Architect update to the obsidian-second-brain CLI tool. Explains how the system was expanded from compiling personal notes to parsing and documenting codebases inside the Markdown vault, detailing pipeline structures and context alignment.
An article detailing a multi-model pipeline compiling technical source materials into a SQLite-managed synthetic wiki.
I Built a Graph Explorer for Obsidian Bases. Here's Why LLM Wikis Need a Visual Layer (Sébastien Dubois)linkedin.comA case study explaining the design of the Graph Explorer Base View plugin. Details why flat markdown lists are hard to curate at scale and how visual nodes colored by confidence, maturity, and explored status aid human oversight.

An engineering case study detailing the construction of MindBase, an open-source, local-first LLM Wiki implementation. Explores architectural mechanisms designed to safely constrain LLM write permissions: strict Zod schema action boundaries preventing edits to raw sources, mandatory human checkbox approvals for ingest plans, tool-restricted builder agents with atomic snapshots and a 400-line cap, deterministic on-disk quote verification for contradiction linting, and single-shot JSON completions optimized for small local models.
I built this for myself. Then 1,374 strangers cloned it: How I gave Claude Code memory without a vector DB (Eugeniu Ghelbur on The AI Operator)theaioperator.ioA follow-up case study on the launch and open-source reception of the obsidian-second-brain CLI tool. Focuses on the compound effect of sharing personal AI integrations and details how to provide Claude Code with long-term memory capabilities without relying on third-party vector databases.
I Found a Full LLM Wiki App. So I Built the Smaller Thing I Actually Needed (Mark Chen on Medium)medium.comA personal implementation reflection contrasting heavy desktop LLM Wiki applications with a minimalist, terminal-first setup. Explains the rationale for avoiding UI overhead and application lock-in by maintaining a lightweight three-layer vault via Claude Code and plain Markdown.

A senior engineering leader's case study introducing his journey to build a compounding personal knowledge system, detailing why vector search RAG decays for agent context and illustrating the initial setup of an LLM-curated memory vault.
I Rebuilt Karpathy's LLM Wiki. Here's What I Learned (Eugeniu Ghelbur on The AI Operator)theaioperator.ioA personal retrospective exploring the construction of the Obsidian Second Brain CLI tool. Details the practical limitations of raw vector RAG for coding agents, outlines lessons learned around prompt routing and context persistence, and offers a blueprint for building folder-native knowledge layers in plain Markdown.

A personal implementation case study detailing how to build an LLM Wiki using Claude Code as a knowledge base maintainer. Explores setting up three-layer directory boundaries (raw sources, living markdown pages, steering rules), running ingest and linting workflows, and managing compounding personal context over time.
Karpathy Plugin for Obsidian (Learning Aloud Blog)learningaloud.comA personal case study exploring the setup, troubleshooting, and model tradeoffs of running a local Obsidian LLM Wiki.

A personal methodology case study reframing the LLM Wiki pattern as a 'single-brain knowledge catalyst' rather than an outsourced second brain. Details a 4-phase workflow (Explore, Learn, Index, Enrich) where human-authored notes maintain cognitive ownership and critical thinking, while the LLM acts as an automated sidecar for indexing, active recall, and quiz generation.
Karpathy's LLM Wiki Works. Here's What Broke When I Ran It at Real Scale. (Michael Tuszynski on The Cloud Codex)mpt.solutionsA six-month personal retrospective running a multi-hundred-page Obsidian LLM Wiki across four domains. Identifies three structural failure modes (internal drift between changelogs and reference docs, silent entity contradictions, orphan page accumulation) and details engineering remedies: a pre-commit drift linter, scoped per-domain sub-schemas, and a companion pgvector index for recall once index context exceeds 100k tokens.
LLM Wiki (Filippo Mameli on Mameli.Dev)mameli.devA practitioner guide and personal workflow detailing a local Markdown LLM Wiki setup based on Karpathy's pattern. Outlines a six-file vault structure (
_Wiki/withraw/,wiki/,outputs/,AGENTS.md,index.md, andlog.md), multi-format ingestion workflows using Obsidian Web Clipper, MarkItDown, and opendataloader-pdf, local search viaqmd, and a four-stage maintenance loop (collect, synthesize, query, and clean).
A detailed case study exploring the setup, best tools, incremental updates, and common failure modes of running a local LLM Wiki.
My Obsidian PKM Setup: Karpathy's LLM Wiki + Claude Code + Local Search (Alfonso Fortunato)alfonsofortunato.comA personal implementation case study detailing an Obsidian PKM setup maintained by Claude Code as a disciplined librarian. Features hierarchical index navigation (
_master-index.mdto topic indexes), four-verb workflows (clip, compile, query, audit), and integration withqmdfor local-first hybrid search.
Persistent Agent Memory with an Obsidian Wiki and Claude Code (Mike Moore on Web of Mike)webofmike.comA production case study detailing a two-layer Obsidian knowledge graph (per-account relationship vaults and a root concepts layer) maintained by Claude Code. Explores Pocket AI voice memo ingestion into immutable
_raw/storage, provenance steering with[UNVERIFIED]flags, 5-15 page fan-out targets per source, and parallel agent Git merging.
A production retrospective and practical guide detailing the operation of a 127-page LLM knowledge layer without vector databases. Outlines real-world operating numbers, architectural lessons across the three-layer setup, common failure modes in agent maintenance, and setup blueprints for business research and project workflows.
The Karpathy/Wyndo Wiki: How I Built a Knowledge Base That Maintains Itself (Lee Shand on Wired Differently Substack)shaggyrs6.substack.comA personal case study detailing the construction of a self-maintaining Obsidian wiki using Claude Code, Readwise sync, and a custom skill toolkit, exploring cold-start sparse connections, import command challenges, and LLM-tagging behaviors.
The Lint Ran Every Week. The Dead-Link Count Never Went Down. (Michael Tuszynski on The Cloud Codex)mpt.solutionsA follow-up case study analyzing 10 weeks of scheduled linting telemetry across a 334-page Obsidian LLM Wiki. Explores why automated detection failed to reduce dead links (which grew from 109 to 146) or orphan pages, tests automated fuzzy repair scripts (resolving 0 of 185 broken links), evaluates why auto-generated stub cards merely convert dead links into hollow orphans, and argues for budgeting human entity-triage decisions over automated checkers.

A comprehensive 3-month operations report detailing the maintenance of a 490-page, 880-commit LLM Wiki in Obsidian. Explores six discrete agent operations (ingest, update, query, lint, enrich, reorganize), deterministic 20-check pre-commit lint gates, split human-agent ownership models, and healthy page-count consolidation passes.

A detailed production review of an LLM Wiki operating over a 450-page Obsidian vault, outlining folder structures, pre-commit linter checks, and split-ownership boundaries.

Part 2 of a 4-part series detailing why the system instructions and formatting rules (the schema) represent the true value of an agentic wiki, explaining how to write rules that force agents to capture raw facts, structure page layouts, and log contradictions.

Part 4 of a 4-part series exploring how an external, flat-file Markdown wiki acts as a load-bearing memory layer that compensates for the small context windows of local on-device models, enabling private and cost-efficient knowledge work.

A personal retrospective by a CERN physicist compiling 3,300+ Kindle highlights across 141 books into an Obsidian LLM Wiki using Claude Code. Audits subtle hallucinated quotations and epistemic error cascades as text moves further from source evidence (10% error in quotes, 17% in book facts, 33% in syntheses), highlighting why mechanical deterministic verification scripts are essential over model self-checking.
This Open-Source App Turns Your Documents Into a Self-Building Wiki (Kristopher Dunham on Medium)medium.comA review and user experience case study of the nashsu/llm_wiki desktop app. Runs through the core components of automated document ingestion, entity graph linkage, and automated markdown linting, outlining a lightweight manual recipe for testing the compiler loop before installing local tools.

A follow-up case study on separating agentic reasoning from deterministic linting rules to prevent consistency drift.

A case study outlining a low-friction voice capture and structuring workflow. Details using mobile dictation interfaces to dump unstructured ideas into raw files, using local LLM agents to extract concepts and links, and rendering the results inside an Obsidian wiki.
What My Personal Knowledge Base Made Visible That I Couldn’t See (Robin Good Substack)robingood.substack.comAn essay exploring how converting personal writings into a local LLM Wiki exposes hidden cognitive patterns, expert positioning, and missing research areas.
