Key Takeaways

  • Web chat interfaces fail at long-term personal operations because context windows drift, memory features are opaque, and session resets wipe operating context.
  • Scott Farrar runs Claude Code locally against an Obsidian vault, giving the coding agent direct read and write access to plain-text Markdown files on his hard drive.
  • The setup lets Farrar dump raw, unstructured details like appliance serial numbers and warranty PDFs into the vault, while the coding agent organizes and cross-references the records.
  • Storing the data in local Markdown files prevents vendor lock-in, keeps family data private, and allows spouses to edit the vault directly or sync it to displays like DAKboard.

The Method

Most people interact with LLMs through a browser window. You paste a chunk of text, ask a question, and get an answer. When you close the tab, the working context disappears. Even with hosted memory features enabled, you cannot see what the model retained, what it discarded, or why it hallucinated an answer three weeks later.

Farrar, a math educator, bypassed web interfaces entirely for running his home. He pointed Claude Code at a local folder managed by Obsidian. Instead of treating the model like a conversational companion, he treats it as a software engineer maintaining a database of plain-text files.

“I wanted to build a system where I was confident that AI was grounded in my truth and knowing what it knows about me so that I could be confident in what advice it was giving me back,” (play from 3:45) Farrar explained.

His workflow follows three steps:

First, Farrar creates a single directory of plain-text Markdown files inside Obsidian. This vault holds dedicated files for home maintenance schedules, insurance policies, contractor notes, and appliance warranties.

Second, he launches Claude Code in his terminal and directs it to that folder. When an issue comes up, such as comparing auto insurance quotes or tracking down the warranty on a broken dishwasher, Farrar asks Claude Code to inspect the existing files. If new information arrives, he pastes the unstructured mess directly into the terminal.

Third, Claude Code reads the repository, executes the requested analysis, and writes the results directly back to the Markdown files on disk. As Farrar put it: “I built this system that would read and write files, read and write markdown files, so they would have the memory of what we were talking about, and it would be living on my computer so that if I come back later and start a new Claude Code session, it can reference those same files.” (play from 4:09)

Host Claire Vo highlighted the core advantage of this structure over hosted chat apps: “The problem that you're trying to solve here really is when you're using like the web ChatGPT or Claude, you're like, what does it really know about me? What does it really remember? What information does it have? That is very opaque to most people.” (play from 5:00) With local files, the user owns the index. If something is wrong, you open the Markdown file, edit the sentence, and the agent's memory updates instantly.

Farrar also uses the agent to bridge the gap between messy inputs and neat documentation. “I was able to dump in a lot of unstructured information and rely on the coding agent to structure it and maintain that structure,” (play from 13:41) he noted.

Where This Breaks Down

This workflow requires technical comfort that ordinary family members rarely have. Running terminal commands, managing directory permissions, and resolving git merge conflicts are natural for a developer, but asking a non-technical spouse to fire up a command-line interface to check when the furnace filter was last replaced will fail immediately.

Farrar solves part of this friction by keeping the files in Obsidian, where anyone can read normal text or push updates to a shared family dashboard like DAKboard. But the moment the agent makes an unauthorized file edit or misfiles an insurance rider, someone has to audit git diffs. If you do not already live in a terminal, running a coding agent to track household chores creates more maintenance overhead than it removes.

What to Do With This

Create a dedicated folder on your computer named household and open it inside Obsidian. Drop your current lease or mortgage summary, your primary insurance declaration page, and a text list of your major home appliances into that folder. Open your terminal, run Claude Code in that directory, and instruct it to extract all replacement schedules and policy deductibles into a new file called maintenance-calendar.md.