←Back/How I AI
How I AI

Claude Code and Obsidian: Scott Farrar’s Household Brain

Scott Farrar uses Claude Code, Markdown files, and Obsidian to manage appliance records, insurance questions, and household repairs. Here’s what makes the system useful—and where he deliberately keeps himself in charge.

Featuring Scott Farrar · Math educator and parent

Claire Vo's profile picture

Claire Vo

November 16, 2026·10 min read

Episode chapters

  1. 2:34Replacing forgetful chats with household recordsScott explains why he moved household context into Markdown files that Claude Code can read and update, then shows how that supports insurance questions.
  2. 8:22Appliance photos and inspectable memoryAn Obsidian vault connects appliance records, source photographs, warranties, and project files so new conversations can retrieve household-specific context.
  3. 14:08A shared to-do list without household pull requestsScott turns a family brainstorming session into a prioritized task file and explains the trade-offs between Dropbox and GitHub for sharing.
  4. 19:02Making tasks and wins visible on DAKboardA wall display surfaces projects, trips, suggested tasks, and completed work; Scott describes the behavioral impact and starts an earthquake-readiness idea.
  5. 26:14Keeping students responsible for the thinkingScott connects his preference for human agency to math education and describes a chatbot that asks students to teach it.
  6. 31:04When persistent memory gets messyScott explains how he handles unhelpful sessions and a to-do file that grew too verbose despite his instructions.
Episode outline

Scott Farrar says he has run out of pending household fires. Not because his house stopped needing work, but because he finally sent the emails, asked the questions, and moved the projects forward. Now, more often, he’s waiting on someone else.

That is a more interesting result than an AI agent ordering something automatically. Scott built a system that gives him work to do—and makes that work specific enough to finish.

A math educator and parent, Scott was frustrated by a familiar uncertainty in AI chat: What does it actually remember? Did it retain the insurance details? Is its advice grounded in his house, or in a plausible-sounding generic household?

His answer was a folder of Markdown files that Claude Code can read and update, with Obsidian providing a way to inspect the records. Appliances, warranties, project notes, and next steps live outside any single conversation. The clever part is less the storage format than the division of labor: Claude organizes the context and proposes the next move; Scott remains responsible for deciding and acting.

Give the next conversation something real to remember

Scott’s starting requirement was straightforward:

“I wanted to build a system where I was confident that AI was grounded in my truth.”

Instead of repeatedly pasting household context into a chat, he used Claude Code to read and write local Markdown files. A new session can reference those same files rather than requiring him to reconstruct the previous conversation.

Obsidian makes that memory visible. Scott can open a project file, inspect a checklist, or look at the graph connecting his household information. But he doesn’t spend every interaction navigating the vault. Often, he simply opens Claude Code and asks a loose question.

The appliance inventory shows how he got useful context into the system. He asked Claude what information he should gather about his house. It suggested taking pictures of every appliance. Those pictures became source material for an appliance list, including identifying information such as serial numbers.

The fridge example is painfully practical. Scott can start a new conversation with “what’s going on with the fridge?” and have Claude consult the existing records. In the example he described, it can bring back warranty information, including a warranty ending in 2027, because that information was entered previously.

Organizing household appliance records in Obsidian, showing a visual representation of a stove and detailed note categories.

This is not a demonstration of AI independently diagnosing a refrigerator. It is a demonstration of retrieving household-specific context before making a plan. That distinction matters. A remembered warranty date is useful; a confident repair recommendation still needs checking.

For someone copying this, I would start with one troublesome appliance, its identifying photo, the warranty, and any repair history—not a comprehensive digital model of the house. That is an adaptation of Scott’s approach, rather than a setup sequence he demonstrated. The test is whether a fresh session can find the right facts without you retelling the story.

Insurance became a sequence of questions, not a document pile

Car insurance was one of Scott’s clearest examples of the system earning its keep.

He put insurance documents into the folder and worked through quotes with Claude. The project file retained the state of that work: what information had arrived, what remained unresolved, and what questions needed answers before he could make a decision.

That persistent state made a vague question useful. Instead of asking a new chatbot to explain car insurance from scratch, Scott could ask what to do next on the project. Claude identified the blocker: email the agent and ask specific questions.

The demo used the name Dana Whitfield, which Scott explicitly said was a fake name. The important artifact was the checklist, not the contact. It turned “I should probably deal with insurance” into a concrete next action.

Scott’s claim was not that Claude proved he had found the cheapest policy. It was that he could read the documents with AI assistance, combine them with his household context, and ask the agent much more detailed questions. Previously, he would have depended more heavily on the agent’s interpretation because he would not have read everything closely himself.

That is the part worth copying: use the model to improve the questions you bring to a professional. Ask it to identify unanswered questions and the information needed for a decision, rather than treating its recommendation as the decision.

For usability, I would also ask it to tie each important claim back to the relevant document and flag anything it cannot substantiate. Scott did not demonstrate that exact prompt, but it follows from his central requirement: being able to inspect what the advice rests on.

The boundary is important here. The system can help Scott understand coverage and maintain momentum. The episode does not establish that it can verify every policy interpretation or make the final coverage choice for him.

Let the agent organize the mess—but inspect what it writes

You do not need to love filing to benefit from files. Scott’s system sits somewhere between meticulous organization and dumping a pile of information into a folder.

He described relying on the coding agent to structure unorganized inputs and maintain that structure. One particularly useful input was a brainstorming session with his wife: every household task they could think of, large and small.

That included the tub spout in the kids’ bathroom and a possible multi-thousand-dollar dry rot problem. Crucially, the latter was still a question: was there a problem, and how would they find out?

Scott asked the system to compile the tasks with priorities, cost estimates, and references. The resulting to-do file became one of the records he returned to manually most often. A list that mixes small repairs with uncertain, expensive projects needs judgment about urgency; a generated estimate is a planning aid, not a contractor’s quote.

The system also has a failure mode familiar to anyone who has asked an AI to maintain a document: it keeps adding words. In an earlier version, Claude began putting updates into the to-do table until the table became unwieldy, despite instructions to keep it brief.

Scott added a linting script to help enforce the structure and constrain verbosity. He also asks why the model made a mistake, partly to discover whether a bad assumption has been written into memory. Sometimes he simply closes the chat and starts a new session.

That is the maintenance cost of persistent AI memory. A bad answer in a disposable conversation can disappear when you leave. A bad fact or sprawling update written into your files can affect the next conversation, too.

Sharing added another constraint. Scott wanted his wife’s Claude sessions to access the same household knowledge. They first used a shared Dropbox folder copied to both laptops. The risk was simultaneous edits and sync conflicts. GitHub was another option, but making household collaboration depend on a coding workflow was not an obvious improvement.

They landed on a compromise: Dropbox handles much of the shared material, while Scott uses GitHub for some scripts and related pieces. They generally avoid editing simultaneously.

There is a useful product lesson here: occasional collaboration does not automatically require real-time collaborative infrastructure. It also does not justify asking your spouse to submit a pull request to manage the refrigerator.

Scott initially scoped the system to household information, excluding medical information and finances. Later, he mentioned financial-adjacent planning tasks such as retirement consolidation. Anyone adopting this approach should make their own boundary explicit. Local files make the stored records inspectable; that alone does not settle every question about sensitive information or how an AI tool processes it.

Put the next action somewhere people will see it

An Obsidian graph is interesting to Scott, who likes graphs. It is not necessarily a useful interface for everyone else in the house.

To get information off his laptop, he used DAKboard, a wall display that can include a custom web page. Claude built a script to extract tasks, upcoming trips, and other information from the household records into a display designed to catch someone’s attention while walking past.

The display surfaces current projects, things that need attention, a suggested task, and recent wins. That last category is a deliberate choice: finished work should be visible alongside unfinished work.

A detailed look at a personalized DAKboard setup, showcasing tasks, projects, and reminders for efficient household management.

Scott described the effect in behavioral terms. Checklists no longer disappear into an internal “I should do that.” The system retains them, and later conversations can bring overdue items back into view. He gave the example of being reminded that he had been waiting on something since September 24th.

The wall display is not essential to the underlying architecture. Its contribution is visibility. A perfectly maintained file cannot help much if nobody opens it.

The same system also accommodates ideas that are not ready to become commitments. During the episode, Scott asked, “Hey, do I need to plan for earthquakes?” Claude could draw on existing house context and capture the topic as an idea. Scott explained that ideas and full projects have different places in his system; projects collect information and checklists, and completed projects can move into the wins.

My suggestion of recording a narrated video walkthrough of the house was an idea to try afterward, not a demonstrated workflow. Scott’s demonstrated intake method was appliance photography and accumulated household records. Those should not be conflated.

The useful boundary: AI prepares the work; people own it

Scott’s views on math education explain a lot about his household setup. He wants AI to support agency rather than take it away.

At home, that means surfacing recommendations while leaving him responsible for action. With students, it means preserving the cognitive effort of learning. He described flipping the usual tutoring relationship: instead of a chatbot explaining a linear equation, the chatbot asks the student to explain it, then follows up with questions such as why subtracting from both sides works.

That discussion was a teaching approach, not a live product demonstration. Readers interested in his related work can explore Algebuds: Student Tutors Chatbot.

The connection is practical. Sometimes a useful assistant should make you more capable of doing the work, rather than remove you from it. Scott wants the household system to help him understand what needs doing and why. He wants a tutor to draw out what a student knows, without supplying every answer.

The verdict: copy the persistent, inspectable records; the project-specific next actions; and the habit of showing completed work. This approach is best for households with recurring repairs, documents scattered across folders, and someone willing to maintain a lightweight system. Scott’s household assistant write-up is a useful starting point.

Do not mistake the vault for verified truth merely because it is readable. Someone still needs to check extracted facts, resolve conflicting information, judge estimates, and make decisions about insurance, repairs, and safety. Scott’s strongest result was not an autonomous home. It was a household manager who could finally see the next action—and take it.

Thanks for supporting How I AI

Sponsors

Build your next product with ChatPRD

Turn an idea into a PRD, user stories, and a plan.

Try ChatPRD free

Start shipping better products.

Join 250,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.

Free to start · 14-day Pro trial · SOC 2 Type II