Context
Scott Farrar is a math educator and a parent who juggles kids, a career and a house. He found that complex questions in an AI chat, like whether his car insurance was right, left him unsure what the AI actually knew or remembered as the conversation went on. So he moved his household life into Claude Code, where the agent reads and writes a folder of Markdown files that he views in Obsidian. In this How I AI episode, host Claire Vo has him demo the system and talk about AI and kids learning math. The episode is about home life, but the structure of the system (a visible, file-based memory, project files with checklists, and a shared display) is a pattern PMs can reuse for any long-running work they run with an agent.
The Big Idea
An AI is only as trustworthy as the context you can see. Keeping that context in plain Markdown files that the agent reads and maintains turns loose questions into grounded, actionable next steps.
Scott can ask something as vague as "what should I do next?" and get a specific answer, such as which person to email to unblock his insurance decision. That works because every session reads the same files, not a chat history he cannot inspect.
Key Insights
Chat memory is opaque
Scott's starting problem was trust, not capability:
- In a long chat, he had to decide whether to dump all his information in, and then wonder whether the conversation was drifting.
- He could not tell what the AI really remembered about him, so he could not be confident in its advice.
- Claire agreed this is opaque to most people using AI on the web.
His fix was to switch from a chat interface to a coding agent. In Claude Code, the agent reads and writes Markdown files on his computer, so a brand-new session can reference the same files as the last one.
Files make memory inspectable
Because the memory is a set of Markdown files, Scott can open any of them and see exactly what Claude is working from. Claire described this as a "very transparent source of truth." In practice he rarely reads the files directly. He usually starts a new Claude Code session and asks a loose question, such as what is going on with the fridge, and the agent looks up the relevant files itself. The point is that he could check, which is what gives him confidence in the answer.
Ask the AI what to capture
When Scott set the system up, he asked Claude what information he should gather about his house. Claude told him to walk around and photograph every appliance. Those photos now sit in the vault and are summarized in an appliances list with serial numbers. Other entries include device details, such as which model and OS version a kid's iPad runs, so he can ask when it will lose software updates. Later, a fridge question came back with the fact that its warranty runs out in 2027, because he had logged that earlier.
Grounded context sharpens negotiations
The car insurance project showed how the system changes a real decision:
- Before, Scott would have trusted his insurance agent and not read the policy documents in detail.
- Now, Claude reads the quote documents and meshes them with his own home context.
- It summarizes the quotes, flags gaps in coverage, and lists the questions he needs answered before making an informed decision.
- When he asked a loose "what should I do next?", it named the blocker: email the agent with a specific question.
The result was that he could ask far more detailed questions and know exactly what he was asking and what he was being told.
Dump loosely, let the agent structure
Claire joked that people split into very organized types and "type B" people like her who just dump in unstructured context. Scott said the system works across that spectrum. He dumps in a lot of unstructured information and relies on the coding agent to structure it and keep it structured. Claude enforces the rules of how the vault is organized, and agents like Claude Code or Codex can go through a directory of long-ignored files and organize them.
Checklists change behavior
Claire asked whether the system actually changes what he does. Scott said yes:
- Next steps are recorded in a checklist, not left as an "I should" thought in his head.
- In any chat, the agent may remind him that he is overdue, for example that he has been waiting on something since September 24th.
- He has moved forward on so many projects that he has run out of pending fires.
- He described the new state as the ball always being in everyone else's court: he has sent the email and is waiting on them.
Keep the human doing the thinking
Scott contrasted his system with agents that go out and do things for you. He kept the doing role for himself: he wants to know what the AI suggests, but he still wants agency. He applies the same idea to math education, where he wants AI to keep students doing the cognitive effort and productive struggle instead of handing over answers. Claire said she prefers the opposite for herself, letting agents act, which shows this is a design choice that depends on the user and the goal.
Mental Models & Frameworks
The household brain structure
Scott's vault has a few distinct file types, each with a job:
- Reference files: facts about the house, such as the appliances list built from photos, serial numbers, warranties and device details.
- Idea file: a backlog of anything he thinks of, such as earthquake readiness, re-landscaping or repainting. Ideas are not yet full projects.
- Project files: where information is collected, checklists are started and items are checked off, including the current blocking step.
- To-do file: one big list of household tasks, each with a priority level, a cost estimate and a reference.
- Claude instructions: a Markdown instruction file that tells the agent the structure, for example that the to-do list should stay short.
Idea, project, then win
New things follow a small lifecycle. When Scott asked whether he needed to plan for earthquakes, Claude suggested capturing it as an idea first, since it had not yet been promoted to a full project. A project file then holds the research and checklist. When a project is done, he closes it and moves it to the wins column, which also shows on the wall display. He said he really loves closing a project file.
A linter for agent verbosity
Chat agents tend to over-write. Scott runs a linting script that checks the vault to make sure agents have not let files grow too verbose, such as a to-do list bloated with detail. It is a mechanical check on the agent's output, kept alongside the instructions rather than relying on the instructions alone.
Brainstorm everything, then triage
Scott and his wife held one session to name every to-do in their heads, big and small, from a tub spout that does not spray right to a possible multi-thousand-dollar dry rot problem. Claude then compiled the list and added a priority level, a cost estimate and a reference for each item. This to-do file is the one he comes back to most.
Trade-offs & Nuance
Dropbox versus GitHub syncing
Scott wanted his wife to talk to her own Claude and have it know the same things. They tried a few approaches:
- Dropbox: a shared folder copied to both laptops. Claude warned this could cause sync conflicts if both of them were in it at the same time.
- GitHub: Claude suggested committing the Markdown files to a repo instead. His wife was not sure it was worth the effort.
- Where they landed: a mix. Scripts, like the linter and the wall display export, live in GitHub. Most of the household content stays in Dropbox, because they rarely read and write at the same moment.
Claire framed the real question as how real-time the collaboration is. If it is only lightly real-time, occasional conflicts are an acceptable cost for a simpler setup.
Scope limits on sensitive data
At setup, Scott gave Claude his constraints: collaborate with his wife, scope the system to the household, and leave out finances and medical information. Naming what stays out of the system was part of the first prompt, not an afterthought.
Human in the loop versus agents
Scott keeps himself in the loop and lets the AI recommend. Claire prefers agents in the loop and lets them act on her behalf. Neither is presented as correct. The choice shapes how the system is built: Scott's design produces checklists and reminders for him to act on, rather than an agent that acts for him.
Patient tutor versus warm demander
Claire described AI as an infinitely patient private tutor that lets learners ask their most basic questions without judgment. Scott questioned whether a tutor should always be infinitely patient. He suggested it could also be a "warm demander," politely saying that the student already knows this and should give it a try, rather than doing all the work.
Practical Application
Keep agent memory in files
For any multi-week effort you run with an AI agent, such as a vendor evaluation or a launch plan, keep the context in Markdown files the agent reads and updates, not in one long chat. Start each new session from the files, so you can open them and check exactly what the agent knows before trusting its advice.
Give each project a blocker
Mirror Scott's project files at work:
- Create one file per project with a checklist of open items.
- Have the agent keep the current blocking step at the top, such as who you need to email and with which question.
- Ask a loose "what should I do next?" at the start of a session and see whether the file makes the answer specific.
Separate ideas from projects
Keep an idea file for anything not yet committed, and only promote an idea to a project file when you start working on it. Close project files when they are done and log them as wins. This keeps the active list short without losing half-formed ideas.
Ask the agent what to capture
Before building a knowledge base for a new domain, ask the agent what information it would need, the way Claude told Scott to photograph every appliance. Then gather that, and let the agent summarize it into reference files.
Add a mechanical verbosity check
If an agent keeps bloating a shared file despite instructions, add a simple script that checks file length or format, as Scott did with his linter, instead of only rewriting the instructions.
Questions to Consider
- When you get advice from an AI chat about a long-running project, can you see exactly what context it is using, or are you trusting a conversation history you cannot inspect?
- Which recurring work on your team, such as vendor reviews or launch prep, would benefit from a file per project that names the current blocker?
- If two people on your team shared one AI knowledge base in Markdown files, how real-time would their edits be, and would a simple shared folder be enough or would you need version control?
- Where do your team's ideas and active projects sit in the same list, and would separating them, with a clear point where an idea gets promoted, make the active list easier to act on?
- Does your team have a visible "recent wins" record, like Scott's wall display section or the wall of wins Claire's team used at Optimizely, or do finished projects simply disappear?
Bottom Line
Scott Farrar trusts his household AI because its memory is a set of Markdown files he can see, organized into reference files, ideas and project checklists that Claude Code maintains. The same structure, visible context plus a named next step plus a record of wins, is what turns vague questions into progress at home or at work.
Case Studies Mentioned
Car insurance shopping with context
Scott was getting several car insurance quotes. Claude read the quote documents he saved in the folder, combined them with his home context, summarized the quotes, and pointed out coverage gaps. When he asked what to do next, it named the blocker: email the insurance agent with a specific question that he needed answered before deciding. The lesson is that grounded context lets a non-expert ask much sharper questions of an expert, instead of simply trusting them.
The DAKboard wall display
Scott's graph views in Obsidian were not readable for the rest of his household, so he used DAKboard, a wall display that can show a custom web page. Claude built a script that pulls to-do items, upcoming trips and other data from the vault into a page with:
- Current projects and upcoming trips.
- A small suggestion for something to do today, such as a $12 fix.
- Longer-term items that need attention.
- A section for his kids.
- A "recent wins" section, such as fixing the water heater in one day.
The goal was to get the information off his laptop and in front of everyone, so a passing glance prompts action.
Optimizely's wall of work
Claire recalled that at Optimizely, her team stood in front of a physical "wall of work" every Monday, with every epic printed out and moved across the board, plus a wall of wins. She compared Scott's wall display and recent-wins section to that same habit of making work and progress visible.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Claude Code | Anthropic's coding agent, which can read and write files on your computer | Scott uses it as the interface to his household brain, reading and writing the Markdown files across sessions. |
| Obsidian | A note-taking app for Markdown files with a graph view | Scott uses it to view and inspect the vault, including a graph of all the context Claude can reference. |
| Dropbox | File syncing and sharing | Scott and his wife share most of the household folder through Dropbox so both of their Claude setups see the same files. |
| GitHub | Version control and code hosting | Claude suggested it to avoid sync conflicts, and Scott uses it for scripts like the linter and the wall display export. |
| DAKboard | A wall display that can show calendars and custom web pages | Scott shows a generated page of priorities, trips and recent wins on it for the whole household. |
Concepts to Explore
Zone of proximal development
Scott mentioned the zone of proximal development, which he calls the "zone of frustration" with his kids: the space where a task is hard enough that the learner is struggling but can still get there with support. He raised it to ask how patient an AI tutor should be, and whether it should sometimes push the learner to try.
The flipped chatbot tutor
Scott has explored a tutor that reverses the usual roles. Instead of teaching the student, the chatbot asks the student to teach it, for example "help me understand how to solve a linear equation," and then asks why each step is needed. The aim is to draw out what students already know, since many are afraid to share their ideas with people, while keeping them doing the thinking.
People to Follow
Scott Farrar
Scott is a math educator and parent who built a household knowledge system in Claude Code and Obsidian. He cares most about student agency in math education and designing AI that supports students' thinking instead of short-circuiting it. He shares his work at sfxmath.com and said he is always eager to talk about math education.
Notable Quotes
"I wanted to build a system where I was confident that AI was grounded in my truth." (Scott Farrar)
"I want to know what you're going to suggest and recommend, but I'm going to be the one having some agency still." (Scott Farrar)
"You just dump in unstructured context and you're going to get the next token out." (Claire Vo)
