The Folder Is the Memory

The opening question

Dirk opened with a practical problem: he writes a board update letter every quarter. He’d iterated twelve pages down to two with Claude and real-world board-member feedback, but felt uncertain about consistency, accuracy, and legal notes each time. Should there be separate writing and checking skills?

Lou’s answer came in two parts. First, the research foundation: feed in real board letters so the model has a feel for the register. Second, use the interview method — tell Claude what you’re trying to produce and say: Ask me the questions I need to answer to get this right. Go back and forth, cover the legal angles, refine the tone. When the output is good, say: Review everything we’ve done and turn that into a skill. Then — crucially — after every quarterly use, add: Learn from everything we did this session and update the skill. The skill gets smarter each quarter.

The process-first correction

Don Back brought two connected observations. He’d used AI to produce a governance brief for a not-for-profit board facing a three-year festering issue — loaded the bylaws, the governing Act, the meeting transcript, then ran a Socratic back-and-forth that surfaced both the framing and sub-issues he hadn’t consciously named. One session; a structured, credible document ready for the chair and secretary.

The broader pattern he’s seeing across the companies and PhD grads he advises: 2024–25 was the year of tool-first AI adoption — pick a platform, push everyone to use it, see what happens. What’s now emerging is the corrective. Document the process, optimize the process, then let that drive your selection and implementation of the tool. The tool shapes work around its defaults when it leads; the process shapes the tool when it leads. Scott’s framing from the chat landed perfectly: I have a hammer — let’s find nails is the exact trap.

Context architecture for high-stakes documents

Joanna is self-representing in formal complaints across Ontario Small Claims Court and several regulatory bodies. Her documents now exceed 150 pages. She’d been working through threads, building in citation-checking loops, but still felt 70–80% manual. Lou’s diagnostic: What are you trying to produce, and where do the citations exist?

The context problem first. Joanna was uploading files into conversations, which filled context with content rather than leaving room for intelligence. Lou’s reframe: Mentally separate the conversation from the library of documents related to your project. The folder is the memory; the conversation is disposable. Make a project folder on your desktop, drop every relevant document in, write a CLAUDE.md at the root that tells Claude what each file contains and where the regulatory URLs live. Use Claude Code, point it at that folder, and let Claude manage its own reading. When a thread gets long, start a fresh one, tell it to read CLAUDE.md, and you’re immediately back at full context without re-uploading anything.

The hallucination problem second. Two grounding mechanisms. NotebookLM: give it the URLs of the relevant regulatory pages, and it answers only from those sources, returning the reference for every claim. Claude Code with a ground-truth directive: add to CLAUDE.md — Consider this regulatory URL and its subfolders ground truth. Only provide regulatory information from that source. Division of labor: the locked source verifies which citations are real; Claude writes the argument.

Lou closed with a reframe worth keeping: It can already think — you’re just not using it the way that causes it to think yet. The shift is from Q&A to delegation. Here’s my case, here’s what I’m up against. What do I need to know, what questions do I need to answer, what positions should I take? Then: do that for me.

Opus plus Codex

Donald shipped a Kindle-style reader for his book manuscript using a workflow that landed with the group: Claude/Opus for planning and architecture, Codex for execution, with Donald asking Codex to report back at checkpoint gates and Claude to review the work. Twenty to thirty commits, all well documented. For the first time I feel like a real programmer.

Lou picked up the thread with his own discovery: pointing Codex at the same folder as Claude for a code review found half a dozen things Claude had missed. His tentative conclusion: one Claude subscription plus one Codex subscription is better than two Claude subscriptions. Claude wins on knowledge, planning, and brainstorming; Codex may outperform Sonnet on implementation once the spec is detailed enough — and comes with far higher usage limits.

What to try before next session

Run the interview-to-skill loop on your most repeated task. Pick the thing you do most repetitively with AI — a weekly report, a client follow-up, a research brief.

Open Claude Code, point it at the folder for that task, and say: I want to produce [X]. Interview me — ask me every question you need to answer to get this right. Work through the interview until the output is good. Then: Review everything we did and turn it into a skill. Use it once for real. Fix anything that’s off. Close with: Learn from the corrections we made today and update the skill.

You now have a skill that’s already better on its second run than it was on its first.