AI Wiki and Context Farmers — How a Solo Founder Builds Living Memory in 2026

The problem most founders are still solving wrong

You open Claude. You explain who you are. You explain your business. You explain your offer. You explain your ICP. You explain the project you're working on.

You get an answer. You close the tab.

Tomorrow you open Claude again. You explain who you are. You explain your business. You explain your offer.

This isn't a productivity problem. It's a memory problem. The model has no idea who you are between sessions. Every conversation starts at zero.

Most people patch this with a "magic prompt" — a giant block of context they paste into every session. It works for a week. Then the prompt gets out of date. Then it gets too big to maintain. Then they stop using it.

The fix is not a better prompt. The fix is durable context files plus farmers that keep them current.

What an AI wiki actually is

An AI wiki is a folder of .md files that describes you, your business, your offer, your audience, and your decisions in a form an AI can read at the start of every session.

It's not for humans. Humans don't need to read it. It's a standing-context layer for Claude, ChatGPT, Gemini, or any other AI you give access to.

The wiki has three layers:

1. Vault — your private working memory. Notes, drafts, raw transcripts, internal logs. Nobody reads this except you and your agents. It changes every day.
2. Brand wiki — curated source-of-truth. Your brand identity, voice, offer architecture, audience, milestones, proof. Stable, sanitized, internal. Updated when something meaningful changes.
3. Public wiki — sanitized, publishable, crawlable. Build-in-public pages that double as SEO and proof. Updated when proof exists.

Vault feeds brand wiki. Brand wiki feeds public wiki. Each layer answers a different question. Each layer has different sanitization rules.

Together, the three layers are what I call the AI memory. It is the production layer that everything else gets built on top of — the book, the funnel, the content, the follow-up. Without it, every asset starts from a blank chat.

What context farmers do

A farmer is a small program that watches a source — your inbox, a Substack, a YouTube channel, a Skool community — and pulls new content into your vault automatically.

The farmer captures. You curate. The agent uses the curated layer.

You don't sit at the keyboard processing every email or transcribing every video. The farmers do that overnight. When you open your laptop in the morning, the new material is already in the right folder, classified, scored, and ready to synthesize.

This is the part most "AI agency" guys leave out when they sell you a course about running an agency. They sell you the agent. They don't sell you the farmers that keep the agent fed.

The 1HBP stack (real, running, today)

Three farmers running on a VPS, capturing into a vault that's read by three agents.

Farmer 1 — watches my inbox for new emails from a few creators I follow and learn from. Classifies + routes them to the right folder with a confidence score, so a launch or a new framework never sits buried in the inbox.

Farmer 2 — watches Skool and YouTube for new videos in tracked channels. Pulls transcripts and metadata. Drops them into the right folder for synthesis.

Farmer 3 — watches a few Substacks I read regularly. Captures new posts. Scores them for relevance.

All three run nightly on cron. I don't touch them.

The agents read the vault, plus a high-fidelity voice profile (compiled from a 100-question voice interview). They start every session by reading the standing-context files. They ask clarifying questions before they execute.

This is roughly the same pattern Brad described in his "self-improving AI wiki" Loom, the same pattern BraveBrand teaches under "Brand Wiki / Intelligence Wiki," and the same pattern Ruben Hassid lays out in "you're just a text file" and "Magic." Three independent sources arriving at the same architecture.

The architecture is converging because it's correct. Context files plus questions plus human taste beat prompt libraries every time.

How to start

You don't need a VPS. You don't need three farmers on day one. You need three things:

1. A .md voice profile. Run a voice interview on yourself — 50 to 100 questions. Compile the answers into a high-signal context file. Keep it under 5,000 tokens. The Voice Compiler pattern from Ruben Hassid is the cleanest version of this I've seen.
2. A .md brand identity file. Who you serve. What you sell. What you refuse to do. What you sound like. What you sound like NOT.
3. One read-then-ask habit. When you start any meaningful Claude session, paste a one-line instruction: "Read these files completely before responding. Then ask me clarifying questions before executing." That single line collapses 80% of bad AI output.

That's the minimum viable AI wiki. One context file. One brand file. One habit.

Add farmers when capture volume becomes the bottleneck — usually after a few months of trying to track sources manually.

Add the public wiki layer when you have real proof to share — working systems, shipped artifacts, sanitized screenshots. Public wiki pages double as SEO assets that compound over time.

What I'd do differently

Two things, in retrospect.

I'd start the voice profile sooner. I waited until question 22 of my own interview before realizing the compiled file mattered more than the raw answers. The interview is the raw material. The compiled file is the deliverable.

I'd skip the prompt libraries entirely. Most of mine sat in a notebook unused for months. The agents only got useful when the context layer existed — at which point most of the prompt-library work was redundant.

If you're starting today, write the voice profile first. Build the prompt library never.