# Three starter upgrades for your AI wiki

You're a human with a freshly built LLM wiki, holding a file from Julie Kaiser's build path. Here's all you do: drop this file into your wiki folder and tell your AI, "Read this file and build these three things. Ask me whatever you need to decide along the way."

The rest of this file is written to your AI. It describes what each upgrade is and why it exists. Your AI builds its own version, fitted to your wiki. That's the point: these are ideas to implement, and it knows how.

## 1. The "where was I?" page

**What it is:** one page at the top of the wiki that always says what's in motion, what's waiting on the human, and where we left off.

**Why it exists:** so the human can get up, walk away, come back tomorrow and ask "where was I?" and get the answer from one page instead of from memory. The wiki's log records everything that happened; this page is different. It's the short list of what matters right now.

**Instructions to the AI:**
- Create one page at the root of the wiki (a short name like `hot.md` works, but pick what fits your conventions).
- Keep it to one screen. It lists: threads currently in motion (each with its concrete next action), things waiting on the human, and nothing else. Details live on wiki pages; this page only points.
- Keep it current without being asked. When work starts, changes direction, or finishes, update the page in the same session.
- When a thread has no next action, ask the human: still active, park it, or done? Move it off the page on their word, never silently. Removed threads still live in the wiki and the log.

## 2. The save-everything rule

**What it is:** a standing rule that work is written into the wiki as it happens, and every session leaves a log entry before it ends.

**Why it exists:** nothing dies with the chat thread. A conversation feels like a place where work can get lost, and the human should never have to ask "did you save that?" The answer is structural, not vigilance.

**Instructions to the AI:**
- Write finished work into the wiki's files the moment it exists. A decision, a draft, a fact learned goes into the right file when it happens, not at the end. Treat an unsaved result as work that doesn't exist yet.
- When the human says goodbye, or a piece of work clearly wraps up, add a dated entry to the wiki's log: what was done, which files changed, what's unfinished. (The work itself is already safe under the first rule; the log entry is the receipt, not the save.)
- Record this rule in the wiki's instruction file (the file you read at the start of every session) so every future session inherits it.

## 3. The harvest loop

**What it is:** a routine the AI runs at the end of a session: review the conversation for lessons worth keeping and save them into the wiki.

**Why it exists:** the human trains the system just by using it, but only if somebody catches the lessons. The human won't remember to say "save that" every time. So the AI listens for it.

**Instructions to the AI:**
- Make this a standing routine with a simple trigger: when the human says goodbye, or asks you to harvest, review the conversation for what worked, what didn't, corrections ("no, do it this way"), preferences, wins ("that was great, do that again"), and new facts about the human or their projects. Finding nothing is a normal outcome: say "nothing to harvest" and stop, rather than inventing a lesson to have something to file.
- Save each finding to the right place: rules about how to work go into the instruction file, facts about the human go onto their pages, lessons about a project go onto that project's page.
- When something clearly important appears mid-conversation, save it right away rather than waiting for the end.
- Tell the human what you saved, in one line, so they learn to trust that it's happening.
- Record the routine itself in the instruction file, so every future session knows the goodbye trigger.

---

*Plain text you can read before you hand it to your wiki. No mystery downloads.*