Automation & Agents

Make AI Write Your Knowledge Base After Every Session: A One-Page-Per-Project Wiki System

Follow this for 30–60 minutes and you'll have a system where AI automatically distills your work into notes every time you finish a session — from then on, looking things up means flipping a page, not re-asking the AI or digging through folders. Zero cost, using whatever AI tool you already have (Claude Code, Cursor, or ChatGPT all work) plus any note app (Obsidian, free, recommended).

09/15/2026 ｜ 3 min read

One Core Idea

Conversations are one-off; knowledge should compound. An AI chat log is not a knowledge base — it has no structure, isn’t browsable, and goes stale. Three things are actually worth pulling out on purpose: (1) where the output lives and how to use it, (2) why you decided what you decided, (3) what pitfalls you hit. Everything else — progress, todos, the back-and-forth of the conversation — should stay out, or the knowledge base turns into a second messy log.

Step by Step

Step 1 | Open a dedicated area

Create a folder in your note app, e.g. ai-projects/. Keep it separate from your reading notes or journal — this area holds only knowledge about “projects AI helped you with.”

Step 2 | Build one index page

Inside that folder, create a README.md (or “Overview”) with a table: note name | last updated | one-line summary. From then on, you (and the AI) check this page first, and only open the individual note if it’s a hit.

Step 3 | Lock down a note template

One page per project, always the same five sections (copy this):

# Project Name

**One line**: what this project is, for whom.

## 1. Goal & Positioning
## 2. Output Map (where deliverables live, how to use them)
| Output | Path | Purpose / How to use |
## 3. Key Decisions & Rationale (including what was rejected, and why)
## 4. Pitfalls (symptom → root cause → how to use it next time)
## 5. Related Links

> Progress and todos are NOT written on this page (tracked elsewhere).

Step 4 | Give the AI a end-of-session rule

Every time you finish something substantive with AI, hand it this (save it as a snippet):

“Write the reusable knowledge from this session into my ai-projects/ folder: use my five-section template; one page per project — update the existing page if there is one, never create a new file for the same project; only write knowledge that will still hold true in three months, no progress or todos; update the README index table when done. If this session was just Q&A with no substantive knowledge, just tell me ‘nothing to save’ — don’t force it.”

Step 5 (Advanced) | Make it fully automatic

If your tool supports custom instructions or memory (e.g. Claude Code’s CLAUDE.md or a skill), bake Step 4’s instruction into it and bind it to your “wrapping up” trigger phrase — you won’t even need to paste it again. If you don’t know how to set this up, just ask the AI (see the rescue prompt at the end).

Step 6 (Optional, cost-controlled) | Backfill old projects

Want to bring past projects in? First have the AI just list them (“scan what projects we’ve done together, just list them, don’t write anything yet”), pick the ones you want, then fill them in one at a time — and explicitly tell it which files to read (a whitelist). Never allow “re-read the entire history” — that’s slow and expensive.

Pitfalls You Will Hit

1. Pitfall 1: AI stuffs progress logs into the notes. Things like “finished X today, doing Y tomorrow” leak in, and the knowledge base is worthless within a week. Fix: keep the template’s closing line “progress and todos are not written here,” and tell it to delete anything that leaks in.

2. Pitfall 2: AI creates a new file every time. The same project ends up with v1, v2, and final notes — which is the same as having no notes. Fix: never drop the rule’s wording “one page per project, update the existing one, never create a new file.”

3. Pitfall 3: backfilling re-reads everything and burns through tokens. Fix: use Step 6’s whitelist — explicitly say “only read these 3 files,” nothing else.

4. Pitfall 4: trying to save everything makes the notes worthless fast. Pure Q&A or small tweaks aren’t worth saving. The rule’s “just say nothing to save” line exists specifically to block this — better sparse than bloated.

Universal Rescue: How to Ask AI When You’re Stuck

If you get stuck on any step, don’t guess — hand the situation to an AI (Claude / ChatGPT / Gemini all work) with this template:

I'm a beginner following a guide to build an "AI auto-writes my knowledge notes" system,
using [Claude Code / Cursor / ChatGPT / ...] as the tool and [your note app] as the note app.
I'm stuck at [paste the step you're stuck on], and here's what's happening: [paste the exact text / describe the screen].
Please walk me through it one step at a time, and wait for me to confirm before moving on.

One-Line Summary

Lock the rules, fix the template, one page per project — the knowledge base grows itself. More AI-workflow recipes coming; follow along so you don’t miss them.