“I’m lost, what we need to work on?”
It wasn’t the first time I opened a session in Claude typing exactly that. So I went back through six weeks of my transcripts and counted: 8 times. And that was just the tip of the iceberg. In the same timeframe, I asked Claude to hunt (and fix) stale info in my docs 77 times.
Here is what those numbers tell me.
Output used to be the constraint: plans, docs, fixes, all slow because you had to invest time doing them, or you were waiting for someone else’s work, approval, confirmation.
Now AI lets anyone produce all of that fast. And that’s where we (me included) need to pay careful attention: AI generates and edits work faster than we can internalize its meaning and impact.
- If we don’t internalize the meaning, we don’t understand it.
- Without understanding, we don’t remember it.
- And without remembering it, we lack knowledge and learning.
The cause: AI has clouded our working and thinking environment. Like never before, we must know where things stand, which doc leads, what’s done, what’s stale, what’s pending.
AI literacy and training need to address the elephant in the room, so big yet so silent that many don’t see it: learning to use AI means doubling down on building systems that work as your North Star. Otherwise you’re lost in a sea of information you don’t know whether you can trust.
The new constraint isn’t resource availability or speed.
It’s orientation: knowing where you are, where you’re headed, and why you’re moving in that direction.
Treating orientation as its own system means automated checks, scheduled updates, documentation consolidation rituals. Here’s my approach:
1. One lead doc. Everything else links FROM it.
2. Done docs get archived with a pointer left behind.
Never deleted, never left lying around as live.
3. AI doc updates are untrusted: I audit them before
planning on them.
4. A /session-end command. Every session closes with one ritual:
update the session history, sweep the docs it touched for
stale info, archive what's complete.
5. Drift checks. Any doc that mirrors something live gets an
automated test that diffs the two.
6. A daily log. One plain-english line per session,
written at the end of each day.Rule 6 is the newest and my favorite. I run several AI sessions a day, often in parallel. By evening they blur in my head. Now a slash command sweeps the day and writes one line per session in a doc, using my own words. It helps me internalize the key things I did that day. It also marks the end of it: when I run /daily-log, work is done for the day.
None of this is documentation discipline for its own sake. It’s the price of speed and the cost of our memory being prone to error.
So what, Matteo?
- With AI, output stopped being the bottleneck. Orientation is the bottleneck now.
- Every doc your AI writes is a claim, not a record. Claims need verification before you build on them.
- Git covers the code. The decisions, dead ends, and parked questions need their own memory system, and it won’t build itself. You need to do it.
- Before adding a new skill.md from a random GitHub repo, build your system first to never lose your North Star.