# I wrote the filing system for AI agents in 2005. I just called it a website.

- Link: https://jcody.com/blog/filing-system-for-ai-agents-2005/
- Published: 2026-10-26T11:55:00-04:00
- Author: Jeffrey Cody

So, funny story. I am ramping up on Claude with my home server, and the more I do,
the less I sleep, and the more I do. Somewhere in there I had notes and documents
piling up in folders all over the place, and I got to a point where I asked my AI
agent a real question: _where does this note belong?_

Then I did something odd. I went and found a PowerPoint from the summer of 2005.

It was my plan for rebuilding a corporate website. I was not a writer then and I
still am not. What I had was a process for telling a marketing team how to organize
and write well. It had three parts.

## Levels, Tiers and one Source of Authority

**Levels (L1 to L6).** Every page lives at one level: homepage, product line, product,
application, feature, detail. The top levels are almost all message. The bottom 
levels are almost all detail. I even had a table for the ratio: 90/10 marketing-
to-detail at the top, 5/95 at the bottom.

**Tiers (T1 to T5).** Every page tells its story five ways: intro, benefits, examples,
technical, call to action. Same fact, five angles, repeated on purpose so each reader
lands on the version they need.

**Source of Authority.** One page is the definitive answer. Everything else points
to it. Nobody copies it.

Add one more rule: a holding area for **Orphan Pages**, the things that do not fit
yet and need a decision instead of a bad guess.

## Same outline, new reader

 * My agent’s memory index is one line per entry. The real documents hold the tables
   and numbers. That is the 90/10 ratio.
 * One idea gets written up as a process doc, a note, a future item and a maintenance
   item. That is tiers: the same fact from several angles, each reinforcing the 
   others.
 * One file is canonical and the rest point to it. That is a Source of Authority.
 * “Does not fit yet” gets its own list. Those are the orphans.

My [video archive work](https://jcody.com/blog/safe-archival-compression-reclaim-space-keep-the-raw/)
follows the same logic. A bad file gets logged, parked, and the run moves on. The
original is never touched. Nothing is deleted until the new copy is verified.

None of this is new. It is what a good outline has always done. The only thing that
changed is the reader. In 2005 the reader was a prospect. Now it is a machine that
has to find the right paragraph in a pile of documents.

## The cheap lesson

If you are learning to build with AI, before you pick a model, a framework or a 
tool, build the outline: folders, files, indexes, chapters, headlines, paragraphs.
Decide what lives where, what points to what, and where the orphans go. The agent
will only be as good as the structure you hand it.

I have been waiting my entire career for this moment. The tools finally caught up
with the way I think. I am not smarter than anyone else in the room. I just know
where things go, and I am completely sucked into this.

_JCody Ventures builds the systems behind video, drone, networks and marketing, 
with the filing system done first. See the [AI Projects](https://jcody.com/ai-projects/)._

P.S. The 2005 deck also had a rule about which color blouse the on-camera host should
wear. Not every rule aged well.