Context is the
Foundation.

The total information available to the model when you prompt it.

Data, documents, emails, conversations, preferences, rules.

Key Concepts

Context

The total information available to the model when you prompt it. Everything it can "see."

Context Drift

When that knowledge goes stale. Old notes, outdated instructions, scattered files. The AI reads all of it and can't tell what's current.

Stale Context

Information that was true once but isn't anymore. The AI acts on it confidently.

The Demo

Same project. Same information. Same tools.
One maintained. One drifted.

Prompt (sent to both)
> Draft a post welcoming the beta cohort and post it as a draft to Circle.
While these run, let me show you the difference.

What's Under the Hood

✓ Maintained
CLAUDE.md — the map (loads every session)
wiki/ — current knowledge: people, features, decisions
skills/ — task instructions (posting, broadcasting)
services/ — API tools (Circle MCP server)
raw/transcripts/ — recordings, organized by date
✗ Drifted
NOTES.md, circle api notes.txt, tim notes.md...
18 files at root, no structure
Same tools buried underneath
 
CLAUDE.md says "Alpha is the current program"
Skill file says use the wrong API format

Results

✓ Maintained

  • Named all 7 founders and their companies
  • Got the curriculum, weekly structure, Week 3 cutoff
  • Draft posted to the live Circle community

✗ Drifted

  • No founder names
  • "Alpha members rolling into the cohort" (stale note)
  • Generic post — could be about any cohort anywhere
  • Also posted — but based on wrong context
Both "worked." Only one was right.

They Didn't Build
Better Prompts.

They built a system.

Organized context. Current information. Tools connected to both.
That's the foundation everything else depends on.
Patrick McHeyser