Watch the idea
The concept
Rule
Early on the desk
Pages overflow
The window is finite
Pin the ask at the end
What is near now is memory
The desk has a size
Long chats forget the beginning. The model is not getting bored. The early tokens left the window. System rules, sources, and the actual question compete for the same finite desk.
What to pin
If the desk is full, put the rule, the allowed sources, and the question near the end. Summarise old turns instead of pasting the whole thread. RAG is one way to keep only useful pages on the desk.
Overflow looks like personality
People say the model 'ignored my constraints'. Often the constraints fell off. Treat context like memory with a hard cap. Measure tokens. Design for the cap, or the lesson will fail in production.
Worked example
The lost honor code
You paste a 40-page PDF, then a one-line honor rule at the top, then chat for an hour. The rule is gone. Move 'cite or refuse' next to the last question. The desk remembers what is near the now.
Picture to keep
Fit the desk
- 1Pin the brief
- 2Drop noise
- 3Ask last
