Six rules for context engineering
The whiteboard version — six rules for laying out a context window, each with a number measured on this site's chat.
llmcontext-engineeringprompts
Everything I learned building this site’s chat compresses into six rules. Each one has a number behind it, measured on the system running in the corner of this page.
| # | Rule | One line | Measured |
|---|---|---|---|
| 1 | Separate reference from rules | Reference gets searched when relevant. Rules must be obeyed every time. Sort every token into one of the two buckets before placing it. | A 5,600-token corpus and a one-line length rule needed opposite positions |
| 2 | Position is priority | Identity at the top, bulk in the middle, must-win behavior at the end. The middle is a black hole. | Same rule moved from top to end: 230 words down to 75 |
| 3 | Move before rewriting | When the model disobeys, relocate the instruction before strengthening its wording. | Rewriting got me to ~110 words. Moving got me to 75 |
| 4 | Budget tokens like money | Every token dilutes attention on every other token. Not “does it fit” — “does it earn its place”. | Instruction drift was measurable behind just 5,600 tokens |
| 5 | Stable prefix, volatile suffix | Byte-stable content first and cached. Everything that changes goes after the cache marker. | Warm questions cost 0.15 cents against 0.8 cold, and tuning the suffix is free |
| 6 | Measure behaviors separately | They fail one at a time. One assertion each, run on every change. | The same prompt scored 10/10 on grounding while running 3× over length |
The full system these six rules run on — corpus, citations, caching and the ten-question eval suite — is written up in the chat on this site, explained.
If you keep one sentence: the model gives you a fixed attention budget — lay it out like a designer laying out a page, not like a lawyer adding clauses.