Prompt Engineering
Five ideas that change how you build with AI.
Not a tutorial. A set of lenses — architectural decisions that determine whether a system behaves reliably or just seems to work.
4 concepts · viewing 1 of 5
01
Layering
Your prompt is a stack of sources with different authority and lifespan — not one blob of text.
Design choices
- 01PrecedenceWhat wins when two layers conflict — a user instruction versus a retrieved document, a system rule versus a session preference.
- 02ProvenanceWhether the model can say where a claim came from — which layer it lives in, so contradictions can be traced back to their source.
- 03Budget allocationWhich layer gets cut first when the context window fills — and whether that decision is made deliberately or left to chance.
- 04Layer orderingWhether the sequence of layers in the prompt actively signals relative importance, not just logical grouping.
- 05Refresh cadenceWhich layers are fixed for the session and which are rebuilt per turn — and whether stale content persists past its useful life.
Typical root causes
- 01Instructions ignored after a long conversationThe task was stated once, near the top, and gets deprioritized as later turns fill the window — the "valley of meh." The model never forgot; the instruction just lost effective weight.
- 02The model contradicts something said early onA later layer — a retrieved document, a follow-up message — implicitly overrides an earlier constraint. No conflict resolution rule exists, so recency wins by default.
- 03Retrieved content overrides an explicit user ruleRetrieval results appear after the user instruction in the prompt, so the model weights them as more recent and therefore more authoritative.
- 04Can't tell a user's rule from a random file's suggestionWithout provenance markers, the model has no basis for deciding which source carries more weight — everything looks equally authoritative.
- 05Behavior changes just from reordering the same contentThe prompt structure isn't encoding importance intentionally, so the model infers it from position — and that inference is fragile.
Context window34% used
System
always present
Developer
Tools
Retrieved context
History
Memory
Task, restated
always present
Recognize this?
1 / 5