1:1 mentoring with Big Tech AI engineers
LLM & Agentic

Prompt Engineering Foundations

Where an instruction belongs — tool description, system prompt, or user turn — why emphasis that worked on older models now causes over-triggering, and the prompt patterns that have expired.

Last updated

Foundations6 min readFirst readLLM Lifecycle

After this section you can

  • Place an instruction in the right surface — tool description, system prompt, or user turn — and say what breaks when it is in the wrong one
  • Strip emphasis that made older models comply and now causes over-triggering
  • Recognise the prompt patterns that expired, and name the API feature that replaced each
01

Prompt Engineering Foundations

Where an instruction lives decides whether it is followed, cached, or quietly ignored. And the current models follow the system prompt closely enough that prompts written to overcome older ones now overshoot.

THE CENTRAL IDEA

Most prompt problems are not wording problems — they are placement problems. A prompt is three surfaces rendered in a fixed order (tools, then system, then messages), and each has a different lifetime, a different cost, and a different moment at which the model reads it. Put an instruction in the wrong surface and it either gets ignored at the moment it mattered or destroys your prompt cache. The second half of the job is volume: emphasis was a workaround for models that under-followed instructions, and on current models the same emphasis causes over-triggering.

Three homes for an instruction — chosen by how often it changes, not how important it is
RENDER ORDER 1 · TOOLS name · description · schema read when the model picks a tool changes almost never 2 · SYSTEM role · standing rules · contract read on every turn stable across the conversation 3 · MESSAGES the task · the data · this turn read as the conversation changes every request cache breakpoint CACHED — written once, re-read at ~10% of input price processed fresh, every call Anything that varies per user or per request belongs on the right of that line. Splice a user’s name into the system prompt and you have just given every user a private cache.

Related

More in LLM & Agentic

Get full access to all 74+ sections with code examples, diagrams, and interactive animations.

Unlock Premium