System Prompt Altitude

Definition

The "right altitude" (or "Goldilocks zone") for a system prompt: the level of specificity that is "specific enough to guide behavior effectively, yet flexible enough to provide the model with strong heuristics" — avoiding both over-specific brittleness and over-general vagueness.

Key points

• Two failure modes to avoid:
  • Too low / over-specific (verbatim): "hardcoding complex, brittle logic in their prompts" — fragile and high-maintenance.
  • Too high / over-general (verbatim): "vague, high-level guidance that fails to give the LLM concrete signals" — assumes shared context the model lacks.
• Target (verbatim): "specific enough to guide behavior effectively, yet flexible enough to provide the model with strong heuristics."
• Practical guidance: use clear sections (e.g. <background_information>, <instructions>, tool guidance, output description); start minimal and add instructions/examples based on observed failure modes; keep it "informative, yet tight."

Why it matters for the exam

• System-prompt scenarios test recognizing the two failure modes and choosing the balanced "right altitude." Watch for answers that are either a wall of brittle rules or a vague one-liner.

Common gotchas

• The fix for a misbehaving prompt is not always "add more rules" — over-specification is itself a failure mode. Prefer strong heuristics plus a few canonical examples.
• Start minimal and iterate on real failures rather than pre-emptively enumerating every edge case.

See also

• Context Engineering
• Context Rot
• Lost In The Middle
• Just-in-Time Retrieval

Sources

• blog-context-engineering