What is the right altitude for a system prompt?
Right altitude is Anthropic's term for system-prompt guidance pitched between two failures: specific enough to steer the model's behaviour, yet general enough to give it heuristics rather than brittle, hardcoded if-else rules.
Anthropic's "Effective context engineering for AI agents" (2025-09-29) describes system prompts as failing in two opposite directions. At one extreme, engineers hardcode complex, brittle logic to force exact behaviour, which is fragile and expensive to maintain. At the other, they write vague, high-level guidance that gives the model no concrete signal, or assumes context it does not have.
The right altitude sits between them, a Goldilocks zone. The post's test is the minimal set of information that fully defines the expected behaviour, adding that minimal does not necessarily mean short.
What each altitude looks like
Too low
"If the customer mentions a refund and the order is under 30 days and the item is not electronics, say X; if it is electronics, say Y..." Every unlisted case falls through, and every policy change means editing branches.
Too high
"Be helpful and follow company policy." The model has to guess which policy and what helpful means here.
Right altitude
State the goal, the policy and the reasons, the few hard rules that never bend, and what to do when unsure. The model can then handle cases nobody listed.
How to find it
Start minimal
Anthropic recommends testing a minimal prompt on a capable model first, then adding instructions and examples to fix the failure modes actually observed.
Organise into sections
Distinct sections, marked with XML tags or Markdown headers, such as background, instructions and tool guidance. Anthropic notes exact formatting is likely to matter less as models improve.
Examples over rule lists
A few diverse, canonical examples carry more than a long list of edge cases.
Remove contradictions
OpenAI's GPT-5 prompting guide (2025-08-07) found contradictory or vague instructions more damaging to GPT-5 than to earlier models, because it spends reasoning tokens trying to reconcile them.
The altitude moves with the model
The same OpenAI guide reports that Cursor's instruction to be thorough when gathering information, useful for older models, made GPT-5 call search repeatedly on small tasks; softening it fixed the behaviour. OpenAI's later guidance for GPT-6 Astra goes further: step-by-step recipes written for earlier models can limit a stronger one, and strong "ask first" wording can make it stop work it could safely finish.
A prompt pitched right for one model can be too low for the next. Re-checking the altitude belongs in every model upgrade.
How Motiontic's engineers apply it
Motiontic's engineers write the goal, the reasons and the hard limits first, then add instructions only when an eval shows a failure that needs one. Each instruction is tied to the failure it fixes, which makes it clear what can be removed when the model changes.
System prompts at three altitudes
| Too low | Right altitude | Too high | |
|---|---|---|---|
| Style | Hardcoded if-else rules | Goals, reasons, heuristics and a few hard rules | Vague advice |
| Unlisted cases | Fall through or break | Handled by judgement | Handled by guesswork |
| Maintenance | Grows with every edge case | Changes when policy changes | Little, but little effect |
| Typical failure | Brittle, contradictory behaviour | Needs re-tuning per model | Inconsistent behaviour |
Common questions
What does right altitude mean in prompt engineering?
It means writing system-prompt guidance at a level of detail between brittle, hardcoded rules and vague advice, so the model has clear heuristics and can handle cases the prompt did not list.
Should system prompts be short?
Minimal, not necessarily short. Anthropic's target is the smallest set of information that fully defines the behaviour, which for a complex agent can still run to several sections.
Why does a prompt that worked on one model fail on a newer one?
Instructions that compensated for an older model's habits can over-steer a newer one. OpenAI's guides show thoroughness and ask-first instructions causing over-searching and early stops in later models.
Related
System prompt
The standing instructions that frame every turn of a conversation or agent run.
Context engineering
Curating everything in the model's view, not just the wording of the prompt.
Context Engineering vs Prompt Engineering
Wording the instructions, compared with curating every token the model sees.
Few-shot prompting
Teaching a model the task by showing a few examples, not by retraining it.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.