Prompt best practices

Cura 1T follows instructions closely, which cuts both ways: precise prompts get precise behavior, and vague prompts get confident improvisation. These patterns keep clinical outputs on rails.

Cura 1T is a research model, not a medical service, and not a substitute for a clinician. Benchmark scores do not establish safety for unsupervised clinical use.

Structure the system prompt

Role, task, rules, output contract — in that order. Put non-negotiables (safety behavior, what to do when information is missing) in the rules, not the user turn:

System prompt
You are a clinical documentation assistant for licensed clinicians.## TaskSummarize the encounter note the user provides.## Rules- Use only facts present in the note; never infer diagnoses.- Flag any medication-allergy conflicts under a "Safety" heading.- If information is missing, say what is missing instead of guessing.## OutputMarkdown with headings: Summary, Medications, Safety, Follow-up.

Delimit injected content

Wrap documents, notes, and retrieved passages in explicit markers so instructions and data can't blur:

Python
messages=[    {"role": "system", "content": SYSTEM_PROMPT},    {"role": "user", "content": "Note:\n<note>\n" + encounter_note + "\n</note>\n\nQuestion: draft the follow-up plan."},]

Checklist

  • One task per request. Chain calls rather than stacking extraction + reasoning + formatting into one mega-prompt.
  • Show, don't describe, formats. A single worked example (few-shot) beats a paragraph of format prose; for machine-readable output use JSON mode.
  • Temperature 1.0 is the default and the recommendation for reasoning-heavy clinical work; lower it only for deterministic formatting tasks.
  • Leave headroom for thinking. Reasoning tokens count against max_tokens — budget generously or answers truncate mid-plan.
  • Stable prefixes cache. Keep the long, shared part of the prompt (system + document) byte-identical across calls so it's served from cache.
← Automatic retryBenchmarking →