GUIDE: CREATING HIGH-PERFORMANCE CLINICAL FORMATTING PROMPTS


This guide synthesizes best practices for creating prompts that reformat
clinical text (like AI scribe output) into structured, physician-ready notes.
It is based on a "show, don't tell" philosophy.

CORE PRINCIPLES

1. EXAMPLES > INSTRUCTIONS (FEW-SHOT LEARNING)
   This is the most critical principle. Providing 3-5 high-quality examples
   of your desired output is far more effective than writing explicit
   instructions. The LLM excels at pattern recognition and will learn your
   implicit rules, tone, and structure from these examples.

2. BREVITY = QUALITY
   Concise prompts and concise outputs are better. Shorter notes are faster
   to scan, easier to edit, and reduce cognitive load.

3. ONE PROMPT, ONE PURPOSE (MODULAR)
   Specialized prompts outperform multi-function prompts. Trying to make one
   prompt do two things (e.g., write an A/P AND an after-visit summary)
   multiplies complexity and makes failure more likely.

ANATOMY OF A SUCCESSFUL PROMPT

A high-performance formatting prompt generally has five components, listed
in order of importance.

--- 1. THE FEW-SHOT EXAMPLES (MOST IMPORTANT) ---

This is the core of the prompt. Provide 3-5 complete examples of your exact
desired output. This is where the model learns your desired tone (e.g.,
formal vs. pithy), structure, and clinical shorthand.

--- 2. THE TASK STATEMENT ---

Begin the prompt with a clear, action-oriented instruction. Avoid "role-
prompting" (e.g., "You are a doctor"), as it adds length without improving
practical output.

Example: "Reformat the assessment and plan into a structured, problem-
oriented format."

--- 3. THE OUTPUT STRUCTURE ---

Explicitly show the SHAPE of your desired output. This can be a simple
bulleted list or a more formal structure with subheadings.

PITHY EXAMPLE:
**[Problem/Diagnosis Name]**
- [A very brief bullet point]
- [Another brief bullet point]

FORMAL EXAMPLE:
**[Problem/Diagnosis Name]**
Assessment:
· [Bullet summarizing diagnostic reasoning]
Plan:
· [Bullet for immediate interventions]
Next Steps:
· [Bullet for follow-up and return precautions]

--- 4. CONDITIONAL BOILERPLATE TEXT ---

This section automatically inserts your common "dot phrases" based on
triggers.

How it works: You define a trigger (e.g., "If well child check...") and the
text to add.

Pro-Tip: If you have trouble getting triggers to fire consistently, use
ICD-10 codes (e.g., "If diagnosis includes J06.9...") as they are very
reliable hooks.

Warning: Keep boilerplate text blocks short. Long blocks of text can confuse
the model and degrade output quality.

--- 5. THE FORMATTING RULES (Do this last) ---

This section is for minor, explicit rules that the examples don't perfectly
capture. This is the LAST place you should make edits.

Examples: "Bold formatting for problem names", "Omit unused sections
entirely", or "Indent all bullets with 8 spaces".