PM-301b · Module 2

Format Consistency

4 min read

Inconsistent examples produce inconsistent outputs. This is not a hypothesis — it is a reliable property of in-context learning. The model infers format from the examples. If your examples have different structures, the model infers that format is variable and produces variable outputs. Inconsistency in examples is permission to be inconsistent in outputs.

Do This

  • Use identical format across all examples: same labels, same order, same punctuation
  • If examples differ in length, ensure the length difference reflects a real input-output relationship (longer input → longer output)
  • Use the same separator between examples in every instance
  • Test your format specification by asking: could someone learn the format from examples alone, without any instructions?

Avoid This

  • Mix "Input:" / "Question:" / "User:" as labels across examples
  • Use "---" in some examples and "===" in others as separators
  • Include optional sections in some examples but not others without explanation
  • Allow examples to vary in structure and hope the model picks the right one
# WRONG: Inconsistent labels and structure
Question: What is CIPHER's domain?
Answer: CIPHER handles data analysis and statistical modeling.

Input: Who manages proposals?
Response: FORGE is responsible for proposals and SOW drafting.
She also handles scope gap analysis.

User: What does CLOSER do?
Output: Sales coaching.

# RIGHT: Consistent structure throughout
Input: What is CIPHER's domain?
Output: CIPHER handles data analysis, statistical modeling, and pipeline reporting.

Input: Who manages proposals?
Output: FORGE owns proposal writing and SOW drafting. She flags scope gaps and
structures pricing for every client engagement.

Input: What does CLOSER do?
Output: CLOSER provides sales coaching, discovery methodology, and deal strategy for
B2B enterprise accounts.

Input: {{USER_INPUT}}
Output: