The Prompt Lab — Exemplar Priming
Learn the exemplar priming prompting technique with concrete before/after examples.
Exemplar Priming
The Technique
Exemplar Priming means giving the model one or more concrete examples of the exact output format and tone you want before asking it to do the real task. It works because frontier models are exceptional pattern-matchers — they’ll reverse-engineer your implicit style rules from a sample far more reliably than they’ll infer them from a description alone.
The Naive Prompt
Write a changelog entry for our new feature: users can now
export their dashboard data as a CSV file.Why It Falls Short
Without a reference point, the model defaults to generic changelog prose — often stiff, overly formal, or structured nothing like your product’s voice. You’ll get something technically accurate but stylistically foreign, which means editing time or a brand that feels inconsistent across releases. “Write a changelog entry” is an instruction with a thousand valid interpretations.
The Improved Prompt
Write a changelog entry for our new feature in the same style
as this example from our product:
---
EXAMPLE:
🔗 Connect your calendar in one click
You can now sync Google Calendar or Outlook directly from
Settings → Integrations. Events show up automatically —
no manual imports, no copy-paste.
---
NEW FEATURE TO WRITE:
Users can now export their dashboard data as a CSV file
(available under Settings → Export).Why It Works
The example communicates at least five things simultaneously that would be painful to describe in words: the emoji-first format, the conversational second-person voice, the em-dash rhythm, the parenthetical navigation hint, and the “no X, no Y” compression pattern. Claude Sonnet 4.6 and GPT-5.4 will both extract those patterns and apply them faithfully — the output lands production-ready instead of needing a full rewrite.
When to Use This
- You have an established style that’s hard to articulate. If your brand voice lives in your gut or in existing docs rather than a formal style guide, one good example teaches it faster than a paragraph of adjectives (“warm but professional, concise but not terse…”).
- You’re building a pipeline or automation. When prompts run without human review — generating social posts, release notes, support macros — exemplar priming is your consistency safeguard. Embed 1–2 real examples from your corpus directly in the system prompt.
- The output has a rigid micro-format. JSON schemas, structured summaries, code comments with specific conventions — any time the shape of the output matters as much as the content, show don’t tell.
Next edition: we’ll cover Stepwise Decomposition — breaking a complex task into an explicit sequence of substeps so the model can’t skip the hard parts.