Use Few-Shot Prompting for Consistent Output Formatting

When you need the LLM to follow a specific, non-obvious output format, abstract instructions alone are often not enough. Including 2-3 concrete examples (few-shot prompting) in your system prompt dramatically improves format consistency and reduces parsing failures.

Real-world example: A workflow extracts structured event data from free-text calendar entries. Without examples, the model inconsistently formats times, handles multi-day events differently, and sometimes omits fields.

System prompt with few-shot examples:

Extract event details from the text and return a JSON object.

Example 1:
Input: "Team lunch at Olive Garden next Tuesday from noon to 1:30pm"
Output: {"title":"Team Lunch","location":"Olive Garden","start_date":"next Tuesday","start_time":"12:00","end_time":"13:30","duration_minutes":90,"is_all_day":false}

Example 2:
Input: "Company offsite Dec 15-17 at the Marriott downtown"
Output: {"title":"Company Offsite","location":"Marriott Downtown","start_date":"Dec 15","end_date":"Dec 17","start_time":null,"end_time":null,"duration_minutes":null,"is_all_day":true}

Example 3:
Input: "Quick sync with Sarah tomorrow at 3"
Output: {"title":"Sync with Sarah","location":null,"start_date":"tomorrow","start_time":"15:00","end_time":"15:30","duration_minutes":30,"is_all_day":false}

Now extract from this input:
{{ $json.calendar_text }}
```text
Key patterns demonstrated by the examples:

```text
- Time normalization: "noon" → "12:00", "3" → "15:00"
- Default duration: 30 minutes when no end time specified
- Multi-day handling: start_date/end_date with is_all_day: true
- Null fields: explicitly null instead of omitted
- Title cleaning: removed filler words, proper capitalization
```text
> **Note: Example Selection Matters**
>
> Choose examples that cover your most common edge cases. If 30% of your inputs are multi-day events, at least one example should demonstrate that format. If ambiguous times are common, show how you want them resolved.

Few-shot prompting typically reduces output format errors from ~15-20% to under 3%, which directly translates to fewer failed workflow executions.

**Related:** [Use Manual Trigger During Development Instead of Webhook or Schedule](../api-cost-optimization/01-use-manual-trigger-during-development-instead-of-webhook-or-schedule.md) | [Flatten Deeply Nested API Responses](../code-node-mastery/01-flatten-deeply-nested-api-responses.md)