🤖 AI Engineering · Prompt Engineering

Select few-shot examples that move accuracy (Classification tasks)

Choose and order examples covering edge cases so the model generalizes instead of copying surface patterns.

intermediate~30 minAI EngineersML EngineersLLM App Developers

Steps

1Cover each output field's edge cases once: empty, unicode, long input, ambiguity
2Order examples easy-to-hard so difficulty ramps within context
3Match example format byte-for-byte to the required output format
4Prefer real production samples over invented ones
5Rotate in failure cases you fixed, turning regressions into teachers
6Measure: remove one example at a time, drop any whose removal costs nothing
7Include one deliberately ambiguous item with the correct tie-break label
8Balance classes in examples; imbalance biases predictions

Common Pitfalls

▲Six near-identical easy examples teaching style but not judgment
▲Whitespace mismatches between examples and instructions

Commands

Install with skills CLI

$ npx skills add aniruddhaadak80/skills --skill prompt-engineering-few-shot-selection-cls

Install globally

$ npx skills add aniruddhaadak80/skills --skill prompt-engineering-few-shot-selection-cls -g

Tags

#prompts#few-shot#ai-engineering#prompt-engineering

Install this skill

$ npx skills add aniruddhaadak80/skills --skill prompt-engineering-few-shot-selection-cls -a opencode

SKILL.md

# Select few-shot examples that move accuracy (Classification tasks)

> Choose and order examples covering edge cases so the model generalizes instead of copying surface patterns.

## When to Use This Skill

Choose and order examples covering edge cases so the model generalizes instead of copying surface patterns.

## Steps

1. Cover each output field's edge cases once: empty, unicode, long input, ambiguity
2. Order examples easy-to-hard so difficulty ramps within context
3. Match example format byte-for-byte to the required output format
4. Prefer real production samples over invented ones
5. Rotate in failure cases you fixed, turning regressions into teachers
6. Measure: remove one example at a time, drop any whose removal costs nothing
7. Include one deliberately ambiguous item with the correct tie-break label
8. Balance classes in examples; imbalance biases predictions

## Common Pitfalls

- Six near-identical easy examples teaching style but not judgment
- Whitespace mismatches between examples and instructions

View source on GitHub ↗

Organize identity, rules, tools, and output contracts into sections that evolve without breaking behavior.

🤖 AI Engineering·~30m

Open →

Select few-shot examples that move accuracy (Extraction tasks)intermediate

Choose and order examples covering edge cases so the model generalizes instead of copying surface patterns.

🤖 AI Engineering·~30m

Open →

Guarantee structured outputs from LLMsintermediate

Get machine-parseable JSON reliably using schema-first prompting, constrained decoding, and repair loops.

🤖 AI Engineering·~35m

Open →