Few-Shot Example Curation: Quality, Rotation, and Counter-Examples, Part 1

Chain-of-thought is not a universal upgrade; it helps on reasoning-bound tasks and is overhead everywhere else.

What AI does well here

• Use CoT on tasks requiring multi-step reasoning (math, complex logic, multi-constraint problems)
• Use few-shot CoT examples on reasoning tasks where the structure of reasoning matters
• Hide CoT from end-user output when the reasoning isn't user-facing value
• Evaluate with and without CoT to confirm benefit on YOUR task

What AI cannot do

• Make non-reasoning tasks better with CoT (it just adds tokens)
• Make CoT a substitute for fine-tuning on hard reasoning tasks
• Trust the reasoning trace as ground truth (models can produce plausible-but-wrong reasoning)

Few-shot examples teach the model your edge cases and your style — pick them like you're picking a teaching set, not filler.

What AI does well here

• Cover the three most common input shapes you actually see
• Include one tricky edge case your last release got wrong
• Match the exact output format you expect, character for character
• Surface the reasoning step if you want the model to externalize one

What AI cannot do

• Generalize to a regime far outside the chosen examples
• Replace a clear instruction — examples and instructions reinforce each other
• Stay relevant when your data distribution shifts

Showing the model an explicit wrong answer alongside the right one prevents specific failure modes — but can also seed them if done sloppily.

What AI does well here

• Pair every negative example with the corrected version
• Label them clearly: 'BAD' and 'GOOD' tagged blocks
• Use sparingly — one or two negatives, not ten
• Reserve them for failures you have actually observed

What AI cannot do

• Substitute for clear positive examples
• Stop the model from repeating the bad pattern in adjacent contexts
• Generalize 'don't do X' beyond the literal example

Maintain an example pool larger than what fits in the prompt, and sample N examples per call with a stable hash for reproducibility.

What AI does well here

• Reduce mimicry of one phrasing
• Surface examples evenly over time
• Detect example-set bugs faster

What AI cannot do

• Replace evaluation against held-out cases
• Compensate for biased pool composition
• Guarantee any single output's quality

Few-shot examples teach the model your output shape and your edge-case handling. Examples that all look alike teach only the easy case; the model fails on anything off-distribution.

What AI does well here

• Match the format of provided examples in new outputs
• Generalize patterns shown across diverse examples
• Handle cases similar to ones you demonstrated

What AI cannot do

• Generalize from examples that all look the same
• Recover gracefully from a case unlike any example
• Tell you when an example was a poor choice

Teams over-invest in fine-tuning when 5-10 strong few-shot examples would solve the task; they also avoid fine-tuning when the cost arithmetic actually favors it.

What AI does well here

• Score whether the task is style or knowledge
• Estimate prompt-token cost with examples included
• Compare against fine-tune training and inference cost
• Recommend evals to compare both

What AI cannot do

• Account for hidden ops cost of maintaining a fine-tune
• Predict whether the model provider will release a better base model
• Replace a real eval comparison

Few-shot examples teach the model the shape of the answer. Choosing diverse, edge-leaning examples beats stacking similar ones.

What AI does well here

• Suggest covering edge cases in examples.
• Help format examples consistently.
• Spot when examples contradict each other.

What AI cannot do

• Know which examples your model will weight most.
• Replace systematic eval.
• Guarantee a behavior change from one swap.