Why is sarcasm so hard for LLMs to detect?

Sarcasm inverts literal meaning using context, tone, and exaggeration rather than keywords. A model that reads only the words sees praise where a human sees mockery. The prompt teaches it to weigh contextual cues, not vocabulary.

How do the labeled examples help?

Few-shot examples anchor the model to the kind of irony common in your domain. Reviews use praise contradicted by defects; social media leans on emoji and exaggeration; literature uses mock-authority. Domain-matched examples sharply improve accuracy.

What does the confidence threshold do?

It tells the model not to commit to a sarcastic or ironic label unless its confidence meets your bar. Below it, the text is labeled sincere with a note — useful when false positives are costly.

Does it need an API key?

No. The prompt is built locally in your browser with no network call. You paste your text into the prompt and run it in ChatGPT, Claude, or any model.

Which output format should I choose?

JSON gives a label, confidence, and cue list ideal for pipelines; the explained format adds a one-line rationale for human review; label-only is fastest when you just need the classification.

Socratic Irony Detector Prompt

Socratic irony detector prompt

Sarcasm and irony are where literal sentiment analysis falls apart: the words say one thing and mean the opposite. “Best day ever” after a three-hour queue is mockery, not praise. The Socratic irony detector prompt builds a few-shot prompt that teaches a model to read the contextual cues — contradiction, exaggeration, scare quotes, emoji — that reveal non-literal intent, tuned to the domain you’re working in.

How it works

You pick a domain (social media, product reviews, or literature), which loads labeled examples matched to that style of irony along with an explanation of the cue each one relies on. You set the output format and a confidence threshold below which the model defaults to “sincere” rather than guessing. The tool assembles a prompt listing the contextual cues to weigh — literal-versus-situation contradiction, hyperbole, tonal markers, mock-praise, sentiment-fact mismatch — the few-shot examples, and your threshold and format. It runs locally with no API key; you paste in the text to classify.

Tips and notes

Match the domain. Review sarcasm and literary irony look different; the wrong example set hurts accuracy.
Raise the threshold when false positives cost you. Moderation pipelines often prefer to miss subtle sarcasm than to flag sincere praise as mockery.
Use JSON output with a cue list. Reading why the model labeled something sarcastic is the fastest way to spot and fix misclassifications.
Add your own examples. The domain sets are a starting point; pasting a few of your hardest real cases as extra few-shot examples lifts accuracy further.