Rate sentiment with evidence

Sentiment labels each backed by the exact phrase that decided them, with a real “mixed” category instead of a mushy neutral.

When to use it

Labelling a batch of feedback, reviews, or messages where you need the call to be auditable — and where “mixed” is a real answer, not a cop-out.

What it prevents

The unaccountable label. A bare sentiment score gives you no way to tell a right call from a wrong one, and models quietly average a mixed review into a bland “neutral”. Requiring the deciding phrase, verbatim, and a real “mixed” category makes every label checkable and stops the genuinely mixed cases from disappearing.

The prompt

Analyse the sentiment of each review below and back every label with evidence.

### Output format
A markdown table with three columns: the review number, a label (positive, negative, or mixed), and the exact phrase that decided the label.

### Requirements
- Use "mixed" when a review praises one thing and criticises another; do not average them into neutral.
- Quote the deciding phrase verbatim; do not paraphrase it.
- If a review is sarcastic, judge the intended sentiment and note the sarcasm in the phrase column.

### Reviews
"""
1. Fast, once it finally installed.
2. Beautiful UI. Crashes daily.
"""

Why each part is there

Every section of that prompt exists because a specific failure happens without it. These are the checks it passes, and what each one is protecting you from:

• States a concrete task — why this matters
• Avoids vague language — why this matters
• Specifies an output format — why this matters
• Avoids dangling references — why this matters
• Separates data from instructions — why this matters
• Avoids over-aggressive phrasing — why this matters

How to adapt this one

Swap the reviews and keep the two disciplines: quote the deciding phrase verbatim, and use “mixed” rather than averaging a praise-plus-complaint into neutral. The sarcasm clause generalises to any domain where surface words invert the meaning — tell the model to judge intent and show its evidence, so you can overrule it when the quote does not support the label.

The mistake this prompt avoids

Trusting a label with no evidence. “Negative, 0.8” is unfalsifiable — you cannot tell whether the model read the review or guessed. Demanding the exact phrase that decided the call turns sentiment from an opaque score into a claim you can check in seconds, and the checkable version is the only kind worth acting on at volume.

A variation

Have a rating or a segment attached to each response? Ask for sentiment split by that dimension — separate the one-star comments from the five-star ones — rather than pooled. The most useful finding in feedback is usually that the same feature delights one group and frustrates another, and pooling the sentiment averages that signal into invisibility.

Check your version

Once you have adapted it, run it through the checker. The same nine checks that graded this page will grade yours, free and in your browser — paste it in, or build one from scratch by answering a few questions.