Extract Review Insights

This skill reads customer reviews for one product and pulls out the patterns that matter: what customers consistently like, what they consistently dislike, the specific language they use, and whether the reviews support or undercut the product's marketing claims.

The skill works from the reviews only. It does not invent themes, fabricate customer segments, estimate counts beyond what the data shows, or guess at root causes. When evidence is thin or mixed, it says so.

For reference on the expected output, see references/example-output.md.

Voice and Approach

Be direct and concise. Report what the reviews say without editorializing. Use plain language. Do not narrate your internal process or over-explain your methodology. When transitioning between steps, keep it brief and natural. The user wants the analysis, not a walkthrough of how you arrived at it.

Conversation Flow

Turn 1: Collect Reviews

The skill needs reviews for one product. Accept any format: pasted text, CSV export (Shopify, Yotpo, Bazaarvoice, PowerReviews, Judge.me, Stamped, or similar), or a document (PDF, Word, text file).

Optionally, the user may also provide:

• Product/brand name
• Product data (feed entry, PDP content, or product description), which gives the skill concrete claims and features to check reviews against
• Review metadata (star ratings, dates, verified purchase flags)

Let the user know what you need and what's optional. Don't over-explain the process.

Turn 2: Clarify (if needed)

Only ask follow-up questions if something is genuinely ambiguous:

• CSV columns aren't obvious (which column is the review body?)
• Reviews appear to cover multiple products
• Something else prevents you from starting

If everything is clear, skip this turn and go straight to the analysis.

Turn 3: Deliver the Analysis

Produce the full analysis as a Markdown document using the output structure below. Let the user know a few ways the output can be useful: the Useful Customer Language section is good raw material for PDP copy and ad creative, the Claims section can inform how confidently a product page leans into specific features, and the Likes/Dislikes sections can surface product improvement opportunities or FAQ content. Offer to adjust groupings, go deeper on a theme, or reframe anything.

Turn 4+: Revise

Edit individual sections in place. Do not regenerate the entire document for a single correction.

Analysis Instructions

Core principles

• Use only what the reviews say. Every insight must trace back to specific reviews. Do not infer themes that aren't explicitly stated or clearly implied by multiple reviewers.
• Focus on repetition. A single reviewer's opinion is an anecdote. A pattern appears when multiple reviewers independently say the same thing. Note when a theme appears in many reviews vs. a few.
• Report the evidence, not the cause. If customers say the zipper breaks, report that. Do not speculate on why the zipper breaks.
• Be honest about weak evidence. If only 2-3 reviews mention something, say so. If reviews contradict each other on a point, report the split. Do not smooth over mixed signals to make the analysis feel cleaner.
• Preserve customer language. When quoting or paraphrasing, stay close to the words customers actually used. Their phrasing is often more useful than a polished summary.

How to identify themes

1. Read all reviews. Note every distinct positive and negative point.
2. Group points that describe the same thing, even when worded differently. "Runs small," "had to size up," and "tight through the shoulders" are the same theme (sizing).
3. Count how many reviews touch each theme. Use plain language for frequency: "mentioned in many reviews," "a few reviewers noted," "one reviewer mentioned." Do not fabricate exact counts unless you can actually count them accurately from the data.
4. Rank themes by frequency. Lead each section with the most-repeated patterns.

How to handle product data

When product data (feed entry or PDP content) is provided:

• Extract the product's stated claims, features, and selling points.
• In the Claims Supported / Claims to Be Careful With section, cross-reference each claim against what reviewers actually say.
• A claim is "supported" when multiple reviewers independently confirm it.
• A claim needs caution when reviewers contradict it, when evidence is mixed, or when no reviewers mention it at all (absence is worth noting but is not contradiction).

When no product data is provided:

• Work from claims implied in the reviews themselves (e.g., if many reviewers say "this is waterproof," treat waterproofness as an implied claim).
• Note in the Claims section that you're working without the brand's own product data and that providing it would strengthen the analysis.

Output Structure

# Review Insights: [Product Name]

## Overview
[Product name, review count, rating distribution if metadata is available.
One paragraph summarizing the overall picture: what the dominant sentiment
is and what the key takeaways are. Keep it to 3-5 sentences.]

## What Customers Like
[Grouped by theme, ordered by frequency. Each theme gets a short heading,
a plain-language description of what reviewers say, and a note on how
common the theme is. Include short review snippets only when they add
something the summary doesn't. Do not list every positive comment --
group and summarize.]

## What Customers Don't Like
[Same structure as above. If a negative theme is minor or mentioned by
very few reviewers, say so. If a theme has mixed signals (some love it,
some don't), note the split.]

## Useful Customer Language
[Specific words, phrases, and descriptions customers use that are worth
borrowing for product copy, PDP content, ads, or email. Group by theme
if helpful. These should be the customers' actual words, not polished
marketing rewrites.]

## Claims Supported / Claims to Be Careful With
[If product data provided: cross-reference each identifiable claim against
review evidence. If no product data: work from claims implied in the
reviews. For each claim, note whether it's supported, contradicted, mixed,
or not mentioned. Be specific about the evidence.]

## Confidence Notes
[Flag which parts of the analysis are based on strong patterns (many
reviews, consistent signal) and which are based on thin evidence (few
reviews, mixed signals). If the review set is small, note that the
analysis may not be representative.]

Important Behaviors

• Produce the analysis as a single Markdown document.
• Use the product name in the document title. If no product name is provided, use "Untitled Product" and ask the user to confirm.
• When quoting customer reviews, use their actual words. Do not clean up grammar or rephrase unless the original is unintelligible.
• When editing, change only the requested section.

Edge Cases

Small review set (fewer than 10 reviews)

Produce the analysis but shorten it. With fewer than 10 reviews, most "themes" are really just individual opinions. Note this prominently in the Confidence Notes section: "This analysis is based on N reviews. Patterns identified here may not hold across a larger sample." Keep What Customers Like and What Customers Don't Like to the points that appear more than once.

Large review set (more than 500 reviews)

Use up to 500 reviews, prioritizing the most recent when dates are available. Let the user know how many reviews were included and that older reviews were excluded. If the user wants to focus on a specific time period or segment instead, offer to re-run with a different subset.

Mixed or contradictory reviews

When reviewers disagree on the same point (e.g., half say it runs large, half say it fits true to size), report the split. Do not average conflicting opinions into a lukewarm summary. Note the disagreement and, if possible, note whether different reviewer contexts (use case, body type, expectations) explain the split.

Reviews with no clear patterns

If the reviews are all over the place with no repeated themes, say so. Produce the analysis with whatever individual points are most notable, but be clear in Confidence Notes that no strong patterns emerged. This is a valid finding, not a failure.

CSV with unexpected columns

If the CSV doesn't have obvious review body, rating, or date columns, ask the user which columns to use. Common column names to look for: "Review Body," "Review Text," "Comment," "Content," "Body," "review_body," "review_text." For ratings: "Rating," "Stars," "Score," "review_rating."

Reviews in multiple languages

If reviews are in multiple languages, analyze all of them but note which language each quoted review is in. If translation is needed for the user to understand a quote, provide it in brackets.