Paste your open-text feedback into Claude, tell it what the text is and what decision you are trying to make, and ask for recurring themes with sentiment attached to each one - not an overall score. Below roughly a few thousand documents a month this beats setting up a dedicated sentiment classifier, because you get the reasons rather than a label. The catch is that the model will hand you a confident label on text it has misread, so check a sample against your own judgement before you present any distribution as fact.
Tool pricing checked August 2026.
Why is a sentiment score on its own useless?
“72% positive” tells you nothing you can act on. It does not say what the 28% are angry about, whether it is one issue or nine, or whether the positive comments are about the thing you just shipped or the thing you are about to kill.
What you want instead is a theme list: the five or six things people keep bringing up, each with its own sentiment and its own rough frequency. “People like how fast it is, and thirty of them got stuck at the same point in onboarding” is a sentence someone can do something with. A percentage is not.
This matters most because real feedback is rarely one-sided. “Love the product, hate the price” averages out to neutral, which is the least useful description of that customer possible. Attaching sentiment to each theme rather than each document keeps the signal.
Should you use a dedicated sentiment tool instead?
Below a few thousand documents a month, usually not - and it is worth understanding why, because most advice on this topic is a decade old.
Amazon Comprehend, the mainstream managed option, charges $0.0001 per unit for sentiment analysis, where a unit is 100 characters and every request is billed at a 3-unit minimum. A 300-character review costs $0.0003. Two thousand of them cost about sixty cents. AWS also gives new and existing customers 50,000 units (5M characters) a month free for the first twelve months. On raw compute, a classifier is essentially free.
What it costs you is everywhere else. You need an AWS account, credentials, a script to push the CSV through the API and pull the results back, and a decision about what to do with the label it returns. And a label is all you get: a sentiment value with a confidence score. It will not tell you that the negatives cluster around a broken password reset flow, because it does not read for topics unless you build that separately.
A general model reading the same 2,000 reviews gives you the topics, the sentiment per topic, representative quotes, and an answer when you ask it a follow-up question - with no pipeline to build. Past several thousand documents a month, running on a schedule, feeding a dashboard, the maths flips and a proper classifier or batch API earns its keep. Under that, you are building infrastructure to save sixty cents.
One warning about the older listicles you will find on this: MonkeyLearn, which appears in most “best sentiment analysis tools” roundups, no longer exists as a standalone product. Medallia acquired it in 2022 and monkeylearn.com now 301-redirects to medallia.com. Check that any tool you are recommended still exists before you plan around it.
How do you actually run the analysis?
Four things make the difference between a useful read and a generic one.
- Say what the text is. “These are 400 support tickets from the two weeks after our pricing change” changes the read completely versus dropping in an unlabelled block.
- Say what decision it feeds. Analysis for a roadmap argument looks different from analysis for a churn post-mortem. Tell it which.
- Ask for frequency and quotes, not just themes. Require a rough count per theme and two verbatim quotes. The quotes are how you audit it - if a quote does not support the theme, the theme is wrong.
- Ask what it could not classify. The residue is often the interesting part: the comments too ambiguous, too short or too sarcastic to place.
If your feedback comes from surveys you designed, the quality of this analysis is decided before you run it - the same point covered in our HR generalist guide, and worth pairing with the wider read in the market research guide.
Once Mira has clustered the feedback into themes, Elena turns the prioritised list into a business case with impact, effort and a recommendation attached.
View Elena →How do you install the Mira skill?
Mira arrives as a ZIP with a SKILL.md file at its root.
- Open the Claude desktop app, click Customize in the sidebar, then Skills.
- Click + and upload the ZIP. Claude reads the
SKILL.mdand summarises what it does. - Toggle it on. Off means Claude ignores it.
- Paste your feedback and describe the decision you are making.
Skills work on every Claude plan - Free, Pro, Max, Team and Enterprise - but they need code execution switched on. If the Skills menu is greyed out or missing, enable Code execution and file creation under Settings > Capabilities. On Team and Enterprise that is an owner-level setting in Organization settings, so you may need to ask.
If you cannot enable code execution, or you work in ChatGPT or Gemini, open the SKILL.md file, copy its contents, and paste them into custom instructions. The skill will not trigger automatically, but it will shape every reply in that conversation.
Where does this go wrong?
This is the section to read twice, because the failure mode is quiet.
- It will be confidently wrong. A model does not return “I am not sure what this person meant.” It returns a clean label in the same tone whether it read the comment correctly or not. Nothing in the output distinguishes a solid classification from a misread one.
- Sarcasm and understatement break it. “Great, another update” and “it is fine, I suppose” are exactly the comments a human reads instantly and a model flattens.
- Frequency counts are estimates, not counts. When a model says a theme appeared in 30 comments, it has not tallied them. If the number is going into a slide, count it yourself or have the model output the row numbers so you can.
- Your sample is probably not your customers. People who leave reviews and answer open-text questions are unusual by definition. The analysis describes who responded, and no tool fixes that.
- Theme boundaries are a judgement call. Run the same feedback twice and you may get six themes one time and four the next, because “slow” and “unresponsive” can reasonably be one theme or two. Neither answer is wrong; both change your chart.
- It cannot tell you what to do. It tells you what people said. The trade-off between fixing onboarding and shipping the feature you already promised is yours.
So spot-check before you present anything. Pull a sample of comments - fifty is plenty for a few hundred documents - label them yourself before looking at the model's answer, then compare. You are not looking for perfection, you are looking for where the disagreements cluster. If they land on one theme, that theme is unreliable and you should say so. If you and the model disagree on more than a handful across the sample, do not report the distribution as a finding. Do this once for a given type of feedback and you will know how much to trust it thereafter.
If most of your unread pile is support tickets, pair Mira's theme read with Sam, who drafts the replies once you know what the customer is actually upset about.
View Sam - Customer Support AI Skill ($14.99) →Why use a skill rather than write the prompt each time?
Two reasons, and neither is that the skill is smarter.
The first is consistency. Comparing this quarter's sentiment to last quarter's only means something if both were read the same way, with the same theme definitions and the same instructions about what counts as a complaint. A saved skill fixes the method so the comparison is real rather than an artefact of how you happened to word the prompt in April.
The second is that the four things listed earlier - context, decision, frequency and quotes, unclassifiable residue - are easy to forget when you are in a hurry, which is exactly when you are pasting 400 tickets in and hoping for the best.
Running quarterly pulse surveys? This designs the survey so the open-text answers are comparable between quarters - the raw material Mira needs to track a shift rather than guess at one.
View Employee Survey Prompt ($9) →What text is worth running this on?
Anywhere you have words at volume and nobody has time to read them: open-text survey answers, product reviews, support tickets, exit interviews, app store comments, sales call notes, social mentions. The method does not change between them. What changes is the context you supply at the top and the decision you are pointing the analysis at.
The one category to be careful with is anything where the sample is tiny. Twelve comments do not have themes; they have twelve opinions, and you should read them yourself.
In summary:
For general open-text analysis - reviews, tickets, survey answers, interview transcripts - use Mira ($14.99). If the pile is mostly support tickets, pair it with Sam ($14.99). If you are designing the survey that produces the feedback, start with the Employee Survey Prompt ($9). Then turn the themes into a case with Elena ($14.99). All work in Claude, ChatGPT and any AI chat, with a 30-day money-back guarantee - and all of them still need you to spot-check the output.
Related Skill Guides
- How to Use Claude for Email Marketing: The Nadia Skill Guide
- How to Use Claude as a Social Media Manager: The Maya Skill Guide
- How to Use Claude as a Growth Marketing Strategist: The Sofia Skill Guide
- How to Use Claude for Marketing Analytics: The Rafa Skill Guide
KissMySkills is a marketplace of 853 AI skills, 158 prompt packs, 55 agents & free tools for Claude, ChatGPT & any AI chat.