Filter comments and reviews before they go live
Every comment, review or community message goes through a filter that detects insults, spam, personal data and off-topic content, and only what passes is published on its own.
- Level
- Low code
- Requires a simple automation or an API call.
- Time
- One afternoon
- Published
What it costs
Through the API, each comment evaluated costs a fraction of a cent with a fast model. For a store with 500 reviews a month, the monthly cost is measured in US cents; the saving is the daily hour of whoever moderates by hand today.
guided mode // I ask, you answer
ready to start
Do it with me, step by step
I will ask you 2 questions about your business, use your answers to get the prompt ready to copy, and then walk you through the 4 steps, one at a time. Like a Thermomix: you just answer and move on.
Who it is for
Online stores with reviews, marketplaces, communities and any site where users publish text that others see.
The problem
Either you moderate by hand and everything takes a day, or you publish automatically and one day an insult, a personal phone number or spam from a competitor shows up under your product. Neither option scales.
What changes
Clean content is published in seconds. Doubtful content goes into a queue for human review with the reason already written. Clearly unacceptable content is rejected on its own. The team reviews the 5% of hard cases instead of 100%.
Ingredients
- The point where the comment is received today before publishing (form, API, database)
- API access to a model
- Your community rules in writing, even if they are five lines
- A "pending review" queue or label in your system
Step by step
- 1
Write your rules as categories, not as principles
"Be respectful" is useless for classifying. "Direct insult at another person", "personal contact details", "promotion of another business", "sexual content" work. Each category must have an action: reject or review.
- 2
Decide what is rejected automatically and what goes to review
Obvious spam and personal data: automatic rejection. Aggressive tone or harsh but legitimate criticism: human review. An honest negative review must be published; removing it costs you credibility.
- 3
Ask for a three-value output
The model replies APPROVE, REVIEW or REJECT, plus the category that drove the decision. Nothing else. Your system acts on that word.
- 4
Test with your last 200 real comments
Include the ones you rejected by hand and the ones you approved. What you need to bring to zero is rejections of legitimate comments; a spam message that goes to review is tolerable.
The prompt
Replace what is in brackets with your company details, or use the guided mode above: it asks you and fills it in for you.
You are the automatic moderator for [SITE NAME], a Chilean [TYPE: product reviews / community / marketplace] platform. Evaluate the following content submitted by a user and decide on an action. Categories that are REJECTED automatically: - spam: promotion of another business, repeated links, meaningless text - personal_data: phone numbers, emails, addresses or RUT numbers of people - sexual_content: explicit or suggestive, aimed at other users Categories that go to human REVIEW: - aggression: direct insult or threat to a person or company - possibly_fake: generic review, suspiciously positive or negative with no concrete details - off_topic: does not talk about the product or service Everything else is APPROVED, including negative criticism as long as it is about the product or service and contains no insults. An honest negative opinion is NOT grounds for rejection. Reply ONLY with two lines: action: <APPROVE|REVIEW|REJECT> reason: <category from the list, or "none" if approved> Content to evaluate: [USER TEXT]
What usually goes wrong
- ✕Rejecting negative criticism. It is the first thing people notice and it destroys trust in your reviews.
- ✕Leaving the model as the only instance. The REVIEW categories exist because human judgment is still needed at the edges.
- ✕Not periodically reviewing what was rejected. Every month, look at 30 random rejections to detect biases in the filter.
When not to use this recipe
If you publish fewer than 10 comments a week, review them yourself: it takes 5 minutes and you build judgment. And if your platform already has built-in moderation, try that first before building your own filter.
Source and real case
E-commerce and communities
Anthropic's guide recommends exactly this three-output design with human review in the middle, instead of a binary filter that forces you to choose between over-censoring and letting too much through.
Official use-case guide:Content moderation — Claude Platform Docs · Anthropic
Last verified:Follows the three-output design with human review in the middle that Anthropic's official content moderation guide recommends.
Share it
If someone on your team could use this recipe, send it in one click. The suggested text copies itself.
Want to apply it in your business? Message Cristián directly.
Telegram