How to Optimize Highlight Extraction
To get high-quality insights from NEXT AI, you need to carefully define two things:
- Product / Company (“Describe your company/product")
- Topic Focus (“Describe the topics on which NEXT AI should focus")
These two fields determine how NEXT AI understands your data and what it extracts from it.
Product / Company
This section provides context. It helps NEXT AI understand:
- what your business does
- how users interact with it
- what kind of data it will receive
If this is unclear, highlights will be generic, misinterpreted, or inconsistent.
How to write it
A strong description answers three questions:
- What do you offer? Be specific about your product or service.
- How do users interact with it? Describe real usage: app, website, store, CRM, support, etc.
- What kind of data will appear? Explain the nature of the input: reviews, transcripts, CRM notes, surveys, etc.
Example (Spotify)
Spotify is a global music and audio streaming platform that allows users to discover, play, and share music, podcasts, and other audio content across devices such as mobile, desktop, and smart speakers.
Users interact with Spotify through features such as search, playlists, recommendations, downloads, and personalized home feeds. Spotify operates on both free (ad-supported) and premium subscription models.
The input data consists of user feedback from app reviews, support tickets, and surveys, often describing issues, feature requests, or general user experience.
What makes a good description
A strong Product / Company description is:
- Concrete → clearly explains product and usage
- Operational → reflects real user behavior, not positioning
- Data-aware → explains what kind of input the AI will process
What not to include
Avoid:
- marketing language
- internal jargon
- irrelevant company history
Focus on what helps interpret user feedback.
Common mistakes
Too generic
We are a company that offers digital products.
→ No useful context for interpreting feedback
Too marketing-focused
We deliver world-class experiences that delight our users.
→ Sounds good, but contains no actionable information
Missing data context
Spotify is a music streaming platform.
→ Doesn’t explain what kind of input the AI will process
How to improve over time
If highlights are:
- too generic → add more detail about user interactions
- misinterpreted → clarify how the data is written (e.g. informal, incomplete)
- off-topic → better define input sources
Topic Focus
This is the most important part.
It defines:
- what insights are extracted
- how structured they are
- what is ignored
How to structure it
A strong Topic Focus typically includes:
- Data context
- Focus topics
- Rules and constraints
Example (Spotify)
You will receive user feedback from app reviews, surveys, and support tickets related to Spotify’s mobile and desktop apps.
Your task is to extract clear and structured insights from each input.
Focus on:
- Playback issues (e.g. buffering, crashes, offline mode)
- Search and discovery (e.g. finding songs, recommendations)
- Playlist management (creation, editing, sharing)
- Ads experience (frequency, relevance, interruptions)
- Subscription and pricing (premium vs free, billing issues)
- App performance and bugs
- User interface and usability
- Content availability (missing songs, regional restrictions)
Rules:
- Capture all relevant points mentioned in the feedback
- Ignore general praise unless a reason is given
What makes a good Topic Focus
A strong Topic Focus is:
- Explicit → clearly lists what matters
- Aligned with analysis → matches how you use the data later
- Constrained → prevents noise and hallucination
Common mistakes
No focus topics
Summarize the feedback.
→ Leads to shallow, inconsistent output
Too many or vague topics
Focus on everything related to the product.
→ Results become noisy and unfocused
Redefining the output format
Output:
- A short title
- A structured description
→ This is unnecessary and will likely reduce quality.
NEXT already defines the highlight structure internally. Overriding the output format can lead to inconsistencies and negatively impact downstream AI features such as Tagging Jobs, AI Chat, and analysis.
Align topics with your analysis
Your topics should reflect the questions you want to answer.
Start from real questions
- “Why are users unhappy?"
- “What drives churn?"
- “Where do users struggle?"
Translate into topics
| Question | Topics Needed |
|---|---|
| Why are users unhappy? | pain points, product areas |
| What drives churn? | pricing, reliability, UX |
| Where do users struggle? | journey steps, usability |
How many topics should you define?
- Too few → insights will be too generic
- Too many → highlights become noisy and inconsistent
As a rule of thumb:
- Start with 5–10 core topics
- Expand only when needed
Remove noise
If highlights are messy:
- remove rarely used topics
- merge overlapping ones
- add explicit exclusions
Examples:
- “Ignore general praise unless a reason is given"
- “Do not include individual product names"
- “Exclude timestamps or technical metadata"
How to improve over time
Refine based on real output:
- If highlights are too vague: add more specific topics, clarify what should be captured
- If highlights are inconsistent: define stricter structure, tighten rules
- If insights are missing: check if the topic is explicitly listed, if not, add it
- If highlights contain noise: add exclusions, narrow the scope