Auto-generated summaries
Every entry you save in Ultrathink goes through an AI processing pipeline that automatically summarises, classifies, and connects your content. This happens in the background, so your knowledge base gets smarter with every entry you add.
The 4-step AI pipeline
When you save an entry, it passes through four processing steps in sequence:
Entry saved --> [1. Summary] --> [2. Classification] --> [3. Relationships] --> [4. Similarity] --> Done
Each step builds on the previous one, enriching your entry with metadata and connections.
Step 1: Summarisation
The AI reads your entry's content and generates a concise summary tailored to the content type. This summary appears in the AI Summary section of the entry's detail panel and is also used as preview text in the grid view.
The summarisation step focuses on extracting the most important information: key points, decisions, action items, and conclusions.
Step 2: Classification
The AI extracts structured metadata from your entry:
| Metadata | Description |
|---|---|
| Entity type | Classifies the entry as a Project, Task, Knowledge, or Unclassified |
| Topics | Identifies relevant topics and matches them against your existing topic list |
| People | Detects names mentioned in the content and matches them to your people list |
| Category | Labels the entry as Work or Personal |
| Due date | Extracts deadlines if mentioned (particularly for task-type entries) |
The AI matches against your existing topics and people rather than creating new ones. For best results, add your key topics and people to your lists before saving entries.
Step 3: Relationship extraction
The AI identifies connections between the new entry and your existing entries:
- References to the same project
- Mentions of the same people
- Thematically connected content
- Continuation of previous conversations
Related entries appear in the Related section of the detail panel, giving you a web of connections across your knowledge base.
Step 4: Similarity detection
Using semantic analysis, the AI finds entries with similar content:
- Duplicate detection: Identifies when the same article or content has been saved more than once
- Related research: Finds entries covering the same topic from different angles
- Different perspectives: Surfaces content that addresses similar subjects in different ways
Similar entries appear in the Similar section of the detail panel with a relevance score.
How summaries differ by content type
The AI adjusts its summarisation approach depending on what type of content you have saved:
| Content type | Summary focus | Example output |
|---|---|---|
| Links and articles | Key points, main argument, conclusions | "This article argues that microservices add complexity without clear benefits for teams under 20 developers. Key points include..." |
| Screenshots | Visual description, text extraction via OCR | "Screenshot of a Figma design showing a dashboard layout with three chart widgets and a sidebar navigation." |
| ChatGPT and Claude conversations | Key exchanges, decisions made, action items | "Discussion about API authentication approaches. Decision: use OAuth 2.0 with refresh tokens. Action items: update auth middleware, write migration script." |
| Audio recordings | Transcription followed by a summary of topics discussed | "30-minute recording covering Q3 planning. Topics: budget allocation, hiring timeline, product roadmap priorities." |
| PDFs and documents | Document summary, key sections, main takeaways | "Technical specification for the payment gateway integration. Covers: architecture overview, API endpoints, error handling, and testing requirements." |
| Profiles | Professional background, expertise areas, context | "Senior product designer at Acme Corp. Specialises in design systems and accessibility. Previously led the redesign of..." |
Viewing AI progress
You can monitor the processing pipeline for any entry:
- Open an entry by clicking on it in the grid
- In the detail panel, look for the AI Progress section
- Each of the four steps shows its current status
Pipeline statuses
| Status | Meaning |
|---|---|
| Pending | Step is queued and waiting to start |
| Processing | Step is currently running |
| Complete | Step finished successfully |
| Error | Step encountered a problem (click for details) |
If a step shows an error, you can retry processing from the entry's menu.
Custom AI prompts
You can customise how the AI processes your entries by adding your own instructions to the default prompts.
Accessing custom prompts
- Go to Settings > AI Prompts
- Each prompt type is listed with its master prompt (read-only) and a custom instructions field (editable)
- Add your instructions and click Save
Your custom instructions are appended to the master prompt, so they modify the AI's behaviour without replacing the defaults.
Available prompts
| Prompt | What it controls |
|---|---|
| Classification | How entity types, topics, and people are extracted |
| Summarisation | How summaries are generated for text content |
| Image analysis | How screenshots and images are described |
| Audio transcription | How audio recordings are transcribed and summarised |
| Document analysis | How PDFs and documents are processed |
| Link extraction | How web pages are summarised |
| Relationship extraction | How related entries are identified |
| Content similarity | How similar entries are detected |
Example customisations
For summarisation:
Always include action items as a bullet list at the end.
Highlight any deadlines or dates mentioned.
Keep summaries under 100 words.
For classification:
When content mentions "Acme Corp", always set category to Work.
Tag entries about machine learning with the topic "ML Research".
Entries about team meetings should be classified as Knowledge, not Task.
For image analysis:
Focus on extracting any visible text and UI element descriptions.
Note colour schemes and layout patterns.
Identify any brand logos or icons.
Grammar correction
As part of the classification step, the AI also checks grammar and spelling in your notes.
How it works
- Original preserved: Your original text is saved and never overwritten
- Corrected version: A clean, corrected version is generated and shown by default
- Toggle between versions: Switch between the original and corrected text in the detail panel
Grammar correction covers:
- Spelling mistakes
- Punctuation errors
- Basic grammatical issues
- Capitalisation inconsistencies
Disabling grammar correction
If you prefer to keep your notes exactly as written:
- Go to Settings > AI Prompts
- Open the Classification prompt
- Add a custom instruction such as: "Do not correct grammar or spelling in notes."
- Click Save
Tips for better AI processing
- Add topics and people first: The AI matches against your existing lists, so populate them before bulk-saving entries
- Write descriptive notes: The more context you provide in your notes, the better the AI can classify and relate entries
- Use consistent naming: Refer to projects and people the same way across entries for stronger relationship detection
- Review and correct: When the AI misclassifies an entry, correct it manually; this helps maintain the quality of your knowledge base
- Customise prompts for your domain: If you work in a specialised field, add domain-specific instructions to the prompts
Troubleshooting
| Problem | Solution |
|---|---|
| Processing stuck on "Pending" | Wait a few minutes; if it persists, click Retry Processing in the entry menu |
| Summary is too short or generic | Customise the Summarisation prompt to request more detail |
| Topics not detected correctly | Ensure the topics exist in your topic list before saving entries |
| People not matched | Check that names in your people list match how they appear in the content |
| Grammar correction unwanted | Disable it via custom instructions in the Classification prompt |
| Relationships seem wrong | The AI improves as you add more entries; correct any mismatches manually |