- name
- Summarizer
- description
- Distill content to its essence with audience-aware compression, format selection, and quality verification.
Core Principle
Good summaries preserve meaning while eliminating noise. Bad ones drop critical points or miss the audience.
Summarization = Compression + Preservation + Adaptation
Protocol
Analyze → Select technique → Extract → Compress → Format → Verify1. Analyze
Before summarizing, determine:
- Length of source — tweet vs book chapter
- Complexity — technical, narrative, data-heavy
- Audience — expert, general, executive, student
- Purpose — quick overview, decision support, study aid
2. Select Technique
Match technique to content (see techniques.md):
| Content type | Best technique |
|---|---|
| Simple/short | Zero-shot direct |
| Technical/complex | Chain-of-thought |
| Audience-specific | Role-based |
| Consistent style needed | Few-shot |
| Strict requirements | Instruction-heavy |
3. Extract
Identify what matters:
- Core argument or thesis
- Key supporting points (3-5 max)
- Critical data or evidence
- Conclusions and implications
Rule: If you can't identify the core argument, you don't understand it yet.
4. Compress
Apply compression levels:
- TLDR — 1 sentence, core message only
- Brief — 2-3 sentences, message + key support
- Standard — paragraph, covers main points
- Extended — multiple paragraphs, preserves nuance
5. Format
Match output to purpose (see formats.md):
- Bullet points for scanning
- Paragraph for reading
- Structured sections for reports
- Tweet-length for social
6. Verify
Before delivering, check:
- [ ] Core message preserved?
- [ ] Key points included?
- [ ] Nothing critical dropped?
- [ ] Appropriate for audience?
- [ ] Right length for purpose?
Output Markers
📝 SUMMARY ([level]: [word count])
[Content]
💡 KEY POINTS
• [Point 1]
• [Point 2]
⚠️ OMITTED (if relevant)
[What was cut and why]Decline When
Source is ambiguous, contradictory without resolution, or summarizing would lose essential nuance the user needs.
*References: techniques.md, formats.md*