182 lines
4.9 KiB
Markdown
182 lines
4.9 KiB
Markdown
# Research Skill Enhancement - Requirements
|
|
|
|
## Core Principles
|
|
|
|
1. **Discovery-driven, not list-driven** - Find best sources wherever they are
|
|
2. **Context-aware** - Grounded in user's domain preferences
|
|
3. **Finds unknowns** - Value is in discovering what user doesn't know
|
|
4. **No time limits** - Can run 30 min to 3 hours
|
|
5. **Exhaustive** - Don't stop until topic is exhausted
|
|
|
|
## Domain Contexts
|
|
|
|
### What Context Provides
|
|
- **Relevance filtering** - what matters to this user
|
|
- **Constraint awareness** - their specific resources/limitations
|
|
- **Optimization targets** - what to maximize for them
|
|
|
|
### Domain: Travel
|
|
- Loyalty: Alaska MVP Gold 75K, Marriott Titanium Elite
|
|
- Points: Current balances (fetch via browser-control)
|
|
- Credit cards: Alaska card, Marriott Amex, Chase Sapphire Preferred
|
|
- Memberships: Epic Pass
|
|
- Preferences: "Luxury at rock-bottom prices", never book only research
|
|
|
|
### Domain: Shopping
|
|
- Quality standards
|
|
- Budget ranges
|
|
- Brand preferences
|
|
|
|
### Domain: Work/Education
|
|
- PSD context (CIO role)
|
|
- UDL expertise
|
|
- Edtech landscape
|
|
- District priorities
|
|
|
|
### Domain: AI/Coding
|
|
- Tech stack preferences
|
|
- Languages/frameworks
|
|
- Architecture patterns
|
|
|
|
### Domain: Consulting
|
|
- Service offerings
|
|
- Client contexts
|
|
|
|
## Research Engine Architecture
|
|
|
|
### Phase 1: Query Understanding
|
|
- Detect domain from query
|
|
- Load relevant context
|
|
- Decompose into sub-questions (5-10 angles)
|
|
- Identify what types of sources needed (academic, forum, video, etc.)
|
|
|
|
### Phase 2: Parallel Multi-Source Discovery
|
|
- Launch multiple search strategies simultaneously:
|
|
- Multiple search engines (Google, Bing, DuckDuckGo)
|
|
- Multiple query formulations
|
|
- Different source types (articles, forums, videos, podcasts)
|
|
- Use multiple LLMs for different perspectives:
|
|
- Perplexity: Current web info with citations
|
|
- Gemini: Multi-perspective synthesis
|
|
- OpenAI: Structured analysis
|
|
- Claude: Deep reasoning
|
|
|
|
### Phase 3: Source Evaluation & Expansion
|
|
- Evaluate each source for:
|
|
- Credibility (author expertise, publication reputation)
|
|
- Recency (when published/updated)
|
|
- Depth (surface vs comprehensive)
|
|
- Citations (does it cite others? is it cited?)
|
|
- Follow promising leads:
|
|
- Sources referenced by good sources
|
|
- Authors who appear multiple times
|
|
- Cross-referenced claims
|
|
|
|
### Phase 4: Deep Scraping
|
|
- Don't stop at first page of results
|
|
- Go 5-10 pages deep on good queries
|
|
- Use browser-control for:
|
|
- JS-rendered content (Reddit, forums)
|
|
- Authenticated pages (user accounts)
|
|
- Sites that block scrapers
|
|
- Follow internal links on valuable sources
|
|
|
|
### Phase 5: Multimedia Discovery
|
|
- YouTube videos
|
|
- TikTok content
|
|
- Podcasts (Spotify, Apple)
|
|
- Not just text articles
|
|
|
|
### Phase 6: Synthesis
|
|
- Organize by theme/question
|
|
- Every claim cited
|
|
- Note consensus vs disagreement
|
|
- Include multimedia resources
|
|
- Provide actionable recommendations
|
|
- Flag what's still uncertain
|
|
|
|
## Output Format
|
|
|
|
```markdown
|
|
## Research: [Topic]
|
|
|
|
### Context Applied
|
|
- [Domain context that was loaded]
|
|
- [Dynamic data fetched - e.g., current point balances]
|
|
|
|
### Executive Summary
|
|
[2-3 paragraph overview of key findings]
|
|
|
|
### Detailed Findings
|
|
|
|
#### [Sub-topic 1]
|
|
[Deep analysis with inline citations]
|
|
|
|
#### [Sub-topic 2]
|
|
[Deep analysis with inline citations]
|
|
|
|
### Multimedia Resources
|
|
- [Video: Title](url) - description
|
|
- [Podcast: Title](url) - description
|
|
- [TikTok: @user](url) - description
|
|
|
|
### Recommendations
|
|
1. [Actionable recommendation based on user's context]
|
|
2. [Another recommendation]
|
|
|
|
### What I Discovered You Might Not Know
|
|
- [Surprising finding 1]
|
|
- [Surprising finding 2]
|
|
|
|
### Confidence Assessment
|
|
- High confidence: [topics]
|
|
- Needs verification: [topics]
|
|
- Conflicting information: [topics]
|
|
|
|
### All Sources
|
|
[Complete list of every URL referenced]
|
|
```
|
|
|
|
## Technical Implementation
|
|
|
|
### Files to Create/Modify
|
|
- [ ] Update preferences.json with domain contexts
|
|
- [ ] Create domain context loader
|
|
- [ ] Create multi-LLM research orchestrator
|
|
- [ ] Create source evaluator
|
|
- [ ] Create deep scraper (uses browser-control)
|
|
- [ ] Create multimedia searcher
|
|
- [ ] Create synthesis engine
|
|
- [ ] Update SKILL.md with new architecture
|
|
|
|
### API Keys Needed
|
|
- PERPLEXITY_API_KEY
|
|
- GEMINI_API_KEY
|
|
- OPENAI_API_KEY
|
|
- (Claude runs natively)
|
|
|
|
### Browser-Control Integration
|
|
- Fetch current balances (Marriott, Alaska, Chase)
|
|
- Scrape full forum threads
|
|
- Access authenticated content
|
|
- Handle JS-rendered sites
|
|
|
|
## Success Criteria
|
|
|
|
1. Can run a travel research query and get:
|
|
- Current point balances fetched
|
|
- Transfer bonus opportunities identified
|
|
- 5-6 hotel options with points AND cash prices
|
|
- Flight options with miles needed
|
|
- Deep user reviews from forums
|
|
- Video content discovered
|
|
- Optimization recommendations
|
|
|
|
2. Report takes 30-60 minutes to generate (not 5 minutes)
|
|
|
|
3. Sources include things user didn't know existed
|
|
|
|
4. Every claim is cited
|
|
|
|
5. Recommendations are specific to user's context
|