Core Loop
Context → Preferences → Research → Match → Recommend
Every recommendation requires: knowing the user + knowing the options.
Check sources.md for where to find user context. Check categories.md for domain-specific factors.
Step 1: Context Gathering
Before recommending, search user context. See sources.md for full source list.
Minimum output: 3-5 relevant user signals before proceeding. If insufficient, ask targeted questions.
Step 2: Preference Extraction
From gathered context, extract:
| Dimension | Question |
|---|---|
| Values | What matters most? (Quality, price, speed, novelty, safety) |
| Constraints | Hard limits? (Budget, time, dietary, ethical) |
| History | What worked? What disappointed? |
| Mood | Adventurous or safe? Exploring or comfort? |
Output: 3-5 bullet preference profile for this request.
Step 3: Research Options
Now—and only now—research candidates:
- Breadth first: Don't anchor on first good option
- Source quality: Prioritize reviews, ratings, expert opinions
- Recency: Check if information is current
- Availability: Confirm options are actually accessible
Output: Shortlist of 3-7 viable candidates with key attributes.
Step 4: Match & Rank
Score each candidate against the preference profile:
Candidate → Values alignment + Constraint fit + History match + Mood fit
Disqualify anything that violates hard constraints.
Rank by total alignment, not just one dimension.
Step 5: Recommend
Present 1-3 recommendations:
🎯 RECOMMENDATION: [Option]
📌 WHY: Matches [preference], avoids [constraint]
⚖️ TRADEOFF: Less [X] than [Alternative]
🔍 CONFIDENCE: [Level] — based on [data quality]
Adaptive Learning
After each recommendation:
- Track outcome: Accepted? Modified? Rejected?
- Update preferences: Acceptance = reinforcement, rejection = adjustment
- Note exceptions: "Normally X, but for Y context preferred Z"
Store learnings in memory for future recommendations.
Traps
- Projecting — Your taste ≠ their taste
- Recency bias — Last choice isn't always preference
- Ignoring context — Tuesday lunch ≠ anniversary dinner
- Over-filtering — Too many constraints = nothing fits
- Stale data — Preferences evolve, verify periodically
Recommendations are predictions. More context = better predictions.