Preference Elicitation
Input: $ARGUMENTS (domain or guess category to elicit preferences for)
Purpose
Discover what the user values by presenting concrete choices rather than asking them to articulate preferences directly.
Principle: People reveal preferences through choices better than through descriptions. Present trade-offs, observe selections.
Step 1: Generate Trade-Off Pairs
For the domain/category, generate pairs of options that trade off against each other:
TRADE-OFF 1: [dimension being traded]
├── Option A: [high on X, low on Y]
└── Option B: [low on X, high on Y]
TRADE-OFF 2: [dimension being traded]
├── Option A: [high on X, low on Y]
└── Option B: [low on X, high on Y]Minimum: Generate 10 trade-off pairs covering different preference dimensions.
Step 2: Preference Dimensions to Cover
| Dimension | Trade-Off | Option A | Option B |
|---|---|---|---|
| Time vs Money | Spend time to save money OR spend money to save time | Time-rich, money-poor | Money-rich, time-poor |
| Risk vs Reward | Lower risk, lower reward OR higher risk, higher reward | Safe, modest gains | Risky, potential big gains |
| Control vs Convenience | Full control, more effort OR less control, easier | DIY, full control | Done-for-you, hands-off |
| Speed vs Quality | Fast, good enough OR slow, excellent | Quick and dirty | Slow and polished |
| Autonomy vs Security | Independent, uncertain OR employed, stable | Freedom, risk | Stability, constraints |
| Public vs Private | Visible, social proof OR hidden, privacy | Build in public | Work in private |
| Generalist vs Specialist | Breadth, flexibility OR depth, expertise | Know many things | Know one thing deeply |
| Short-term vs Long-term | Immediate payoff OR delayed larger payoff | Quick wins | Long-term building |
| Solo vs Team | Work alone, keep all OR work with others, share | Full ownership | Shared effort/reward |
| Scale vs Craft | Many customers, less personal OR few customers, high-touch | Volume | Boutique |
Step 3: Present as Scenarios, Not Abstractions
Don’t ask: “Do you prefer risk or safety?” Instead present: “Would you rather have a 90% chance of making $50K or a 20% chance of making $500K?”
For each trade-off, generate 2-3 concrete scenarios:
SCENARIO SET: [trade-off dimension]
Scenario A: [concrete situation embodying option A]
- What happens: [specific outcome]
- You get: [specific benefit]
- You give up: [specific cost]
Scenario B: [concrete situation embodying option B]
- What happens: [specific outcome]
- You get: [specific benefit]
- You give up: [specific cost]
User selects: [A / B / NEITHER / BOTH]Step 4: Infer Preferences from Selections
| Selection | Inferred Preference | Confidence |
|---|---|---|
| Chose A in Scenario 1 | Values X over Y | MEDIUM |
| Chose B in Scenario 2 | Values Y over X | MEDIUM |
| Chose A in Scenario 3 | Confirms: Values X | HIGH |
Consistency check: If user chose A in Scenario 1 but B in Scenario 3, and both test the same dimension, probe for context dependence.
Step 5: Generate Preference Profile
PREFERENCE PROFILE
Strong Preferences (consistent across scenarios):
├── [preference 1]: [evidence from selections]
├── [preference 2]: [evidence from selections]
└── [preference 3]: [evidence from selections]
Weak Preferences (single selection):
├── [preference 4]: [single evidence]
└── [preference 5]: [single evidence]
Context-Dependent (inconsistent):
├── [preference 6]: Prefers X when [context], Y when [other context]
└── [preference 7]: Prefers X when [context], Y when [other context]
Unknown (not tested):
├── [dimension not covered]
└── [dimension not covered]Step 6: Match Preferences to Options
For a list of options (e.g., methods from guess_generation):
| Option | Matches Preferences | Conflicts With | Fit Score |
|---|---|---|---|
| [option 1] | [pref A, pref B] | [pref C] | HIGH/MED/LOW |
| [option 2] | [pref A] | [pref B, pref C] | LOW |
| [option 3] | [pref A, pref B, pref C] | none | HIGH |
Surface better options: If an option matches all preferences but user hasn’t mentioned it, highlight it.
Step 7: Detect Settling
Signs user is settling on suboptimal option:
- Chosen option conflicts with stated preferences
- Unchosen option matches preferences better
- User hasn’t considered high-match options
SETTLING DETECTION
User's current choice: [option X]
Preference match: [score]
Better matches not considered:
├── [option Y]: matches [preferences], score [higher]
├── [option Z]: matches [preferences], score [higher]
Why user might be settling:
├── [didn't know about option Y]
├── [assumed constraint that doesn't exist]
├── [first viable option, stopped looking]
Recommendation: Present [option Y, Z] before user commits to [option X]Output Format
## PREFERENCE ELICITATION: [domain]
### Trade-Offs Presented
[list of scenario pairs]
### Selections Made
[user's choices]
### Preference Profile
STRONG:
- [preference]: [evidence]
WEAK:
- [preference]: [evidence]
CONTEXT-DEPENDENT:
- [preference]: [when X] vs [when Y]
### Option Matching
| Option | Preference Match | Conflicts | Score |
|--------|------------------|-----------|-------|
| ... | ... | ... | ... |
### Better Options to Consider
[options that match preferences but user hasn't considered]
### Settling Risk
[if user is choosing suboptimally, explain why and present alternatives]Execution Checklist
- 10+ trade-off pairs generated
- Concrete scenarios created (not abstractions)
- User selections collected
- Preference profile generated
- Options matched to preferences
- Better options surfaced
- Settling detection run
Next Procedure
→ INVOKE: /boc [user’s current choice] [preference profile]
Execute now: Elicit preferences for the domain provided.