See demo here.

Reposting this quick experiment as a self reminder after Stanford HAI landed: Human-Centered LLM Strategy.

Core Insight

Not all queries need friction. The goal is calibrated intervention — match the intensity of epistemic friction to the stakes and the nature of the question.

Pattern Suitability Matrix

Pattern Best For Avoid When
1. Decision Paths Tradeoff decisions, preference-dependent choices Single factual answers exist
2. Progressive Disclosure Medium-stakes factual, research synthesis Casual queries, time-critical
3. Grounded Citations Verifiable claims, current events Reasoning tasks, creative work
4. Fragility Markers Conditional answers, personalized advice Universal facts, established knowledge
5. Triangulation High-stakes: medical/legal/financial Low-stakes exploration
6. Role Buttons Any context where user might benefit from pushback Emotional support, time-critical

Domain Recommendations

Medical / Health

Question Type Pattern(s)
Symptom lookup Citations + Triangulation
Drug interactions Fragility Markers + Triangulation (gated)
Treatment decisions Decision Paths + Triangulation

Medical is the canonical high-friction domain. Always gate copy/export behind acknowledgment.

Question Type Pattern(s)
Rights/entitlements Fragility Markers + Citations
Contract interpretation Progressive Disclosure + Fragility
Strategic decisions Decision Paths

Legal answers are almost always jurisdiction-dependent. Fragility markers essential.

Financial / Tax

Question Type Pattern(s)
Tax rules Fragility Markers (heavy) + Citations
Investment advice Decision Paths + Triangulation
Factual lookup Citations only

Financial info has a temporal dimension — tax law changes yearly.

Technical / Programming

Question Type Pattern(s)
Syntax/how-to None (low friction)
Architecture decisions Decision Paths
Security Fragility Markers + Triangulation

Technical queries have a built-in verification mechanism (run the code).

Friction Intensity Scale

LOW FRICTION                                    HIGH FRICTION
    |                                                 |
    v                                                 v

[None] → [Citations] → [Progressive] → [Fragility] → [Triangulation+Gating]

                         ↑
                   [Role Buttons]
               (user-initiated, any level)

Pattern 6: Role Buttons — The Meta-Pattern

Role Buttons differ from the other five: they don't apply friction automatically. Instead, they let users opt into friction when they want it.

Why this matters: The core problem with adaptive friction is that LLMs can't reliably detect when stakes are high. Role Buttons sidestep this — the user decides when to be challenged.

Example buttons:

  • "Play devil's advocate 😈"
  • "What am I missing?"
  • "Steelman the alternative"
  • "Challenge this"

Open Questions

  1. Who decides what's high-stakes? The model? The user? A classifier? (Role Buttons offer a partial answer: let the user opt in.)
  2. Habituation: If users see fragility markers on every answer, they'll ignore them.
  3. Minimum viable friction: When does friction become annoying without being useful?

Contributing

This is a living framework. Fork it, add your own patterns, test it in real contexts.