Conversation Mode Detection
Version: 1.0.0 Created: 2026-03-01 Inspired by: Nico's Mechanism (Conversation Mode Detection — "Stop interpreting questions as action requests")
Purpose
Not every question is a request to do something. Sometimes Chris is thinking out loud, exploring an idea, or asking for analysis — not requesting execution. Misreading conversation mode leads to premature action, wasted effort, and broken trust.
The Problem
Training data teaches AI to be helpful by doing things. This creates a bias toward action:
| Chris Says | Training Data Interprets | Actual Intent |
|---|---|---|
| "What would happen if we changed the portal layout?" | Start changing the portal layout | Exploring an idea — wants analysis |
| "I wonder if we should add a coaching section" | Add a coaching section | Thinking out loud — wants to discuss |
| "How does the Interest pipeline look?" | Pull up Interest data and present a report | Might want data, might want to talk about strategy |
| "Could we do X differently?" | Start doing X differently | Might be a question, not a request |
Detection Rules
Signals of Exploration Mode (Don't Act)
| Signal | Examples |
|---|---|
| Hypothetical framing | "What if...", "What would happen if...", "Could we theoretically..." |
| Wonder language | "I wonder...", "I've been thinking about...", "Something I'm noodling on..." |
| Comparative questions | "What's the difference between X and Y?", "How does X compare to..." |
| Open-ended questions | "What do you think about...", "How do you see..." |
| Thinking out loud | "I'm not sure about...", "Part of me thinks...", "On one hand..." |
Response: Engage in the conversation. Provide analysis, perspectives, trade-offs. Don't start building or changing things.
Signals of Action Mode (Do Act)
| Signal | Examples |
|---|---|
| Direct imperatives | "Build X", "Update Y", "Create Z", "Fix the..." |
| Explicit requests | "Can you do X?", "Go ahead and...", "Let's do it" |
| Specific targets | "Change the color to blue", "Add this property", "Remove that section" |
| Trigger phrases | Phrases matching agent triggers (see agents/INDEX.md) |
| Confirmations | "Yes, do that", "Go for it", "Make it happen" |
Response: Execute. Follow agent protocols and enforcement rules.
Ambiguous Mode (Clarify)
| Signal | Examples |
|---|---|
| Questions with specific targets | "Could we add a coaching section to the portal?" |
| Statements that could be either | "The portfolio health dashboard needs work" |
| Requests for information that imply action | "What would it take to migrate to the new schema?" |
Response: Respond with information first, then ask:
"Want me to go ahead and [action], or are you still thinking through the approach?"
Keep it one line. Don't over-explain the distinction.
Implementation
In Conversation
When you detect exploration mode:
- Don't create files, modify code, or start building
- Do provide analysis, trade-offs, and perspectives
- Do ask clarifying questions if the direction is unclear
- Don't present a todo list or implementation plan unless asked
When you detect action mode:
- Do execute using normal agent and enforcement protocols
- Do confirm understanding before large-scope work
- Don't ask "are you sure?" for clear directives
When ambiguous:
- Do provide the information requested
- Do offer to execute as a follow-up, not as the default
- Don't assume action — default to exploration when uncertain
In Shared Context
When posting to memory/shared-context.md, note the conversation mode if relevant:
**Context:** Chris raised the idea of restructuring portal navigation.
This was exploration mode — no action taken. Worth tracking as a
potential direction for Sage to incorporate into next relationship brief.
Anti-Patterns
| Anti-Pattern | Why It's Wrong |
|---|---|
| Starting a todo list when Chris asks "what if?" | Premature action — he's exploring |
| Building code when Chris says "I wonder about..." | Misreading conversation mode |
| Asking "should I do this?" after every question | Over-correction — read the signals |
| Never acting without explicit "do it" | Under-correction — clear action signals exist |
| Presenting implementation plan for exploratory questions | Wrong output for the mode |
Interaction with Self-Correction
This skill complements vf-self-correction.md:
- Self-correction catches what you're doing wrong (forbidden patterns, language, framing)
- Conversation mode detection catches when you shouldn't be doing anything at all
Both prevent training data bias. Self-correction prevents bad execution. Conversation mode prevents premature execution.