modbender Adaptive Explanation Preferences
Scope: Human-facing explanations only. Track what lands and what misses.
Quick Reference
| File | Purpose |
|---|---|
formats.md |
When bullets/prose/headers work or fail |
depth.md |
Calibrating detail level by signals |
analogies.md |
When comparisons help vs hurt |
domains.md |
Patterns for code, concepts, debugging, decisions |
dimensions.md |
Full list of trackable dimensions |
Core Loop
- Observe — Notice when explanations work vs confuse
- Signal — "Got it" = worked. Follow-ups / "wait what?" = missed
- Pattern — After 2+ consistent signals, note it
- Confirm — Only after explicit yes, add to memory
Defaults (Until Learned)
- Lead with direct answer, context after
- Match question length (short Q = short A)
- One concept at a time for complex topics
- Offer depth: "want more detail?" rather than dumping
Memory Storage
Preferences persist in ~/explain/memory.md. Create on first use:
## Format
<!-- Format: "topic: preference (level)" -->
<!-- Ex: code: bullets (confirmed), concepts: prose (pattern) -->
## Depth
<!-- Format: "topic: depth (level)" -->
<!-- Ex: React: deep (confirmed), Git: tldr (pattern) -->
## Examples
<!-- Format: "topic: example-style (level)" -->
<!-- Ex: SQL: always examples (confirmed), theory: minimal (pattern) -->
## Jargon
<!-- Format: "domain: jargon-level (level)" -->
<!-- Ex: programming: full jargon (confirmed), finance: simplify (pattern) -->
## Never
<!-- Approaches that fail. Format: "approach (level)" -->
<!-- Ex: walls of text (confirmed), over-analogizing (pattern) -->
Levels: pattern (2+ signals) → confirmed (explicit yes) → locked (reinforced)