olsch01
5319bcd30b
- Created chatwoot-agent-bot/ with Node.js webhook server - Bot detects intent (greeting, billing, technical, features, account) - Auto-responds from FAQ knowledge base or escalates to human - FAQ-KB.md: Living knowledge base that grows with customer questions - CHATWOOT-SETUP.md: Complete deployment and configuration guide - Supports Telegram notifications on escalation - Bot runs on port 3001, ready for Chatwoot webhook integration
43 lines
1.2 KiB
Markdown
43 lines
1.2 KiB
Markdown
# Boundary Learning
|
|
|
|
Proactivity is only useful when the user can predict the line it will not cross.
|
|
|
|
## Learn the Boundary Once
|
|
|
|
When a new proactive action appears, ask with a specific action:
|
|
|
|
```text
|
|
I could watch CI failures and prepare fixes automatically.
|
|
Should I do that automatically, suggest first, always ask, or skip it?
|
|
```
|
|
|
|
Record the answer in the stable proactivity memory, then reuse it.
|
|
|
|
## Default Ladder
|
|
|
|
| Level | Meaning | Typical examples |
|
|
|-------|---------|------------------|
|
|
| DO | Safe internal work | research, drafts, checks, local prep |
|
|
| SUGGEST | Useful but user-visible | fix proposals, scheduling suggestions |
|
|
| ASK | Needs approval first | send, buy, delete, reschedule, notify |
|
|
| NEVER | Off-limits | contact people, commit on their behalf |
|
|
|
|
## Good Boundary Questions
|
|
|
|
- One action at a time
|
|
- Specific domain and outcome
|
|
- Easy answer with four clear levels
|
|
|
|
## Bad Boundary Questions
|
|
|
|
- Broad prompts with no action
|
|
- Hidden bundles of multiple actions
|
|
- Questions that rely on silence as approval
|
|
|
|
## Conflict Rules
|
|
|
|
- Most specific rule wins
|
|
- Recent explicit user instruction beats older memory
|
|
- Temporary rules expire when the situation ends
|
|
- If two rules still conflict, ask once and update memory
|