.agents/memory/2026-02-27-onboarding-overhaul-implementation-planning.md

42 lines
2.4 KiB
Markdown

# 2026-02-27 Session Notes
## Onboarding Overhaul Implementation Planning
Nicholai began session to implement the onboarding overhaul (roadmap sections 2.1-2.4), the next priority tier after completing four "on fire" items (defaults, npm gate, hook isolation, non-interactive CLI).
### Problem Context
The current `/onboarding` skill works but suffers from structural issues: no step indicators, no upfront outline, redundant timezone questions, and misleading guidance about manual memory workflows. The `remember` and `recall` skills encourage users to manually capture memories, which duplicates the automatic pipeline and distorts importance scoring.
### Implementation Plan
Five files require modification:
1. **SKILL.md (onboarding)** — Major structural rewrite with new sections:
- Expanded frontmatter describing scope (what it does/doesn't do)
- "What This Skill Does" transparency section
- "Guiding Voice" tone guide (warm, patient, refined; emphasizes "we" language; avoids roleplay to prevent prompt injection warnings)
- "Starting the Interview" section with full 6-step outline before questions
- Step indicators at each phase transition
- Relaxation cues at phase ends ("we can revisit anytime")
- Removed timezone question from Phase 3
- Updated Quick Mode to write "one rule your agent should always follow" instead of memory entries
2. **Remember skill** — Set `user_invocable: false` to hide from users and prevent hook-based auto-remembering that doubles token usage
3. **Recall skill** — Add disclaimer section explaining automatic injection at session start and when manual recall is appropriate (targeted searches, old decisions, debugging)
4. **USER.md.template** — Remove timezone from line 13
5. **AGENTS.md.template** — Remove timezone from line 60
### Key Design Decisions
- Persona guidance framed as tone guidance ("adopt a warm tone") rather than character roleplay to avoid triggering host agent safety heuristics
- "Black wall gateway" aesthetic conveyed through quiet confidence rather than cyberpunk keywords
- Transparency declarations satisfy safety requirements without separate "trust me" blocks
- Remember skill fully hidden; recall remains user-accessible since it's read-only and useful for debugging
### Verification Strategy
After implementation: build packages, verify skill sync, test all three skills in Claude Code, confirm timezone removed from interview flow and templates.