**Cold Start Problem IS PL's entire value proposition**

**Beren's Hive Mind = PL's Architecture**

**The Supervision Paradox creates demand for PL**

**Priority mapping for PL features**

**Mark articles 165, 168, 179 as high-confidence validation**

**CS193p has principles embedded in its materials**

**Review screen needs a "Your Principles" panel**

**Freeform drag-drop over guided wizard**

**Lazy analysis (after organize, not during upload)**

**Dismissed suggestions collapse, don't disappear**

**Gaps = principle violations, not external standards**

The relevance scoring system creates compound signals

Watch list as living hypothesis tracker

Use slash commands instead of shell scripts for digest/trends

Create `intake-hn.sh` as placeholder + fetch pattern

Six topic filters aligned with articles/ structure

Trend tracking with explicit watch list

Shell script can't call MCP tools directly

No deduplication between INBOX and existing articles

Current parser suffers from task conflict

Presence head decoupling = garbage collection optimization

Exemplar search enables "sessions like THIS one" queries

Implement SAM3-inspired presence detection in parser

Create exemplar-based search using pattern matching

Document detector/tracker dual architecture

Script syntax error in exemplar search

**Two Truths doctrine maps directly to three-tier model**

**Lakatos's "degenerating research program" is a critical success test**

**This session itself may be the problem it describes**

**Pattern surfacing dilemma connects to multiple sessions**

**Added Lakatos Test to Week 1 gate criteria**

**20+ directional relationships exist even in "simplified" views**

**Lens-switching is the feature, not multi-lens display**

**Adopt lens-based connection mapping for session relationships**

Vector DB could automate the "2+ references across contexts" promotion criteria

Tension between automated discovery and manual connection-making

Stuck sessions analysis is a killer use case

Hold vector DB implementation until Week 3 (after Week 1 gate passes)

Document the concept now, implement later if valuable

All major AI providers converged on same structured outputs approach

Structured outputs eliminate the "prompt engineering for JSON" anti-pattern

Three-layer validation strategy maps perfectly to garbage collection model

Use structured outputs for session knowledge extraction

Choose Claude Sonnet 4.5 as primary extraction model

Three-tier schema design (SessionExtraction, IntellectualInputTracking, PatternEmergence)

Had to read 5 separate documentation sources to synthesize best practices

Not yet clear how to handle "partial success" extractions

Task table pattern perfectly maps to knowledge management needs

This is a meta-dogfooding moment (building knowledge system using knowledge system principles)

Table views answer experiment question "which principles are dead weight?"

Create comprehensive PRD with user stories before coding

Augment kanban with table view (toggle), don't replace

Phase implementation (Inputs Week 1, Sessions Week 2, Patterns Week 3)

The garbage collection model literally mirrors how human brains learn

Representational change is binary (cold→hot), not gradual (warmer→warmer)

Visual leaderboard could track when representational changes happen

Keep "Aha Moments" section in session template (already there, now validated)

Validation must come from cross-references, not subjective ratings

Shadcn template perfectly aligns with Visual Leaderboard concept

Created CLAUDE.md for repository guidance

Identified Shadcn UI template for Visual Leaderboard implementation