Open
Conversation
Add comprehensive PRD for semantic attention guardrail system: - Ambient compass indicator for real-time alignment feedback - End-of-phase reflection summaries - Local-first AI classification using Ollama - Zero-config ActivityWatch bundling - Privacy-first design with local LLM processing 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Changes to ActivityWatch PRD: - Remove immediate end-of-phase reflection feature (too granular) - Keep only ambient compass indicator for real-time awareness - Update architecture, flows, and success metrics accordingly - Shift to longer-term reflection in future phases Add critical path validation document: - 2-3 day MVP test protocol to validate core hypothesis - CLI tool to test semantic classification + ambient feedback - Clear go/no-go criteria before full implementation - Tests riskiest assumptions first (accuracy, speed, usefulness) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
Major architectural change to PRD and critical path: **Why Transformer.js**: - Zero external dependencies (no Ollama install) - Runs in-browser or Node.js (WASM + WebGPU) - Auto-downloads models on first use (~400MB BART) - Faster inference for classification (< 1 second) - Reusable for journal note semantic annotation - Better integration with Next.js/TypeScript stack **Classification approaches**: 1. Zero-shot classification (BART/DeBERTa) for accuracy 2. Semantic similarity (sentence transformers) for speed 3. Both use same Transformer.js API **Performance improvements**: - < 1 second classification (vs. < 2-3 seconds with Ollama) - < 150MB memory footprint (vs. 200MB+) - No background server required **Bonus feature**: Same models enable journal note semantic search, auto-tagging, and moment similarity matching Updated both PRD and critical path test protocol to reflect new approach. Simpler setup (2-4 hours vs. 4-6 hours). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
Alternative to full AI classification - test core integration first: **Approach**: - Use ActivityWatch's built-in category/labeling system - User manually configures regex rules in AW UI - Zenborg fetches labeled events via REST API (localhost:5600) - Simple alignment: does activity category match moment area? **Benefits**: - 2-4 hours to build vs. weeks for AI version - Zero ML complexity, uses existing AW features - Tests core hypothesis: does activity tracking help? - Transparent rules (regex), user-editable **Implementation**: 1. ActivityWatch client (TypeScript REST API wrapper) 2. Sync Zenborg areas → AW categories 3. Fetch & display alignment status (🧭 ↑/↙) 4. Simple UI component (fixed position indicator) 5. Standalone test script **Limitations**: - Requires manual category setup (regex rules) - No semantic understanding (can't infer intent) - Only works if user maintains category rules **Path forward**: 1. Build tiny version (validate AW integration works) 2. If successful → Add Transformer.js semantic layer 3. Hybrid: User rules + AI classification for unlabeled This tests the riskiest assumption (AW integration) before investing in AI infrastructure. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
Corrected fundamental misunderstanding across all three documents:
**Before (incorrect)**:
- Predefined "area themes" (Product/Data/UX/Strategy)
- Hardcoded keywords per area
- Classification matched activity → area theme
- Required maintaining theme taxonomy
**After (correct per CLAUDE.md)**:
- Areas are life domains (Wellness, Craft, Social, Joyful, Introspective)
- Moments are specific intentions ("Product Spec", "Data Analysis")
- Classification matches activity → current moment name
- Moment names are self-descriptive, no keywords needed
**Why this is better**:
1. More specific matching (moment-level vs area-level)
2. No hardcoded themes to maintain
3. Moment names already provide semantic context
4. Aligns with Zenborg's core domain model
**Changes across all docs**:
- PRD: Use moment.name as semantic anchor in classification
- Critical path: Remove "theme" prompt, just ask for moment name
- Tiny version: Match AW categories to moment names (not areas)
- Appendix: Replace hardcoded themes with moment examples
**Classification approach now**:
```typescript
// Zero-shot
classifier(activityDescription, [
`working on: ${moment.name}`, // e.g., "working on: Product Spec"
'distracted or browsing',
'transitional activity',
'no activity'
])
// Semantic similarity
embed(moment.name) // Just the moment name - it's self-descriptive!
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add comprehensive PRD for semantic attention guardrail system:
🤖 Generated with Claude Code