Taste Engine Handbook
Taste Engine is a canon-and-judgment system that helps teams maintain alignment between their stated doctrine (what they say they believe) and their produced artifacts (what they actually ship).
The Problem
Section titled “The Problem”Creative and product teams accumulate doctrine over time — thesis statements in READMEs, design rules in architecture docs, anti-patterns in post-mortems. But as artifacts multiply, drift creeps in. A feature brief contradicts the product thesis. A naming convention slides. A design rule gets quietly ignored.
Taste Engine closes this gap by:
- Extracting canon statements from your existing doctrine
- Curating those statements through human review
- Reviewing new artifacts against curated canon
- Repairing drifted work with targeted suggestions
- Gating workflows to prevent drift from shipping
- Scaling governance across an entire portfolio
How It Works
Section titled “How It Works”The engine runs locally against Ollama. No paid API. No cloud calls. Your doctrine stays on your machine.
The pipeline flows through these stages:
Source docs → Extract → Curate → Freeze → Review → Repair → Gate ↓ Portfolio → Org → WatchtowerEach stage has dedicated CLI commands under taste <group>.
Key Design Decisions
Section titled “Key Design Decisions”- Rules beat the model. The verdict ladder is deterministic — the LLM scores dimensions, but rules synthesize the final verdict. The model cannot talk itself into a pass.
- Human curation is required. Extracted canon is always “proposed” until a human accepts it. No auto-accept path exists.
- 4 dimensions, not 1. Every review scores thesis preservation, pattern fidelity, anti-pattern collision, and voice/naming fit independently before synthesis.
- 3 repair modes. Drift severity determines the repair path: surface patch, structural repair, or goal redirection.
Handbook Contents
Section titled “Handbook Contents”- Getting Started — Install, configure Ollama, run your first review
- Extraction & Curation — How canon gets extracted and curated
- Review & Verdicts — The 4-dimension scoring and verdict synthesis
- Repair & Redirection — Fixing drifted artifacts
- Workflow Gate — Enforcement modes and CI integration
- Portfolio & Org — Multi-repo governance
- CLI Reference — All 68 commands