Skip to content

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).

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:

  1. Extracting canon statements from your existing doctrine
  2. Curating those statements through human review
  3. Reviewing new artifacts against curated canon
  4. Repairing drifted work with targeted suggestions
  5. Gating workflows to prevent drift from shipping
  6. Scaling governance across an entire portfolio

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 → Watchtower

Each stage has dedicated CLI commands under taste <group>.

  • 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.