Dev surface

Per-persona use cases

Use cases by persona — every Claude primitive lives in one of these scenarios (Customer has 6, the other three have 2 each). Click through to trigger one; /chat also has a Run tour button that auto-plays the two customer scenarios with Studio forced on.

Skills

Loaded automatically by Claude Code when the in-session conversation matches its description. The engineer asks a question; the skill's expertise becomes available.

• Skilltaste-allergen-taxonomy.claude/skills/taste-allergen-taxonomy/SKILL.md ↗

  Canonical 9-allergen taxonomy + ambiguous-ingredient rulings (does tahini count as sesame?).

  Loads when

  • Asking Claude Code to tag a dish with allergens
  • Explaining allergen reasoning while debugging the Haiku tagger
  • Resolving an ambiguous ingredient question

  Use cases that invoke it

  • "What allergens should I tag on a dish that contains tahini and almond butter?"

    → Skill auto-loads. Returns: sesame (tahini) + tree_nut (almond). Cites the taxonomy doc.

  • "The Haiku tagger missed sesame on a hummus dish — what's the rule?"

    → Skill loads, explains tahini→sesame is a hard rule in the taxonomy and proposes the fix in ALLERGEN_TAGGER_USER.

• Skilltaste-eval-judge.claude/skills/taste-eval-judge/SKILL.md ↗

  How the Opus eval judge scores scenarios — deterministic vs qualitative, calibration heuristics.

  Loads when

  • Reviewing an eval verdict that feels wrong
  • Calibrating a rubric in evals/golden-set.json
  • Debugging why a scenario the agent 'clearly passed' got a low score

  Use cases that invoke it

  • "Why did scenario `vegan-peanut-strict` get a 3/5 from the judge?"

    → Skill loads. Walks the two-layer scoring split, then reads the rubric + the judge transcript to attribute the deduction.

  • "Help me write a rubric that the judge won't over-weight on tone."

    → Skill loads. Returns the calibration heuristics + a rubric template that anchors on outcomes, not phrasing.

• Skilltaste-eval-scenario-author.claude/skills/taste-eval-scenario-author/SKILL.md ↗

  Schema, tag taxonomy, and anti-patterns for authoring a golden-set scenario.

  Loads when

  • Adding a new entry to evals/golden-set.json
  • Fixing a scenario the harness rejected as malformed
  • Translating a real bug into a regression scenario

  Use cases that invoke it

  • "Add a regression scenario covering the F-04 spend-cap block."

    → Skill loads. Produces a schema-valid entry with deterministic constraints (cart total, hook verdict) + optional judge rubric.

  • "The harness rejected my new scenario with 'missing rubric.judge_must_cite' — fix it."

    → Skill loads. Identifies the missing field, explains the anti-pattern, returns the corrected JSON.

• Skilltaste-menu-pdf-extract.claude/skills/taste-menu-pdf-extract/SKILL.md ↗

  Layout heuristics that turn Document Intelligence output into structured Dish records.

  Loads when

  • Running /taste-onboard-restaurant
  • Debugging the ingestMenu queue trigger
  • Asking why a dish was extracted with the wrong price or name

  Use cases that invoke it

  • "DocIntel returned this layout JSON — turn it into Dish records."

    → Skill loads. Applies the column-detection + price-anchor heuristics; returns a Dish[] matching lib/types.ts.

  • "Onboarding Thai Fresh missed half the dishes — what went wrong?"

    → Skill loads. Walks the section-header rules, points at the heuristic that failed, suggests a layout fixture to add.

Commands

Slash command the engineer types in Claude Code. Deterministic workflow — runs scripts, calls models, returns a structured result.

• Command/taste-verify.claude/commands/taste-verify.md ↗

  Deterministic pre-commit suite — typecheck, hook smoke-test, fixture sanity. Zero Claude calls.

  Run it when

  • Run before every commit on Phase 1+
  • Run when CI fails and you need to repro locally

  Use cases that invoke it

  • /taste-verify

    → Runs `npm run typecheck`, smoke-tests the PreToolUse hook with a synthetic cart, sanity-checks fixtures. Reports pass/fail per stage.

• Command/taste-eval.claude/commands/taste-eval.md ↗

  Run the golden-set regression suite. Default mode is deterministic; --judge adds Opus rubric scoring.

  Run it when

  • After changing a prompt in lib/claude.ts
  • Before promoting an ADR-gated change
  • Weekly drift check

  Use cases that invoke it

  • /taste-eval

    → Runs deterministic checks against evals/golden-set.json. ~$0.02 (one Sonnet call inside MenuAdvisor).

  • /taste-eval --judge

    → Also invokes Opus judge per scenario. ~$0.03/scenario. Use when calibrating rubrics.

• Command/taste-cost-report.claude/commands/taste-cost-report.md ↗

  Aggregate Claude-call telemetry from .taste/telemetry.jsonl — model/agent breakdown + cache-hit trend.

  Run it when

  • End-of-session cost review
  • After a Phase that added new Claude calls
  • Debugging a cost spike

  Use cases that invoke it

  • /taste-cost-report

    → Prints per-model, per-agent token + $ breakdown for the local telemetry file. No Claude calls — pure file aggregation.

• Command/taste-handoff.claude/commands/taste-handoff.md ↗

  Produce a tight (<500 word) handoff prompt for the next Claude Code session.

  Run it when

  • End of session
  • /compact recommended
  • Switching between Taste workstreams

  Use cases that invoke it

  • /taste-handoff

    → Reads current branch state + TASKS.md + recent commits, emits a paste-ready handoff covering decisions, files, open questions, next action.

• Command/taste-onboard-restaurant.claude/commands/taste-onboard-restaurant.md ↗

  End-to-end restaurant onboarding — upload PDF to blob, create job, enqueue ingest message.

  Run it when

  • Phase 3+ — adding a new restaurant to the catalog

  Use cases that invoke it

  • /taste-onboard-restaurant ./menus/thai-fresh.pdf "Thai Fresh" 78704

    → Uploads PDF → creates jobs row → enqueues ingestMenu message. Returns in <5s; ingestion (DocIntel → Sonnet extract → Haiku allergen tag → AI Search + Cosmos upsert) runs async.

• Command/taste-promote-place-order.claude/commands/taste-promote-place-order.md ↗

  Architectural change: promote placeOrder from runtime-only to a Claude-callable tool. Requires ADR amendment — does NOT execute silently.

  Run it when

  • Explicit decision to let the model construct placeOrder payloads directly

  Use cases that invoke it

  • /taste-promote-place-order

    → Walks the STOP-list (ADR amendment, schema definition, hook re-validation, eval regression). Refuses to mutate TOOL_SCHEMAS until each is checked.

MCP server

A parallel surface to /chat — same underlying tools, different consumer. External MCP clients (Claude Desktop, Claude Code, third-party agents) drive Taste through this server. Read-side only by ADR-006.

{
  "mcpServers": {
    "taste": {
      "url": "https://func-taste-dev.azurewebsites.net/api/mcp"
    }
  }
}