git:20260711.c9add96 to git:20260806.bed1b85

92 added, 64 removed. Audit A to A.

---

name: kelly-restaurant-intel

- description: "Kelly Restaurant Intel: App-in-Skill daily industry intelligence cockpit for restaurants, cafes, and F&B groups. Use when the user asks about restaurants, cafes, food and beverage, menus, local events, daily offers,餐饮, or餐厅 scenes. Prepares news/source signals, buyer-intent interpretation, approved sales actions, and channel drafts for review before any external handoff."

+ description: "Kelly Restaurant Intel: App-in-Skill daily industry intelligence cockpit for restaurants, cafes, and F&B groups (Busabase App-in-Skill). Use when the user asks about restaurants, cafes, food and beverage, menus, local events, daily offers,餐饮, or餐厅 scenes. Prepares news/source signals, buyer-intent interpretation, approved sales actions, and channel drafts for review before any external handoff."

---

# Kelly Restaurant Intel

## Overview

Use this skill as Kelly's daily industry-intelligence operator for **restaurants, cafes, and F&B groups**.

- It turns current news sources, trend signals, competitor movement, customer questions, and buyer-intent clues into a small reviewable batch:

+ It turns current news sources, trend signals, competitor movement, customer questions, and buyer-intent clues into a small reviewable batch, held in Busabase:

- source-backed signals;

- why each signal matters to the buyer;

- sales or operating actions for today;

- draft messages/content for staff brief, IG post, delivery blurb;

- blocked claims that need human, legal, compliance, or domain review.

- Default interaction mode: App UI. Unless the user explicitly asks for chat-only handling, check onboarding/config, prepare or refresh the local batch, start/reuse the local app with `app/start.sh`, and give the actual local URL. Use chat-only mode only when the user says "纯聊天", "chat only", "不要打开 UI", or similar.

+ Signal/action/draft collection is a genuine external operation a browser

+ cannot perform (web browsing, source reading, buyer-intent judgment). The

+ AirApp itself only reads and writes Busabase — it never browses, posts

+ anywhere, or performs an external handoff; Kelly writes today's batch

+ directly into Busabase through `busabase-sdk` as part of the normal

+ workflow below, then reviews and decides inside the app.

+ Default interaction mode: App UI. Unless the user explicitly asks for

+ chat-only handling, check onboarding, collect/refresh today's batch, and

+ give the user the clickable AirApp URL (or the local preview URL when local

+ preview is explicitly requested). Use chat-only mode only when the user says

+ "纯聊天", "chat only", "不要打开 UI", or similar.

+ ## Mandatory Dependencies

+ 1. Read and follow `$kelly-app-skill-creator` for product behavior, visual quality, responsive layout, and the complete canonical `app/` artifact.

+ 2. Read and follow `$busabase` for connection, target Space, node discovery, ChangeRequests, review, and merge behavior.

+ 3. Read and follow `$busabase-app-creator` for resource modeling, AirApp runtime limits, security, validation, and deployment.

+ If a dependency is unavailable, preserve this skill's product contracts,

+ stop before the unavailable Busabase operation, and report the exact

+ missing dependency. Do not invent a second data backend.

## Product Package

- **Buyer**: restaurant owners, cafe operators, F&B marketers, and group managers.

- **Pain**: F&B teams need to decide today's offer and content while reacting to weather, events, and nearby competition.

- **Offer**: daily restaurant intelligence that becomes menu pushes, staff notes, and social/offline offer copy.

- **Demo source mix**: local events, weather, competitor menus, review themes, booking demand, and delivery-platform activity.

Sales framing:

> Every morning, AI watches the sources that affect your business, turns them into today's sales actions, and puts the drafts in a review queue before anything becomes official.

Do not lead with "AI platform", "agent workspace", "database", or model names. Lead with the daily business scene.

## Scene Logic

Use this skill to turn local conditions into restaurant group operating and marketing actions. A signal is valuable when it affects reservations, delivery mix, menu focus, staffing, customer recovery, or group-level promotion timing.

Prioritize signals in this order:

1. weather, events, holidays, transport, tourism, and neighborhood activity that change meal-period demand;

2. competitor menus, offers, delivery ranking, and review themes that alter the guest's choice frame;

3. ingredient, supply, staffing, or operating notices that affect what should be promoted safely;

4. recurring customer questions about wait time, allergens, group booking, delivery, or value.

- Actions should become shift briefs, hero-menu picks, delivery copy, review replies, booking scripts, or Buda/Busabase approval cards. Block allergen/food-safety claims unless sourced, price/menu promises without confirmation, and any health or nutrition advice beyond approved copy.

+ Actions should become shift briefs, hero-menu picks, delivery copy, review replies, booking scripts, or Busabase approval cards. Block allergen/food-safety claims unless sourced, price/menu promises without confirmation, and any health or nutrition advice beyond approved copy.

## Boundary

- - The skill may browse public/current sources, reason over buyer intent, draft actions/content, validate schemas, and write local handoff files.

- - The app reads and writes local files only. It must never post content, send WhatsApp/email, mutate CRMs, scrape private systems, spend money, or perform external side effects.

+ - The skill may browse public/current sources, reason over buyer intent, draft actions/content, and write signal/action/draft/source records to Busabase.

+ - The AirApp reads and writes its own Busabase Bases only. It must never post content, send WhatsApp/email, mutate CRMs, scrape private systems, spend money, or perform any other external side effect.

- Customer-visible drafts, regulated claims, pricing promises, medical/financial/legal advice, and outbound messages are approval-required.

- - Store only the minimal source excerpts needed for review. Do not commit `config.local.json`, env files, `app/.data/`, exports, screenshots of private sources, or raw customer data.

- ## First Run And Onboarding

+ - Store only the minimal source excerpts needed for review. Never store crawl credentials or API keys in Busabase.

+ - Block allergen/food-safety claims unless sourced, price/menu promises without confirmation, and any health or nutrition advice beyond approved copy.

- On invocation, check `app/.data/onboarding.json` and private config readiness. If onboarding is absent/incomplete, guide setup before doing real monitoring.

+ ## Busabase Resources

- Ask for non-secret setup details only:

+ Five Bases under one application Folder (`kelly-restaurant-intel`), declared in

+ `app/app/js/config.js` and `app/resource-map.json`:

- - company/brand name, geography, language, and customer segment;

- - 3-10 public source URLs or source categories to monitor;

- - competitor names/URLs;

- - approved offer, CTA, and forbidden claims;

- - preferred channels among staff brief, IG post, delivery blurb;

- - whether Busabase should be the review provider later.

+ - `signals`: source-backed weather, event, transport, tourism, competitor-menu, delivery, booking, and review signals — evidence link, why-it-matters, buyer-intent interpretation, confidence, risk badges, a suggested action, and the human decision verdict.

+ - `actions`: approved/blocked/reviewable shift briefs, hero-menu picks, delivery copy, review replies, booking scripts, or operating actions linked to signals, with owner, priority, reason, next step, and the human decision verdict.

+ - `drafts`: editable channel drafts (staff brief / IG post / delivery blurb) tied to an approved action; a human edit is stored as `edited-body` on the same row, never a separate file.

+ - `sources`: configured local-event/weather/competitor/trend/delivery source categories, freshness, and coverage gaps.

+ - `settings`: one row (`record-id: "batch"`) holding the current batch's metadata (`batch_id`, `generated_at`, `source`, `vertical`, `buyer`, `offer`) as a JSON payload.

- Never ask for API keys or platform tokens in chat. Secrets belong in env files only.

+ Resources provision lazily through an idempotent Busabase ChangeRequest the

+ first time the app runs in a Space; see `references/ui-schema.md` for exact

+ field shapes.

- When setup is complete and the user confirms, write `app/.data/onboarding.json`:

+ ## First Run And Onboarding

- ```json

- {

- "completed": true,

- "completed_at": "ISO timestamp",

- "config_version": "1"

- }

- ```

+ On invocation, check the `signals`, `actions`, and `drafts` Bases. If all

+ are empty, guide setup before collecting a real batch: ask for

+ company/brand name, geography, language, and customer segment; 3-10 public

+ source URLs or source categories to monitor; competitor names/URLs;

+ approved offer, CTA, and forbidden claims; preferred channels among staff

+ brief, IG post, delivery blurb. Never ask for API keys or platform

+ tokens in chat — Busabase connection happens through the app's own Connect

+ Busabase gate.

## Local App

- Start the cockpit with:

- ```bash

- skills/kelly-restaurant-intel/app/start.sh

- ```

- The app uses local HTTP on `127.0.0.1`, preferring port `3000` through `4000`, or `KELLY_RESTAURANT_INTEL_UI_PORT` when set.

+ Default behavior is AirApp-first — give the user the clickable AirApp URL.

+ Start `pnpm --dir app dev` only when local preview/debugging is explicitly

+ requested.

- Required views:

+ Required app views (hash routes):

- - `#/overview`: human-attention panel, today's top signals, ready actions, blocked items, and source coverage.

- - `#/signals` and `#/signals/<id>`: source-backed signals with evidence links, buyer-intent interpretation, confidence, risk badges, and suggested next action.

- - `#/actions` and `#/actions/<id>`: approved/blocked/reviewable operating or sales actions.

- - `#/drafts` and `#/drafts/<id>`: editable staff brief, IG post, delivery blurb drafts with approve/request-changes/block decisions.

+ - `#/overview`: human-attention panel — today's needs-review/approved/blocked counts, top signals, and ready (approved) actions.

+ - `#/signals` and `#/signals/<id>`: source-backed signals with evidence links, buyer-intent interpretation, confidence, risk badges, and a decision panel (Approve / Request changes / Block) with a review note.

+ - `#/actions` and `#/actions/<id>`: approved/blocked/reviewable operating or sales actions with owner, priority, reason, next step, linked signals, and the same decision panel.

+ - `#/drafts` and `#/drafts/<id>`: editable staff brief, IG post, delivery blurb drafts with the decision panel plus a `Save revision` action that writes an edited body directly onto the draft record.

- `#/sources`: configured source categories, freshness, and gaps.

- - `#/settings`: sanitized config summary, onboarding state, provider, language, and accent color.

+ - `#/settings`: sanitized batch metadata, data provider, onboarding state.

Demo mode:

- `?demo=1`, `?demo=overview`, `?demo=signals`, `?demo=actions`, `?demo=drafts`, and `?demo=detail` load deterministic demo data.

- `lang=en` or `lang=zh` forces UI chrome language.

- - Demo API responses never read/write `app/.data/` or private config.

- ## File Contract

- Read `references/ui-schema.md` before changing the app, scripts, or generated JSON.

- - `app/.data/current_batch.json`: current intelligence batch.

- - `app/.data/decisions.json`: user verdicts and edits keyed by item id.

- - `app/.data/agent_tasks.json`: queued agent work for requested changes or missing evidence.

- - `app/.data/execution_report.json`: dry-run/apply handoff report.

- - `app/.data/onboarding.json`: setup marker.

- - `app/.data/agent.lock`: temporary lock while the skill writes files.

- Validate with:

- ```bash

- node skills/kelly-restaurant-intel/scripts/validate_ui_schema.ts skills/kelly-restaurant-intel/app/.data/current_batch.json

- ```

+ - Demo mode never reads or writes Busabase; decisions only mutate the in-memory demo batch.

- ## Normal Workflow

+ ## Batch Workflow

1. Detect mode. Default to App UI.

2. Browse or otherwise collect current public evidence. For news/trends, use exact dates and source URLs.

3. Build one narrow buyer scene, not a generic AI report.

- 4. Write a batch with signals, actions, drafts, and source coverage. Keep every item tied to evidence or mark it blocked.

- 5. Validate the batch.

- 6. Launch the UI for review.

- 7. Poll `agent_tasks.json` for requested changes and revise only those items.

- 8. On "execute/export approved", re-read decisions and run `scripts/execute_decisions.ts` first as a dry run. Apply only after explicit confirmation.

+ 4. Write the batch directly into Busabase through `busabase-sdk` (`bases.createChangeRequest` for each new signal/action/draft/source row, matching the field slugs in `references/ui-schema.md`; refresh the `settings` "batch" row's JSON payload with `batch_id`/`generated_at`/`source`/`vertical`/`buyer`/`offer`). Keep every item tied to evidence or mark it `blocked`.

+ 5. Open the AirApp for review.

+ 6. Poll for `request_changes` decisions (`decision-verdict` = `request_changes` on a signal/action/draft row) and revise only those items — write the revision back onto the same record.

+ 7. On "execute/export approved", run `node scripts/execute_decisions.mjs` first as a dry run. After performing any real handoff, run `node scripts/execute_decisions.mjs --apply` to mark the approved items done.

+ ## Decisions Workflow

+ 1. Kelly's verdicts write directly onto the item record (`status`,

+ `decision-verdict`, `decision-comment`, `decided-at`) through

+ `busabase-sdk`. From a standalone local preview the write merges

+ immediately (trusted operator); from the deployed AirApp it creates a

+ pending ChangeRequest for the trusted process to merge.

+ 2. Before executing anything, run `node scripts/execute_decisions.mjs`

+ (dry-run). It prints the concrete operation for every decided

+ signal/action/draft: `mark_signal_approved`, `export_action_plan`, or

+ `handoff_content_pack` for an `approve` verdict; `queue_agent_revision`

+ for `request_changes`; `mark_blocked` for `block`; `save_human_revision`

+ for a draft's `revise`.

+ 3. After Kelly confirms the dry-run and performs the handoff somewhere

+ outside this script, run `node scripts/execute_decisions.mjs --apply` to

+ mark the approved items `done`. `request_changes`/`block` verdicts are

+ left exactly as decided — there is nothing further for the script to do.

## Safety Defaults

- Treat outbound messages, regulated claims, medical/financial/legal advice, pricing promises, and publishing as approval-required.

- If source evidence is weak, mark the item `blocked` or lower confidence instead of pretending.

- Preserve source language unless the workflow asks for translation.

- - Use Busabase as the later shared review provider when the workflow needs team approvals; local files remain the reference implementation.

+ - Keep every write idempotent by upserting on the item's own id field so re-running a collection pass never duplicates rows.

+ - Block allergen/food-safety claims unless sourced, price/menu promises without confirmation, and any health or nutrition advice beyond approved copy.

+ ## Useful Commands

+ ```bash

+ node skills/kelly-restaurant-intel/scripts/execute_decisions.mjs

+ node skills/kelly-restaurant-intel/scripts/execute_decisions.mjs --apply

+ pnpm --dir skills/kelly-restaurant-intel/app dev

+ ```