---
name: smart-canteen-agent-native-work-system
description: Load when the user asks to make smart-canteen products or team workflows "AI Agent 必备能力", "Agent 原生", "公司级 Agent", "让 Codex/Claude 操作业务系统", "人和 Agent 一起工作", or to turn PM/design/delivery work into agent-readable executable evidence.
version: 0.1.0
---

# Smart Canteen Agent-Native Work System

Use this skill to turn a smart-canteen product area, project workflow, or team operation into an Agent-native work system: readable by humans, operable by agents, monitored by responsible people, and reusable by future threads.

## Load Order

1. Read `../../../llm-wiki/agent-native-work-system-for-smart-canteen.md`.
2. If the task affects strategy or business priority, also load `../digital-business-investor-strategy-lens/SKILL.md`.
3. If the task affects product roadmap, also load `../smart-canteen-strategy-agent/SKILL.md`.
4. If the task touches Yunxiao, Codeup, WeDrive, Apifox, SLS, manuals, or videos, route to the corresponding specialist skill after the Agent-native contract is defined.

## Core Contract

For every Agent-native task, write a short contract with:

- target workflow or product surface
- human users and Agent users
- source systems and permissions
- read actions, write actions, and forbidden actions
- evidence needed for success
- human supervision points
- recovery path when Agent execution fails

## Design Rules

### Treat Agent As A Product User

Check whether the target surface exposes stable:

- page names, route names, menu paths, and object labels
- statuses, filters, pagination, exports, and empty states
- API contracts and Apifox/OpenAPI entries
- error messages with safe next actions
- operation logs that show initiator, Agent action, result, and rollback path

Do not call a product Agent-ready when the only interface is a vague visual page or a private chat answer.

### Start With Company-Level Agent

Prefer a shared smart-canteen company Agent before many personal Agents. The first useful company Agent should help with:

- project and product knowledge lookup
- bid-control and competitor evidence retrieval
- Yunxiao task intake and status explanation
- manual, screenshot, video, and review-page evidence packaging
- API, SLS, and Apifox triage routing

Record who maintains the Agent context, who reviews outputs, and where logs/evidence are stored.

### Keep Human Supervision Explicit

Never present Agent automation as fully autonomous in sensitive smart-canteen areas:

- payment, refund, subsidy, account balance, card, or wallet logic
- device control, shipment, meal pickup, coffee machine, vending cabinet, or health hardware
- food-safety supervision, school data, production logs, customer acceptance, or release/deploy actions
- contracts, formal bid documents, external customer promises, finance numbers, or legal claims

For these areas, require a human confirmation point and a rollback or manual recovery note.

### Make PM And Design Outputs Executable

When PM or design work is involved, prefer outputs that an Agent can inspect or replay:

- Markdown source for context and reasoning
- CSV maps for fields, roles, states, screens, APIs, and evidence
- HTML review pages for human inspection
- screenshots or short videos when UI behavior matters
- task-index entries and recovery notes

Avoid stopping at a prose PRD when the next team needs a runnable prototype, route map, API map, or validation checklist.

## Output Shape

Create or update the smallest useful set:

- `work/<date-topic>/summary.md`
- `work/<date-topic>/source-index.csv`
- `work/<date-topic>/review.html` when Jack or teammates need to inspect it
- domain wiki under `standards-stack/llm-wiki/` or `modules/product/`
- project skill only when the behavior will recur
- `control/task-index/items/<task_id>.tsv` for reusable or completed slices

Use Markdown for Agent memory, CSV for new structured tables, and HTML for human review. Keep required project indexes as TSV.

## Stop Rules

Stop or mark blocked when:

- the source page needs login and no readable source is available
- the task would require production credentials, private tokens, or raw customer data
- the workflow includes payment/device/food-safety writes but no safe test environment is available
- the user asks for autonomous production action without human approval
- evidence would expose secrets, personal data, full raw payloads, or customer-confidential documents

## Done Criteria

The work is complete only when:

- the Agent-native contract is written
- source and evidence paths are recorded
- human supervision points are explicit
- reusable outputs are wired into the correct project entrypoints
- validation or blocker evidence is reported