---
name: ai-native-knowledge-loop
description: Load when meeting recordings, requirement discussions, solution reviews, PR reviews, launch issues, or retrospectives need to be stored as reusable AI Native project knowledge and templates.
version: 0.1.0
---

# AI Native Knowledge Loop

Use this skill when the task is to make project discussion and delivery experience reusable by Codex/ChatGPT/Agent, not just to summarize a meeting.

## Source Context

- `standards-stack/llm-wiki/ai-native-project-delivery-collaboration/README.md`
- `standards-stack/prompt/governance/MEETING_AND_AI_COLLABORATION_WORKFLOW.md`
- `work/2026-06-01-ai-native-project-delivery-meeting-intake/summary.md`

## Use For

- 会议录音、转写、方案评审纪要入库.
- PR 评审、上线问题、故障复盘沉淀.
- 创建 AI Native 知识库目录与模板.
- 把一次会议中的可复用规则写回 wiki、skill、QUEUE、HEARTBEAT、task-index.

## Knowledge Buckets

Use these six buckets unless a project-specific rule overrides them:

- `01_需求讨论`
- `02_产品方案`
- `03_研发实现`
- `04_测试验收`
- `05_上线与自愈`
- `06_交付复盘`

## Workflow

1. Preserve source metadata: source path, date, participants when known, source type, SHA256 if local file.
2. Convert source to AI-readable Markdown.
3. Extract:
   - decisions,
   - action items,
   - risks,
   - owners,
   - evidence paths,
   - reusable rules,
   - anti-patterns,
   - open questions.
4. Compile stable knowledge under `standards-stack/llm-wiki/` when it will be reused across projects.
5. Create or update the narrowest skill only when a real future task would fail without it.
6. Produce a human-review HTML page when the output needs team confirmation.

## Stop Rules

- Do not store credentials, raw production payloads, private keys, full SQL dumps, or unmasked personal data.
- Do not treat a meeting action item as a completed task.
- Do not absorb criticized or unverified practices as positive skill rules.
- Do not create broad "AI Native everything" skills when a wiki page or checklist is enough.