AI Learning Loop

Run a compact learning workflow that turns a topic or material into a concrete output and a small number of reusable knowledge cards. Optimize for feedback density, not note volume.

Core Loop

Use this sequence:

learning goal
  -> real question
  -> minimal input or research pack
  -> expression card
  -> output review
  -> one necessary card/sink decision
  -> next loop question

Prefer one loop at a time. Do not turn the task into a broad research report or a large note-taking project unless the user explicitly asks.

Workflow

1. Clarify the learning target

Identify the user's intended output before collecting or summarizing information.

If the target is unclear, ask one concise question. Otherwise infer a practical target from context, such as a口播稿、公众号段落、课程小节、playbook、咨询框架, or decision memo.

Read references/question-planning.md when the request starts from a broad topic, has no material, or needs a minimal learning path.

2. Decide the input path

Choose one:

For no-material requests, do not gather many links. Produce a smallest useful pack: 3 must-read resources, up to 5 optional resources, skip list, reading order, and first output task.

3. Build an expression card

Create one publishable or near-publishable expression card around one core judgment. Read references/expression-card.md when drafting a口播稿、短文、课程小节、or other output card.

Treat口播稿 as an expression card:

raw/studio/funnel/formatted/shipin/ = pre-publish expression card
raw/studio/funnel/transcripts/ = post-publish asset card

If working inside kj-llm-wiki, follow that repo's raw/wiki rules. Do not write to raw/ unless the user explicitly asks for a production draft and the repo rules allow that path.

4. Review before expanding

Review the output before polishing or expanding it. Read references/review-and-sink.md when judging gaps, quality, or wiki sink decisions.

The review must answer:

• Is the core judgment clear?
• Is there a real reader pain or use case?
• Which parts are generic or obvious?
• Where is the logic jump?
• What example, counterexample, boundary, or source is missing?
• What is the single most necessary card to add next?

5. Sink only the necessary card

Recommend at most one mandatory sink item by default:

If the user says "沉淀到文档/知识库", invoke the existing wiki-doc-sink workflow rather than reimplementing wiki write rules.

Output Contract

For one loop, return:

1. 今日真实问题
2. 最小输入/调研包
3. 表达卡片
4. 评审反馈
5. 只补哪一张卡
6. 下一轮问题

Keep the answer concise unless the user asks for a full draft or asks to write files.

交付前自检

• 对照 Output Contract 逐项核对回复包含全部 6 项：今日真实问题、最小输入/调研包、表达卡片、评审反馈、只补哪一张卡、下一轮问题
• 无材料请求时，清点调研包数量达标：必读资源恰好 3 个、可选资源不超过 5 个，且含 skip list、阅读顺序和首个输出任务
• 检查表达卡片只围绕一个核心判断展开，评审反馈逐条回答了第 4 步列出的 6 个问题
• 确认沉淀建议默认不超过一张必要卡片，且沉淀目标与 Gap 对照表（wiki/concepts/、wiki/playbooks/ 等）一致
• 在 kj-llm-wiki 中执行时，确认未在用户未明确要求的情况下写入 raw/ 目录