Content Planner

Generates structured topic briefs from multiple signal sources. Does no content creation — only topic judgment and brief output.

Operating mode: routing

Routes to different topic-selection logic by signal source:

signal input
   ↓
[routing decision]
   ├── knowledge base has new/updated entries → new-content topic
   ├── competitor has a major move            → differentiation-response topic
   ├── content calendar due                   → schedule-driven topic
   └── user explicitly requests               → topic by specified direction

Workflow

Step 1: [deterministic] Confirm the topic trigger

A. User explicitly requests → user specified direction/type → go straight to Step 3

B. Periodic check → read the following signal sources:

1. Knowledge-base recent updates — check file modification times, identify recently updated product sections
2. Content calendar — read the current month's schedule, find content due this/next week
3. Published-content index — check which topics already have drafts and which are still missing

The paths above must be configured per the actual project directory.

Step 2: [deterministic] Read context

• Must read: the content calendar for the current month
• Must read: product-line / product-center mapping (configure per actual org structure)
• As needed: the knowledge-base section for the relevant product (use search_content to locate, don't read in full)
• As needed: existing-content inventory (judge whether there is stock content worth re-promoting)

Step 3: [LLM] Generate topic suggestions

Based on signal sources and context, generate 1-3 topic suggestions. Each outputs a structured brief:

## Topic Brief

**Topic title**: [working title, not the final published title]
**Article type**: tech blog / customer case / product update
**Target product**: [product name, use the official site name]
**Target reader**: [specific persona]
**Core message**: [one sentence summarizing the core message]
**Purpose**: [what the article should achieve]
**Material guide**:
  - knowledge-base entry: [which section to point to]
  - internal case: [📋 if any, mark customer name + redaction requirement]
  - published article: [if any reference exists]
**Target channel**: [first channel + syndication channel]
**Approval level**: L1 / L2 / L3
**Priority**: high / medium / low
**Suggested publish time**: [specific date or time window]

Step 4: [LLM] Topic sanity check

• Is the product within the current priority scope?
• Does it duplicate content published/scheduled in the last 2 weeks?
• Is material sufficient (a corresponding knowledge-base entry exists)?
• Is the channel reasonable?

Step 5: [deterministic] Output

Save the brief to content-calendar/briefs/YYYY-MM-DD-brief.md

Output an execution summary:

## Execution summary
- Trigger source: [user request / calendar schedule / knowledge-base update]
- Read: [list files read]
- Topics produced: [N]
- Requires human confirmation before entering content-writer

⚠️ Human-in-the-loop

After a topic brief is produced it must be confirmed by a human before entering content-writer. Do not auto-trigger downstream skills.

Hard rules

1. No content creation. Output only the topic brief; don't write article body.
2. No auto-trigger of downstream. The brief must be human-confirmed.
3. Don't recommend when material is insufficient. If the corresponding knowledge-base entry is too thin, mark "insufficient material, suggest supplementing the knowledge base first".
4. Internal cases default to redaction requirement. Briefs involving internal-source customers must mark 📋.

Failure handling

中文摘要

Content Planner 是基于多信号源（知识库更新、竞品动作、内容日历、效果数据）生成结构化选题 Brief 的 Agent，只做选题判断与 Brief 输出、不写正文。产出后须经人工确认才能进入 content-writer，不自动触发下游。硬性规则：素材不足不推荐、内部客户案例默认标注脱敏要求。