tcms-adapter

Content Adapter

Adapts a reviewed core draft (or published article) into the channel-specific versions needed for publishing. Does not create from scratch — that is content-writer's job.

Two modes

Mode A: multi-channel adaptation of new content Input: core draft from content-writer Output: adapted versions for specified channels

Mode B: re-promotion of existing content Input: URL or local file of a published article Output: lightweight material such as social posts, opinion summaries, data-point extractions

Mode B steps:

1. [deterministic] read the published article
2. [LLM] extract 3 shareable information points (hardest data / sharpest opinion / most vivid scenario)
3. [LLM] generate material per target channel
4. [LLM] self-check (data consistent with original / product-name规范 / no sensitive info)
5. [deterministic] output to content/adapted/repurpose/

---

Step 1: Confirm input and target channels [deterministic]

1. Confirm the core-draft path (or published-article URL)
2. Confirm the list of channels to adapt to
3. Read the core-draft content

When no channel is specified, default to the full set: official-account version / community version / Chinese social post ×2 / English X post ×1 / sales one-pager (only when case or hard data is present)

Step 2: Read channel specs [deterministic]

Read the relevant channel-spec files on demand (references/channel-specs/ directory), and read the brand guidelines (reuse content-writer's brand-rules.md).

Step 3: Adaptation generation [LLM]

Each channel version is generated independently, with no dependency between them:

Official-account version

• Rewrite based on the core draft; reduce technical detail, add scenario description
• Scenario hook at the opening, fixed brand closing at the end
• 1500-2500 words

Developer-community version (deterministic fine-tune)

• When the core draft is already in tech-blog style → use the core draft directly
• Only minor formatting tweaks; may append further-reading links at the end
• Not treated as a separate LLM task

Chinese social post

• Extract 1 most impactful data point or scenario
• hook + core message + topic hashtags, under 300 words
• Produce 2 posts from different angles per core draft

English X post

• Thread format (3-5 posts) or single post
• Written with English thinking, not translated from Chinese
• Keep technical terms in original; each post ≤280 characters

LinkedIn summary

• For management audience, leans business value
• 200-300 words, professional formal tone

Sales one-pager

• Generated only when containing customer case or hard data
• Customer challenge → solution → effect data → applicable scenario, ≤400 words

Step 4: Self-check [LLM]

• Word-count limit
• Key data points preserved (no omission, no fabrication)
• Product-name规范
• Customer-name redaction carried through
• English version avoids Chinglish

Step 5: Output [deterministic]

content/adapted/
├── YYYY-MM-{topic}-wechat.md
├── YYYY-MM-{topic}-community.md
├── YYYY-MM-{topic}-social-cn-1.md
├── YYYY-MM-{topic}-social-cn-2.md
├── YYYY-MM-{topic}-x-en.md
├── YYYY-MM-{topic}-linkedin.md     (if needed)
└── YYYY-MM-{topic}-one-pager.md    (if needed)

Execution summary:

## Execution summary
- Core draft: [filename] ([word count] words)
- Generated: official-account ([words]) / Chinese social ×2 / English X ×1 / ...
- Not generated: [channel] (reason)
- Data-point preservation check: [N] all preserved / [M] excluded due to word limit

---

Hard rules

1. Don't fabricate data. Data in adapted versions must come from the core draft.
2. Don't change the core judgment. Simplify yes, change conclusions no.
3. Consistent product names. All versions use the official names in brand-rules.md.
4. Carry through customer redaction. Redacted in the core draft stays redacted in all versions.
5. Don't cross-reference other drafts. Use only information from the current core draft.
6. English quality. Prefer short and accurate over verbose and stiff. Mark uncertain parts [needs confirmation].

Failure handling

⚠️ Human-in-the-loop

• Mode A: official-account and LinkedIn versions go through L1-L2 approval before publishing
• Mode B: social posts go through L1 light review before publishing
• English versions marked [needs confirmation] need human proofreading before publishing

---

中文摘要

Content Adapter 把已审核的核心稿（或已发布文章）适配为各渠道发布版本，不负责从零创作。支持公众号版、开发者社区版、中文社交帖、英文 X 帖、LinkedIn 摘要、销售一页纸，以及存量内容二次推广物料。硬性规则：不编造数据、不改变核心判断、产品名一致、客户脱敏延续、跨稿不互引、英文宁可简短准确。