cmm-content-writing

name	cmm-content-writing
description	Use this CMM skill to turn a content_plan into platform-specific draft content packages for WeChat, X, Weibo, Xiaohongshu, blog, newsletter, or script-style outputs. It directly creates CMM draft packages using internalized research, blog, social, and script writing patterns while defaulting to local draft-only output. This new CMM layer does not modify legacy writing skills and never publishes, schedules, or writes to external systems unless the user explicitly asks for a later publishing workflow.
version	0.1.2

CMM Content Writing

cmm-content-writing is the writing layer in the CMM content factory.

It consumes a content_plan from cmm-content-planning and produces a content_package for downstream visual, video, and publishing workflows.

Boundary

Use this skill for:

• Turning a content_plan into a draft article, post, thread, note, or script.
• Choosing the right writing pattern for the target platform.
• Producing structured content_package output.
• Repurposing one plan into multiple platform drafts when requested.
• Preserving source discipline, claim caution, and platform-specific structure.

Do not use this skill for:

• Selecting topics. Use cmm-content-planning only after topic selection.
• Generating images or video.
• Publishing, scheduling, posting, or writing to Notion.

Legacy Boundary

Do not invoke or modify legacy writing skills by default. Their useful patterns are internalized here as CMM writing methods. If a legacy skill exposes a useful edge case, record it as legacy_gap and update this CMM skill, its references, or its contract.

Input

Preferred input is a content_plan:

content_plan:
  topic: ""
  audience: ""
  platform: ""
  format: ""
  core_angle: ""
  core_claim: ""
  evidence: []
  risks: []
  expression_strategy: ""
  title_cover_direction: ""
  outline_direction: []
  first_action: ""
  writing_decision:
    user_real_need: ""
    reader_takeaway: ""
    article_type: ""
    target_length: ""
    length_reason: ""
    one_sentence_spine: ""
    must_include: []
    must_exclude: []
    scope_boundary: ""
  drafting_instruction:
    title_promise: ""
    body_must_center_on: ""
    article_shape: ""
    personal_style_role:
      source_profile: ""
      voice: ""
      rhythm: ""
      stance: ""
      lexicon: []
      avoid: []
    must_include: []
    forbidden_drift: []
    length_strategy: ""
    source_boundaries: ""
  style_selection:
    user_profile: ""
    writing: ""
    visual: ""
    layout: ""
  downstream_needs:
    writing: true
    visual: true
    video: false
    publishing: false
  status: draft

If the user only provides a topic, ask to run planning first or build a minimal content plan before drafting.

Read references/content-package-contract.md for exact output fields. Read references/writing-method.md for platform structure, evidence handling, and repurposing rules. Read references/wechat-writing-style.md for the full WeChat long-form quality standard when platform: wechat. Read references/xhs-copy-spec.md for Xiaohongshu/Rednote note structure, card copy, caption, hashtags, and preflight rules when platform: xiaohongshu. For Xiaohongshu titles and cover hooks, also read references/xhs-title-templates.md.

Before drafting, resolve content_plan.style_selection:

• Writing template: cmm-content-factory/references/styles/writing/<id>.yaml.
• User profile: ~/.cmm/profile/<name>.yaml when set; if absent, use the shipped persona-neutral references/styles/default-profile.yaml (then bare neutral only if that is missing) per cmm-content-factory/references/styles/style-system.md.
• The generic anti-AI gate in style-system.md always applies; the user profile may add personal bans on top.

Template rules and user-profile persona/audience/lexicon are mandatory writing constraints, not suggestions. A user instruction in the current turn overrides them for this task only.

No styleless drafting gate: content_plan.style_selection.user_profile, writing, visual, and layout must be resolved before drafting any article. Profiles are runtime configuration. For this local CMM workflow, ~/.cmm/profile/neo.yaml is the local default only when it exists and the user did not request another style. Do not hardcode Neo-specific writing traits into the shared skill; derive drafting_instruction.personal_style_role from the active user profile for each run. If user_profile or writing is empty, stop and return a planning/style-selection block instead of producing正文. Do not write in the model's generic voice and then retrofit style afterward.

Profile-to-constraints gate: Loading a profile is not enough. Before any article正文 is drafted, convert the active user profile, account template, writing template, and this task's personal_style_brief into drafting_instruction.pre_draft_style_constraints. This object is the writing constraint package, not a review note. It must name the opening move, first-300- character judgment, article backbone, section intent, paragraph rhythm, allowed first-person stance, required judgment signals, banned generic shapes, and rewrite triggers for this exact article. If it is missing, generic, or not connected to the selected title promise, block drafting. A later voice_match_evidence.passed: true is invalid unless it proves the draft followed these pre-draft constraints with excerpts or concrete reasons.

Routing

Choose the writing pattern by platform, format, and evidence needs.

Need	Pattern
WeChat long-form analysis, cited article, tutorial, case study	Research article pattern
Personal blog or opinion essay	Blog draft pattern
X thread, Weibo post, Xiaohongshu copy, LinkedIn-style post	Social draft pattern
口播稿 / 短视频口播脚本（真人出镜要念的稿，format=talking_script）	口播稿 pattern → references/talking-script-method.md（钩子先行批量选 + 时长卡字数 + 口播/分镜双轨 + 人设口吻）
其它短视频脚本（纯文本/旁白）	Script draft pattern
Multi-platform adaptation	Repurpose pattern

Writing scope:

• All writing methods are internalized here: evidence-first structure, voice-aware longform, platform adaptation, thread compression, and social copy formatting. No external writing skill is required.
• Any external save, Notion write, schedule, or publish action belongs outside this skill.

Workflow

1. Validate the content_plan.

2. Resolve content_plan.style_selection: load the writing template and the user profile (or neutral mode).

3. Confirm the style gate passed:

   • user_profile_loaded: true
   • writing_template_loaded: true
   • style_selection.user_profile is non-empty
   • style_selection.writing is non-empty If any item fails, block drafting and ask/route back for style selection.

4. For WeChat long-form drafts, run the account-template context gate before generating title or正文. Read references/wechat-writing-style.md, references/platform-copy-specs.md, references/title-spec.md, and, for AI tool / product update / workflow topics, references/wechat-title-templates.md; then read the selected user profile and the resolved WeChat account template. If the active profile declares account_templates.wechat, expand and read that template file; otherwise read the bundled persona-neutral fallback cmm-content-factory/references/templates/wechat-default-template.md. Write wechat_context_gate before drafting:

   wechat_context_gate:
     platform_rules_loaded: true
     user_profile_loaded: true
     account_template_path: ""
     account_template_loaded: true
     article_type_template_loaded: true
     calibration_style_checked: true
     passed: true
     source_files: []
   

   If the resolved account template was not read, block drafting. The account template must be converted into concrete constraints in voice_contract: opening pattern, article structure, paragraph style, practicality pattern, personal judgment pattern, and closing pattern. Do not satisfy this gate by merely naming the profile. Treat the account template as the primary visible article shape. Platform, length, density, and quality gates are backend constraints; they must not cause the draft to grow extra tables, prompt packs, check-standard blocks, or process artifacts that the account template did not call for.

5. Check the writing decision gate for article drafts. content_plan should contain writing_decision.user_real_need, reader_takeaway, article_type, density_level, target_length, one_sentence_spine, must_include, must_exclude, and scope_boundary. If it is missing but the request is clear, build the decision before drafting and record it in the output. If the one-sentence spine is vague, reader takeaway is unclear, or target length is unsupported by the topic payload, stop and route back to planning instead of padding the draft. For WeChat, choose one canonical article type and target length from references/wechat-writing-style.md before drafting: quick_opinion 800-1200, opinion_essay 1200-1800, practical_method 1200-1800, tutorial 1800-2600, workflow_template 1800-2600, trend_explainer 1800-2600, case_study 2200-3200, tool_list_comparison 2200-3200. For product updates, tool introductions, feature explainers, AI workflow updates, and ordinary opinion/news-analysis pieces, prefer 1800-2400 chars unless the article has a concrete long-form deliverable. Any draft above 3200 chars must record a long-form qualification in length_reason; any draft over 3800 chars is a long_special_report exception and requires explicit user confirmation. Also choose one density_level from references/wechat-writing-style.md: compact, standard, or expanded. Default from tight: product updates, feature explainers, short tool introductions, and one-point opinions use compact; tutorials and workflow explanations use standard; real testing, case studies, multi-option comparisons, reusable templates/checklists, and source-heavy analysis may use expanded.

6. Build or validate personal_style_brief before drafting. It must come from the selected user profile plus any loaded account template, and it must shape this exact article rather than restating the profile generically:

   personal_style_brief:
     source_profile: ""
     account_template: ""
     opening_strategy: ""
     core_judgment: ""
     practitioner_perspective:
       stance_type: ""        # observer_judgment | hands_on_experience | workflow_practitioner | user_advisor
       usable_scene: ""
       author_would_do: ""
       likely_misuse: ""
       evidence_boundary: ""
     structure_strategy: ""
     selection_logic: ""
     paragraph_rhythm: ""
     must_sound_like: []
     forbidden_shape: []
     generic_template_risk: ""
     article_type_override: ""
   

   Treat this brief as the writing engine below the selected title's promise. It must be derived from the active user profile for this task, not from any hardcoded creator voice inside the shared skill. For AI tools, product updates, workflow articles, tutorials, and practical tech explainers, practitioner_perspective is required: state whether the author is writing from observation, hands-on use, workflow practice, or user-advisor judgment; name one realistic use scene; say how the author would or would not use it; name one likely misuse; and state the evidence boundary. Do not fake hands-on experience. If there is no real test material, use observer_judgment or workflow_practitioner, not hands_on_experience. If the generic article type would produce a manual, average listicle, or equal-weight tool explanation that conflicts with the user's profile, use article_type_override to preserve the personal article backbone, but do not replace or hide the title's promised subject. Example: a title with "10 tools" may still be practical_method + curated_tool_path, but the 10 tools must remain visible article subjects; the method only explains the selection and order. If personal_style_brief is missing and a user profile is available, stop and build it before writing.

7. Convert the loaded profile and personal_style_brief into drafting_instruction.pre_draft_style_constraints before drafting. This object must be concrete enough to generate from, not merely to evaluate after the fact:

   pre_draft_style_constraints:
     source_profile: ""
     source_account_template: ""
     opening_move: ""
     first_300_chars_must_do: []
     core_judgment_to_state: ""
     practitioner_perspective:
       stance_type: ""
       usable_scene: ""
       author_would_do: ""
       author_would_not_do: ""
       likely_misuse: ""
       evidence_boundary: ""
     article_backbone: ""
     section_intent:
       - section: ""
         job: ""
         style_requirement: ""
     paragraph_rhythm: ""
     first_person_stance: ""
     required_judgment_signals: []
     required_reader_guidance: []
     allowed_lexicon: []
     banned_shapes: []
     rewrite_triggers: []
   

   Block drafting if:

   • core_judgment_to_state is vague or only repeats the topic;
   • AI tool/product/workflow articles lack practitioner_perspective;
   • practitioner_perspective.stance_type claims hands-on use without real user-provided or source-backed experience;
   • opening_move would produce a product-doc or dictionary opening;
   • article_backbone is a generic feature list when the profile calls for a judgment-led article;
   • section_intent does not show how each section advances the profile-shaped judgment;
   • banned_shapes and rewrite_triggers are missing.

8. Choose one primary platform and format.

9. Lock the platform copy budget before drafting. Read references/platform-copy-specs.md and references/title-spec.md, then record the budget in content_package.platform_copy_budget. For WeChat, title_cmm_hard_max: 25 is a hard production standard; the official draft API's 32-character limit is recorded only as title_official_api_max and must not be used to loosen title generation. The upstream content_plan should already contain a title_direction_gate; inherit its selected direction as the title promise, but do not treat it as the final publish title. For WeChat AI tool / product update / workflow topics, use references/wechat-title-templates.md to avoid flat feature-log titles while preserving subject and evidence boundaries. After正文 is drafted, run the final_title_gate: generate 3-5 final title candidates under the CMM hard max, compute the length with a tool, discard over-budget candidates before recommending, and write the result to title_quality_gate. Candidate pattern values must be one of the six canonical patterns in references/title-spec.md; do not invent ad-hoc labels such as "tool + action", "object + result", or "suggestion". Put those explanations in reason instead. Do not rely on model self-counting.

10. Check evidence:

• If the plan has sources, use them and preserve source references.
• If evidence is weak or uninspected, keep claims cautious and add fact_check_status: needs_review.
• Deep backfill: if the outline exposes a claim that needs a number, date, case, or quote not present in the plan, call cmm-web-search in fact_check intent for that specific claim before drafting it. Planning owns breadth research; writing owns claim-level backfill.

11. Before drafting, load or build drafting_instruction from the selected title, outline, writing decision, personal_style_brief, and pre_draft_style_constraints. This package is the generation instruction, not a post-draft rubric. Do not draft until it clearly states the title promise, what the正文 must center on, article shape, runtime-derived personal-style role, must-include items, forbidden drift, length strategy, and source boundaries. It must block drafting when:

• the title promises a counted list but the article shape hides the counted objects;
• personal_style_brief changes the article from the title's subject to a different framework;
• pre_draft_style_constraints are missing or generic;
• the one-sentence spine no longer answers the title promise.

12. Draft the content in the requested platform shape, using drafting_instruction as the primary input. personal_style_brief, writing templates, platform rules, and quality gates are constraints inside that instruction package, not independent engines competing after the draft exists. Enforce the writing decision while drafting: fulfill the one-sentence spine, keep only must_include content, do not expand must_exclude adjacent topics, and keep length inside target_length unless a clear source-backed reason is recorded in review notes. Do not write a generic tool manual and then sprinkle personal phrases on top; also do not over-correct by turning an explicit list/recommendation title into a hidden workflow essay.
13. Build and enforce a voice_contract for the selected style. The contract must turn the loaded writing template + user profile into concrete checks: opening pattern, viewpoint/persona, reader stance, sentence rhythm, required/allowed lexicon, forbidden tones, structure signature, and closing pattern. For WeChat, it must also include any loaded account template as a source file and reflect that template's article shape and practicality rules. It must also include the personal_style_brief fields: core judgment, practitioner perspective, structure strategy, selection logic, forbidden shape, and generic-template risk, plus the pre_draft_style_constraints fields. This rule is universal: it applies to Neo and to any future profile or style. Do not rely on style labels alone.
14. Run voice_match_evidence before output. For each important item in the voice_contract, cite a short excerpt or state the concrete reason it passes. If the draft only uses surface vocabulary but misses structure, viewpoint, or rhythm, mark it failed and rewrite before output. For profile- driven articles, evidence must explicitly check:

• personal_judgment: whether the piece has a clear author judgment;
• first_hand_logic: whether the article explains why the author orders or selects things this way;
• practitioner_perspective: whether the article includes a realistic use scene, what the author would or would not do, a likely misuse, and an honest evidence boundary;
• anti_generic_shape: whether it avoids manual/listicle/explainer drift;
• reader_path: whether the reader gets a usable path, not just item descriptions;
• compression: whether non-core explanations were compressed or removed. Evidence must also prove that pre_draft_style_constraints shaped the opening, first-300-character judgment, backbone, section intent, paragraph rhythm, and rewrite-trigger cleanup. If any rewrite_triggers are present in the draft, voice_match_evidence.passed must be false and the draft must be rewritten before downstream stages.

15. Run the generic anti-AI gate (plus user-profile additions) on the draft and rewrite failing passages.
16. Run the WeChat opening quality gate when platform: wechat and the draft is long-form. The default opening is 80-150 Chinese characters; 150-180 requires a useful context reason; over 180 requires a recorded reason. The opening must state the reader problem, relevance, and article promise without padding. Write opening_quality_gate.
17. Run the WeChat paragraph density gate when platform: wechat and the draft is long-form. Personal spoken cadence does not override platform shape: tutorial/practical articles default to 2-4 sentences per natural paragraph; one-sentence paragraphs are reserved for emphasis, transition, key judgments, section openings/endings, or closing cadence. Count body paragraphs excluding headings, lists, quotes, code blocks, and examples, then write paragraph_density_gate. If the rhythm looks like a talking script, memo, or line-broken note, merge/rewrite before output.
18. Run the WeChat article length gate when platform: wechat and the draft is long-form. Count final正文 with a tool and write article_length_gate. If the draft exceeds the target upper bound by more than 10%, compress it or ask the user to confirm the longer draft. If it is below the lower bound by more than 15%, check whether it under-serves the promised reader takeaway.
19. Run the unified WeChat content quality gate when platform: wechat and the draft is long-form. Write content_quality_gate with concrete evidence for: reader takeaway, one-sentence spine, information density, practical value, structure coherence, evidence support, scope control, no padding, and conclusion payoff. Also check closing density: for compact product updates, tool introductions, feature explainers, and short opinion articles, delete ceremonial sign-offs when the previous sentence already lands the useful judgment. For practical/tutorial/workflow articles, also prove there is an actionable deliverable (step-by-step workflow, template, checklist, reusable prompt, example output, or quality rubric) and that the article is not a principle-only/theory-first explainer. The deliverable may be an article-native method, judgment path, decision sequence, or editing flow; do not force a visible table, prompt block, or checklist just to satisfy this gate. If any item fails, rewrite before output.
20. Run the WeChat information-density compression gate when platform: wechat and the draft is long-form. This gate is separate from article_length_gate: a draft can be inside the length band and still be too loose. Use the selected density_level as the expansion rule. For every draft, at minimum check repeated judgments, generic background or empty significance, and mergeable scenarios/limitations/closing paragraphs. For product updates, tool introductions, feature explainers, AI workflow updates, and timely practical articles, prefer a sharp 2200-2400 Chinese-character draft. If the final正文 exceeds 2500 characters, explicitly check repeated judgments, loose explanations, mergeable scenario sections, over-explained limitations, and padded closing paragraphs. Compress before output unless the extra length is justified by hands-on testing, a detailed tutorial, a comparison table, a reusable template/checklist, or multi-source analysis. Write information_density_compression_gate with the compression decision and evidence.
21. Add review notes:

• source gaps
• claims needing verification
• visual/video handoff needs
• platform constraints
• deviations from the writing decision, especially length or scope changes

22. Add platform metadata, downstream needs, writing_decision, voice_contract, voice_match_evidence, and quality_gate.voice_match. voice_match may be passed only if evidence is present and failures are empty.
23. Output content_package.
24. Stop before visual generation, video rendering, scheduling, or publishing.

Platform Draft Rules

WeChat

• Use body_markdown.
• Prefer clear sections, evidence, and one central claim.
• Keep source references in sources_used.
• Add visual_needs for cover and body illustrations when useful.
• Fill cover_copy: main_title is required — it is the article title verbatim, do not shorten or reword it (the cover renders it across two lines if long); emphasis, subtitle, tags are optional (leave empty when not needed — the cover template enlarges the title and adds whitespace). subject_theme is just a one-line concept hint; the right-side illustration elements are left to the image model (and it renders no text there). Do not let the image stage invent cover copy — decide it here.
• Article title and cover_copy.main_title: follow references/title-spec.md — generate across 2-3 patterns, screen by the 5 铁律 + length red lines (platform-copy-specs.md: WeChat title ≤25; cover_copy.main_title = the article title verbatim) + the 3-tier exaggeration 尺度, then pick by judgment. If the article came from a WeChat title_direction_gate, keep the confirmed direction's promise but rework the final title from the finished正文. Offer the user 3-5 candidates with their canonical 6-pattern label, computed length, and angle reason; never ship a number in the title unless it is verified, unambiguous, and non-misleading.
• Fill illustration_briefs for body images that carry article substance (steps, comparisons, frameworks, key-point summaries): each gives a title (required) and points (word-level items) — this copy is content and belongs here, not invented at image time. Style is not chosen here; it follows the account visual style registry (default + overridable). Default to information-bearing images, not decorative ones.
• Avoid turning uninspected search snippets into factual paragraphs.

X

平台限制：

项目	免费账号	Premium 账号
正文字数	280 字符	25,000 字符
图片数量	最多 4 张	最多 4 张
图片格式	JPG / PNG / WebP / GIF	同左
图片大小	单张 ≤ 5MB（GIF ≤ 15MB）	同左
推荐图片尺寸	1200×675（16:9）或 1200×1200（1:1）	同左
链接计字	每条链接固定计 23 字符（t.co 缩短）	同左

内容规则：

• 默认写单条帖子，不做 thread（多条连发）
• 默认中文写作，除非用户明确要求英文
• 从长文改写时：提炼核心观点为一条帖子，配 1 张带文字的 16:9 封面图
• 280 字符限制下写中文更充裕（一个汉字 = 1 字符），适合信息密度高的表达
• 链接放在正文末尾，计入 23 字符
• 配图需求写入 visual_needs，指定 16:9 封面 + 中文标题文字

Weibo

• Use shorter commentary or article-summary style.
• Keep title and summary length constraints in publish_notes when relevant.

Xiaohongshu

• Follow references/xhs-copy-spec.md as the platform-specific copy contract.
• Choose one note type before drafting: tutorial, checklist, pitfall, tool_review, case_breakdown, or opinion_note.
• Produce separated platform_title, cover_hook, card page copy, caption, hashtags, and visual_needs.
• Keep the publish title display-safe around 20 Chinese characters and keep the prepared body under the Xiaohongshu margin in platform-copy-specs.md / cmm-publishing (<= 950 chars before the trailing hashtag line when preparing a browser publish package).
• Write for a 3:4 card set: one idea per card, default 1 cover + 4-8 content cards, with at least one save-worthy artifact such as a checklist, comparison, prompt, template, step list, or decision rule.
• Keep copy scannable and save-worthy; do not write dense article paragraphs.
• Do not stack emoji, fake urgency, unsupported rankings/numbers, vague attribution, or slogan endings.
• Add visual_needs with type: xhs_card_set; do not leave card text for the image model to invent.

Blog

• Produce a local draft only.
• Do not call or require Notion publication.
• Mark publish_notes with "external publication not requested".

Script

• Write spoken-language script or short-video draft only.
• Add video_needs for later video production.
• Separate hook, main beats, ending, and caption text.

Quality Gate

Before output, check:

Gate	Requirement
claim_integrity	Major claims are supported or marked for review.
platform_fit	Draft shape matches the platform and format.
need_fit	Draft solves the user's real need, gives the stated reader takeaway, honors the one-sentence spine, and stays within the justified length/scope.
drafting_instruction_followed	Light sanity check that the draft follows the pre-draft instruction package; do not add new creative requirements here.
voice_cleanliness	Avoid generic AI filler, empty slogans, and unsupported certainty.
voice_match	Draft follows the style profile's persona, tone, structure rules, and lexicon; must be proven by voice_match_evidence, not asserted.
anti_ai_gate	Draft passes the style profile's anti-AI rules; failing passages rewritten.
structure	Draft has a clear opening, body, and close for its format.
handoff	Visual/video/publishing needs are explicit enough for downstream skills.

Output

Always output a content_package:

content_package:
  plan_id: ""
  platform: ""
  format: ""
  title: ""
  body_markdown: ""
  summary: ""
  variants: []
  sources_used: []
  visual_needs: []
  video_needs: []
  publish_notes: ""
  fact_check_status: pending
  review_notes: []
  writing_decision:
    user_real_need: ""
    reader_takeaway: ""
    article_type: ""
    target_length: ""
    length_reason: ""
    one_sentence_spine: ""
    must_include: []
    must_exclude: []
    scope_boundary: ""
  drafting_instruction:
    title_promise: ""
    body_must_center_on: ""
    article_shape: ""
    personal_style_role:
      source_profile: ""
      voice: ""
      rhythm: ""
      stance: ""
      lexicon: []
      avoid: []
    must_include: []
    forbidden_drift: []
    length_strategy: ""
    source_boundaries: ""
  wechat_context_gate:
    platform_rules_loaded: false
    user_profile_loaded: false
    account_template_path: ""
    account_template_loaded: false
    article_type_template_loaded: false
    calibration_style_checked: false
    passed: false
    source_files: []
  style_selection: {}
  voice_contract:
    profile: ""
    mode: ""
    must_have: []
    must_not: []
  voice_match_evidence:
    checked: false
    passed: false
    evidence: []
    failures: []
  quality_gate:
    claim_integrity: ""
    platform_fit: ""
    need_fit: ""
    drafting_instruction_followed: ""
    voice_cleanliness: ""
    voice_match: ""
    anti_ai_gate: ""
    structure: ""
    handoff: ""
  status: draft

Stop Conditions

Stop and ask before:

• Posting or scheduling anywhere.
• Writing to Notion, Feishu, Google Docs, or external systems.
• Calling paid APIs.
• Generating images or rendering videos.