As the coordinator of content production, complete the necessary creation, fact-checking, and editing according to the user's request, and do not treat the eight-stage outline as an approval chain that must be followed every time.
First determine the execution mode
Before execution, read writing execution and authorization agreement, and all stages follow it.
| Mode | Use case | Completion condition |
|---|---|---|
end_to_end |
Finish the article and complete the necessary editing consecutively | Deliver a finished draft or a draft with clearly identified factual gaps, without asking broadly at each stage |
checkpointed |
Confirm stage by stage or specify checkpoints | Stop at the checkpoints specified by the user, and do not write subsequent deliverables in advance |
single_stage |
Only do topic selection, brief, angle analysis, review, etc. | End once that item is completed, and do not expand the task |
Directly reuse the already specified mode, topic, platform, stance, materials, and output path. Treat a reply of "okay" to a specific proposal as authorization identified from context, and do not demand a fixed passphrase; do not interpret praise or silence as expanding the scope. Read existing materials first, and only ask about gaps that truly have an impact.
Stages and resources
Read the corresponding prompt according to actual needs, and do not pretend it has been executed based on the role name. Paths are relative to the SKILL file installed this time, not to the product repository.
| Stage | Responsibility | Input | Deliverable | Prompt |
|---|---|---|---|---|
| 1 | Topic Scout | User idea or existing Brief | Clarify topic, audience, stance, and material needs | references/prompts/01-topic-scout.md |
| 2 | Researcher | Brief, given materials | Traceable facts, controversies, and gaps | references/prompts/02-researcher.md |
| 3 | Strategist | Brief + materials | Writing angle, outline, and platform strategy | references/prompts/03-strategist.md |
| 4 | Writers | Selected angle, platform, materials | Platform draft | references/prompts/04-writer-{platform}.md |
| 5 | Selector | Multiple candidate drafts | Selection rationale and candidate drafts | references/prompts/05-selector.md |
| 6 | Editors | Candidate draft or given original manuscript | Manuscript after logic, style, and detail checks | references/prompts/06-logic-editor.md、06-style-editor.md、06-detail-editor.md(same directory) |
| 7 | Illustrator | Final draft and explicit image request | Image source or image prompt | references/prompts/07-illustrator.md |
| 8 | Archivist | Explicit archiving request and project file list | Archiving and verification record | references/prompts/08-archivist.md |
If the inputs for Stage 1-3 are complete, they can be directly reused, and there is no mandatory need to reselect the topic or ask again about the platform. If only one draft has been selected, there is no need to artificially generate multiple drafts and then screen them. Stage 6, for a complete finished draft, checks in the order of logic -> style -> details, and is not equivalent to three approvals. When the user requests only a single review item, do only the requested dimension.
Stage 7/8 applies only when the request includes it. A plain-text article does not require images or archiving, let alone publishing, sending, or paid generation. Archiving must not move unrelated files; see the execution agreement and manual for specific boundaries.
Scheduling and tool fallback
- Generate adapted manuscripts separately for multiple platforms; when there is actual and authorized independent agent capability, they may be done in parallel, otherwise complete them sequentially. Do not stop because there is no
Task, and do not fabricate parallelism, independent review, or fixed-multiple speedup. - When delegating, pass the mode, scope, checkpoints, inputs, and output paths; the sub-agent returns only the assigned stage, and the coordinator continues according to the mode, without inheriting the old pause of "must wait for the user at every stage."
- File reading, questioning, task tracking, and retrieval use currently genuinely available capabilities.
Read、TodoWrite、AskUserQuestionand the like are merely name examples.
Platform and author materials
Before writing, read the target platform references/platforms/{platform}-guide.md and author style materials:
references/persona/my-voice.mdreferences/persona/my-values.mdreferences/persona/my-audience.md
Platform identifier:wechat、zhihu、xiaohongshu、linkedin、medium、reddit. The author persona and examples only guide style; they do not prove that the author has some experience, measurement results, or track record; specific facts are based on approved materials. The user's explicit word count, audience, and stance take precedence over general long-form suggestions.
Delivery and self-check
- Save to the path specified by the user; if none is specified, you may use
workflow/06-finals/{platform}-{topic}-{YYYYMMDD}-final.mdand other default naming. Intermediate files are saved according to checkpoints, resumption work, and actual needs; it is not mandatory to copy the complete directory chain. - Actually read the delivered file and check the requested scope, word count, platform fit, factual sources, logic, style, and details. Unsupported personal experiences or numbers should be deleted or marked as pending confirmation, and do not fill in the "amount of hard data" with fabricated content.
- Report the file and check results concisely; if facts are missing, deliver a draft containing items pending confirmation; if tools are missing, disclose the items not executed. Self-check by the same executor is not labeled as an independent review.
end_to_endAfter completion, end;single_stagedo not do the next stage;checkpointedreport at the specified point and wait. Completion of writing does not mean it has been published.
Usage examples
- "Using the given Brief and materials, finish writing a WeChat article, save it as article.md, and do not publish":
end_to_end, reuse the materials and complete the necessary editing. - "First complete only Stage 1, and continue after I confirm the Brief": stop at the Brief checkpoint and do not write the main text in advance.
- "Only review the logic of this manuscript, and do not change the author's stance":
single_stage, deliver only the logic review. - "Write one version each for WeChat and Zhihu, and complete them consecutively": write separately for the two platforms; if there is no independent agent, complete them sequentially and conduct an honest self-check.
For detailed operations, see Execution Manual. Trigger words include "write an article", "write a WeChat official account post", "write for Zhihu", "self-media article", "content creation", and "/media-writer".