1000-2000字重写
Use this skill for the user's 1000-2000 character Douyin copy rewrite workflow when Gemini rewriting must not alter the existing douyin-copy-production-workflow-custom conversation or gemini_session.json.
Core Outcome
Produce final .docx files that contain:
爆款心理学标题包装开头版本一:高阶认知课式开头开头版本二:身份点名式硬核学习开头开头版本三:保留原文开头(来自源文档)正文
Do not add risk suggestions, yellow-highlighted annotations, or [[RISKNOTE:...]] markers.
Default Paths
- Mac source folders are often under
/Users/kin/工作用(同步)/素材文稿/. - Mac final Word output folder: create a content-specific
codex短重写输出_<YYYYMMDD>folder beside the source folder unless the user gives another output path. - Mac Gemini expansion command directory:
/Users/kin/Documents/Codex/2026-07-02/gemini - Mac Gemini expansion command:
./scripts/run_gemini_chat.sh - Mac local command shape:
./scripts/run_gemini_chat.sh --prompt-file <path> --isolated --output-file <path> - Windows fallback source folders are often under
D:\工作用(同步)\素材文稿\orE:\工作用\素材文稿\助理采集文案\. - Windows fallback final Word output folder:
D:\工作用(同步)\素材文稿\codex工作流长文稿orE:\工作用\素材文稿\codex工作流长文稿, depending on where the current synced root lives. - Windows fallback Gemini expansion command directory:
C:\Users\Administrator\Documents\Codex\2026-06-04\gemini3-1pro-api - Windows fallback Gemini expansion command:
.\outputs\run_gemini_chat.cmd - Mandatory isolation rule: always call Gemini with
--prompt-file <path> --isolated; on Mac also pass--output-file <path>so the generated draft is saved for validation. Never run the bare interactive command for expansion, because it can read or write the original long-running Gemini conversation. - Never send
/new; it can overwrite the saved Gemini session. - Opening source of truth: invoke
baokuan-kaitou-sheding/$爆款开头设定; do not hand-write openings from memory.
Reference Files
Load only the reference files needed for the current step:
references/gemini-expansion.md: read before sending source text to Gemini, creating prompt files, retrying failures, or checking generated drafts.references/source-learning-and-style.md: read when processing a batch, learning source viral logic, classifying source styles, or preserving source openings.references/openings-and-titles.md: read after expansion passes checks and before Word export to generate openings, score them, and build title packages.references/export-and-hard-rules.md: read before optional scoring, Word export, final delivery, and post-run workflow updates.
Workflow
1. Queue Every Provided Source By Default
When the user provides a batch of .docx scripts, do not run pre-expansion scoring, ranking, filtering, or top-N selection unless explicitly asked. Preserve stable filename order unless the user gives another order.
Read references/source-learning-and-style.md and create a concise batch learning note in the current workspace outputs folder.
2. Expand With Isolated Local Gemini
Read references/gemini-expansion.md.
For every expansion and retry, build a UTF-8 prompt file containing:
- The full current Gemini instruction block from
references/gemini-expansion.md. - The original source text wrapped as
【原文开始】...【原文结束】.
Then run only the Mac local command by default:
cd /Users/kin/Documents/Codex/2026-07-02/gemini
./scripts/run_gemini_chat.sh --prompt-file /path/to/prompt.txt --isolated --output-file /path/to/expanded.txt
Use the Windows command only as an explicit fallback:
cd C:\Users\Administrator\Documents\Codex\2026-06-04\gemini3-1pro-api
.\outputs\run_gemini_chat.cmd --prompt-file C:\path\to\prompt.txt --isolated
Do not run either bare interactive command:
./scripts/run_gemini_chat.sh
.\outputs\run_gemini_chat.cmd
Do not send /new. Do not use the old web expansion channel.
3. Validate Each Expanded Draft
Read references/gemini-expansion.md for the exact checks and retry language.
Verify length, topic relevance, refusal/API failure text, stale-topic contamination, and required ending. Default target length is 1000-2000 Chinese characters, with 1200-1600 preferred. The minimum acceptable length is 1000 Chinese characters and the default maximum is 2000 unless the user gives a newer threshold.
4. Add Openings And Titles
Read references/openings-and-titles.md.
For every accepted article, execute these subskills in order:
- Invoke
baokuan-kaitou-sheding/$爆款开头设定and use its approved learned opening forms as the source of truth for开头版本一:高阶认知课式开头and开头版本二:身份点名式硬核学习开头. - Extract
开头版本三:保留原文开头(来自源文档)from the source document without rewriting it. - Score versions one and two with the mandatory 20-point opening gate; rewrite with
baokuan-kaitou-shedinguntil both pass. - Invoke
viral-psych-title-wrapper/$爆款心理学标题包装器for the title package. Generate 10 three-line candidates, strongest 3, and recommendation reason. - Maintain a batch title ledger to avoid duplicate mechanism chains or generic interchangeable titles.
5. Export Clean Word Documents
Read references/export-and-hard-rules.md.
Export final .docx files on Mac into a content-specific codex短重写输出_<YYYYMMDD> folder beside the source folder, or the user-specified output folder, with content/topic-based filenames. Do not leave generic names such as risk-note or timestamp-only filenames.
Before final delivery, verify each .docx contains the required headings and has no risk suggestions, yellow highlights, or [[RISKNOTE:...]] markers.
6. Final Reply
Keep the final answer short. Tell the user the Word file path, character count, and confirm that no risk suggestions, yellow-highlighted annotations, or [[RISKNOTE:...]] markers were added.
If optional viral potential review was requested, report it. Otherwise do not add a scoring/pass-fail statement.