outline-revision-annotator（大纲修订注释器）

Outline Revision Annotator

Compare a new outline/script PDF with an older baseline, then produce:

• a UTF-8 markdown report
• a DOCX report generated from that markdown
• an annotated PDF with visible yellow note boxes in the page margin

This skill is for structural revision review, not line editing.

Invocation Hooks

Explicit hook

• $outline-revision-annotator

English trigger phrases

• compare revised outline PDF with old DOCX
• compare new and old script versions
• annotate structural changes in PDF
• generate an annotated diff PDF
• mark only structural additions or rewrites
• compare character bio versions
• create a high-signal change report

中文触发词

• 对比新旧大纲
• 对比新版 PDF 和旧版 DOCX
• 对比剧本改动
• 对比人物小传版本
• 生成对比报告
• 生成标注版 PDF
• 结构性改动
• 只标新增和重写
• 差异标注
• 对比标注版

Auto-follow hook

Run this skill automatically after the agent itself edits any of the following:

• outline
• script
• treatment
• beat sheet
• season map
• episode breakdown
• character bio
• role profile

Automatic follow-up applies when all of these are true:

• there is an older baseline version to compare against
• the revised draft is already a PDF, or can be exported to PDF in the current environment
• the user did not explicitly opt out of comparison output

When the auto-follow hook fires, generate a fresh:

• compare_report.md
• compare_report.docx
• annotated comparison PDF

If no baseline exists, or the revised draft cannot be turned into a PDF, say what is missing and stop.

中文规则：

• 当 AI 自己刚改完大纲、剧本、人物小传、季纲、分集梗概等内容后，如果存在旧版本基线，并且新版可获得 PDF，就自动触发本 skill
• 自动补产物时，不需要等用户再次提醒“再生成一份对比版”
• 如果缺旧版基线，或当前环境无法拿到新版 PDF，就明确说明缺什么，不要假装已完成

Use This Skill For

• Story bibles, season outlines, beat sheets, treatments, scripts, and character bios
• new.pdf versus old.docx review workflows
• new.pdf versus old.pdf review workflows when the old version is also a PDF
• Chinese-language documents where PDF popup comments often render as garbled text
• Deliverables that need to be readable without clicking annotation popups

Do Not Use It For

• punctuation or wording polish
• paragraph-level copy edits that do not change story function
• literal redline review
• mark-everything diffs

Qualification Rule

Mark only changes that alter structure or story function:

• new season engine, thematic spine, or professional framework
• new antagonist, institutional pressure line, or evaluation thread
• a character arc rewritten to serve a different dramatic job
• ending reset from a closed loop to a next-season hook
• a case line upgraded from emotional texture to ethical or institutional conflict

Do not mark:

• sentence rewrites with the same dramatic function
• softened or intensified phrasing
• richer detail that does not change structure

Workflow

1. Make sure the revised draft exists as PDF

If the agent has just edited a DOCX, Markdown, or other text source, first export or obtain a revised PDF. The renderer marks the revised PDF, not the editable source file.

2. Extract the sources

Run the helper script first. It produces plain text and page/block maps for the new PDF, plus plain text for the old baseline.

• Windows commands: references/windows.md
• macOS commands: references/macos.md

3. Compare semantically

Read the extracted text and decide which changes are structural. Then author two files:

• compare_report.md
• annotation_manifest.json

Manifest schema and examples: references/manifest-schema.md

4. Render the final deliverables

Run the same script in render mode. It will:

• create an annotated PDF with visible margin boxes
• copy the markdown report as UTF-8
• convert the markdown report into DOCX
• write a normalized manifest copy into the output folder

5. Visually verify the PDF

Open the annotated PDF and check:

• the right-side yellow box text is visible without clicking
• Chinese text is not garbled
• note boxes do not cover the source paragraph
• only structural changes were marked

If something is wrong, read references/troubleshooting.md.

Non-Negotiable Rules

• On Windows PowerShell, never pipe Chinese Python source directly into python -. Save a real UTF-8 .py file and run the file.
• Do not rely on popup PDF comments for Chinese review notes. Write visible note boxes into the page margin.
• If a change is stylistic rather than structural, leave it unmarked.
• Prefer an ASCII fallback filename if a downstream tool or browser tab cannot handle Chinese filenames cleanly.
• If the auto-follow hook fired, do not silently skip the compare PDF. Either generate it or state the concrete blocker.

Deliverables Checklist

• annotated.pdf
• compare_report.md
• compare_report.docx
• annotation_manifest.normalized.json

Script Entry Points

The bundled script is:

• scripts/build_outline_diff_outputs.py

Supported modes:

• extract: dump source text and PDF block maps
• render: build the annotated PDF and report outputs

What the Script Does Not Decide

The script is deterministic. The agent still decides:

• which changes count as structural
• which blocks should be grouped into one note
• how to phrase the comparison report
• how to phrase each yellow-box explanation

Keep that judgment in the agent. Keep extraction and rendering in the script.