init-doc

name	init-doc
description	Use when initializing a translation project from extraction through glossary, chapter mapping, and progress tracking.
user-invocable	true
disable-model-invocation	true

Initialize Document Translation

Overview

Initialize translation baseline from PDF extraction to chaptered docs, style decisions, glossary, and progress tracker.

Core principle: Build a deterministic, verifiable baseline before any large-scale translation.

Task Initialization (MANDATORY)

Before ANY action, create tasks using TaskCreate:

• One task per major phase (extraction, formatting, images/theme, terminology, chapter split, progress tracker)
• One task for final handoff gate

Interaction Rules

• All user interaction must be Traditional Chinese (zh-TW).
• AskUserQuestion prompts must be Traditional Chinese.
• Do not use Simplified Chinese in user-facing text.

The Process

Step 1: Cleanup and Source Validation

Run cleanup:

uv run python scripts/clean_sample_data.py --yes

Then resolve source PDF from $ARGUMENTS or ask user in Traditional Chinese. Ensure source is under data/pdfs/.

Before extraction, ask user in Traditional Chinese whether to preserve PDF images in the generated docs:

是否要保留 PDF 內的圖片，並在切分後的 Markdown 文件中保留對應圖片連結？

Record this decision as preserve_images: true/false for the rest of the run.

Verification: Source PDF exists under data/pdfs/; preserve_images decision recorded; cleanup script exited 0.

Step 2: Create Tasks

Create TaskCreate items for:

• extraction
• formatting decisions
• image retention decision
• image and theme setup
• terminology bootstrap
• chapter mapping
• progress tracker creation
• final handoff gate

Verification: All tasks created with correct descriptions; task list matches the phases above.

Step 3: Extract and Validate Raw Outputs

Run:

uv run python scripts/extract_pdf.py <pdf_path> --include-images

If preserve_images is false, run instead:

uv run python scripts/extract_pdf.py <pdf_path> --no-include-images

Validate outputs:

• data/markdown/<name>.md
• data/markdown/<name>_pages.md
• data/markdown/images/<name>/（only when preserve_images = true）

Verification: All expected output files exist and are non-empty.

Step 4: Cropping Review and Optional Split

Review readability and completeness. If needed, split large source into parts and re-extract until clean.

Verification: Extracted markdown is readable; no garbled text or truncated sections remain.

Step 5: Confirm Document Formatting Decisions

Summarize content to user in Traditional Chinese:

書本內容概覽：
- 主要內容類型：[規則說明、範例場景、角色選項...]
- 特殊結構：[大量表格、骰表、設計者備註...]
- 建議可使用的格式化元件：[...]

Collect formatting choices (Traditional Chinese):

• layout profile for extraction (auto / single-column / double-column)
• aside mapping (note/tip/caution/danger)
• card/tabs usage
• table/dice-table conventions

Persist via script:

uv run python scripts/style_decisions.py init
uv run python scripts/style_decisions.py set-document-format \
  --layout-profile "<auto|single-column|double-column>" \
  --aside-note "<note_component>" \
  --aside-tip "<tip_component>" \
  --aside-caution "<caution_component>" \
  --aside-danger "<danger_component>" \
  --cards-usage "<cards_usage_note>" \
  --tabs-usage "<tabs_usage_note>" \
  --tables-convention "<table_note>" \
  --dice-tables-convention "<dice_table_note>"
uv run python scripts/validate_style_decisions.py

Verification: style-decisions.json contains document_format section; validate_style_decisions.py exits 0.

Step 6: Select Images, Theme, and Homepage Content

1. If preserve_images = true, ask user to assign extracted images for hero/background/og.
2. If preserve_images = true, copy and resize where needed.
3. If preserve_images = false, skip extracted image assignment and continue with theme-only setup.
4. Ask theme decisions in Traditional Chinese (mode/overlay/palette).
5. Update docs/src/styles/custom.css and persist style decisions.
6. Persist image retention decision via:

uv run python scripts/style_decisions.py set-images --preserve-images <true_or_false>

7. Ask for site meta in Traditional Chinese (all four fields):

   • 網站標題（site.title）：首頁 <title> 及 frontmatter title，例：「Rapscallion 遊戲規則」
   • 首頁描述（site.description）：SEO description，一句話
   • 副標語（site.tagline）：hero 區塊顯示的一行短語
   • 內容簡介（site.intro）：首頁「內容簡介」段落，一到兩句 Persist via:

   uv run python scripts/style_decisions.py set-site \
     --title "<USER_INPUT>" \
     --description "<USER_INPUT>" \
     --tagline "<USER_INPUT>" \
     --intro "<USER_INPUT>"
   

8. Ask for copyright and credits in Traditional Chinese:

   • Copyright notice text（例：© 2024 Author Name. All rights reserved.）
   • Credits entries as role → name pairs（例：原作者、翻譯、美術設計等）
   • Whether to show each section on the homepage

9. Persist via:

uv run python scripts/style_decisions.py set-copyright \
  --text "<USER_INPUT>" \
  --show-on-homepage <true_or_false>
uv run python scripts/style_decisions.py set-credits \
  --entry "原作者:..." \
  --entry "翻譯:..." \
  --show-on-homepage <true_or_false>

11. Ask whether there are any translation-wide notes the translator must always follow. Persist each note via:

uv run python scripts/style_decisions.py add-translation-note \
  --key "<short_key>" \
  --topic "<optional_topic>" \
  --note "<USER_INPUT>"

If the note is specific to one source file or future appended document, use:

uv run python scripts/style_decisions.py add-translation-note \
  --document-key "<pdf_stem_or_doc_id>" \
  --key "<short_key>" \
  --note "<USER_INPUT>"

12. Run:

uv run python scripts/validate_style_decisions.py

generate_nav.py will render these as ## 版權宣告 and ## 製作名單 sections on the homepage. If neither is provided, a generic fallback disclaimer is used.

Verification: validate_style_decisions.py exits 0; style-decisions.json contains site meta, copyright, credits, and image decisions.

Step 7: Build Terminology Baseline

Invoke term-decision skill for terminology bootstrap instead of duplicating the workflow here.

Required handoff to term-decision:

1. Source is the extracted markdown from this init run.
2. First inspect high-signal terminology sources in the original book:
   • source glossary / terminology pages
   • index pages
   • appendix term lists
   • playbook / move summary tables that clearly define recurring mechanics terms
3. Complete one first-pass terminology bootstrap from those sections:
   • prefill obvious term translations into glossary.json
   • keep wording consistent with style-decisions.json
   • ask the user only about uncertain, culturally nuanced, or mechanics-ambiguous terms
4. After that first pass, generate and verify the remaining candidates with:

uv run python scripts/term_generate.py --min-frequency 2
uv run python scripts/term_cal_batch.py
uv run python scripts/validate_glossary.py
uv run python scripts/term_read.py --fail-on-missing --fail-on-forbidden

init-doc must not continue to chapter split until the term-decision handoff completes cleanly.

Verification: validate_glossary.py and term_read.py --fail-on-missing --fail-on-forbidden both exit 0.

Step 8: Chapter Split and Navigation

Invoke chapter-split skill instead of duplicating split logic here.

Required handoff to chapter-split:

1. Source is the extracted _pages.md file from this init run.
2. Reuse the preserve_images decision from Step 1.
3. Reuse the formatting and terminology decisions already completed in Steps 5-7.
4. Generate chapters.json, split docs output, and regenerate navigation.
5. If chapter-split reports unresolved critical issues, stop init-doc and resolve them before continuing.

Verification: chapters.json exists; split docs generated; nav regenerated.

Step 9: Create Translation Progress Tracker

Create data/translation-progress.json from chapters.json:

uv run python scripts/init_create_progress.py --force

Tracker contract:

• chapter ids derived from output file paths
• source page ranges mapped from chapter config
• initial status not_started
• _meta fields (updated, total_chapters, completed)

Verification: data/translation-progress.json exists; contains all chapters from chapters.json with status not_started; _meta fields present.

Step 10: Final Gate and Handoff (Fail-Closed)

Run one-shot handoff gate:

uv run python scripts/init_handoff_gate.py

If any gate fails, stop and fix before completion.

Verification: init_handoff_gate.py exits 0; all tasks marked completed.

Flowchart

digraph init_doc {
    rankdir=TB;
    cleanup [label="Cleanup &\nsource validation", shape=box];
    tasks [label="Create tasks", shape=box];
    extract [label="Extract PDF", shape=box];
    img_decision [label="Preserve\nimages?", shape=diamond];
    extract_img [label="Extract with\n--include-images", shape=box];
    extract_noimg [label="Extract with\n--no-include-images", shape=box];
    crop [label="Cropping review\n& optional split", shape=box];
    format [label="Formatting\ndecisions", shape=box];
    theme [label="Images, theme\n& homepage", shape=box];
    term [label="Terminology\nbaseline\n(term-decision)", shape=box];
    term_ok [label="Term validation\npasses?", shape=diamond];
    split [label="Chapter split\n(chapter-split)", shape=box];
    split_ok [label="Split\nsucceeded?", shape=diamond];
    progress [label="Create progress\ntracker", shape=box];
    gate [label="Final handoff\ngate", shape=box];
    gate_ok [label="Gate\npasses?", shape=diamond];
    done [label="Done →\ntranslate", shape=box];
    fix [label="Fix & retry", shape=box];

    cleanup -> tasks -> extract;
    extract -> img_decision;
    img_decision -> extract_img [label="yes"];
    img_decision -> extract_noimg [label="no"];
    extract_img -> crop;
    extract_noimg -> crop;
    crop -> format -> theme -> term;
    term -> term_ok;
    term_ok -> split [label="yes"];
    term_ok -> fix [label="no"];
    fix -> term;
    split -> split_ok;
    split_ok -> progress [label="yes"];
    split_ok -> fix [label="no"];
    progress -> gate;
    gate -> gate_ok;
    gate_ok -> done [label="yes"];
    gate_ok -> fix [label="no"];
}

Progress Sync Contract (Required)

1. Keep tasks updated via TaskUpdate at every step.
2. Mark blockers immediately and include failing command/context.
3. Close tasks only after final gate passes.

When to Stop and Ask for Help

Stop when:

• source extraction repeatedly fails
• chapter-split cannot produce a usable config
• glossary validation cannot be resolved safely
• docs build fails with unclear root cause

When to Revisit Earlier Steps

Return to earlier steps when:

• user changes formatting/theme policy
• user changes proper noun strategy
• source markdown changes enough to invalidate page mapping

Red Flags

Thought	Reality
"Validation failed but it's probably fine, keep going"	Fail-closed means stop. Fix the failure before proceeding.
"Skip chapter-split handoff and split manually"	Chapter-split skill ensures deterministic structure. Never bypass.
"User didn't answer about proper nouns, I'll just pick one"	User confirmation is required for formatting and proper noun policy.
"Progress tracker can be created later"	Tracker must be initialized before handoff. No exceptions.
"Skip terminology baseline, we can add terms during translation"	Terminology drift across chapters is costly. Bootstrap first.
"I'll reuse the old glossary without re-validating"	Glossary changes between runs. Always validate.
"One quick formatting change doesn't need style-decisions.json"	All formatting decisions must be persisted. No ad-hoc overrides.

Next Step

Continue with /translate or /super-translate.

Example Usage

/init-doc
/init-doc data/pdfs/rulebook.pdf