• The final main product is a single HTML article, not a web application.
• The article can have a

  Raw

  free layer (any HTML / CSS / JS / React: interactions, layout typesetting, animations, small tools, on-demand SVG / canvas illustrations), but it must serve reading, explanation, argumentation, rhythm, or aesthetics.
• Do not generate: backends, forms, drag-and-drop workbenches, complete dashboards, product prototypes, or general web apps.
• Information density is confirmed by the user; 100% information retention is default, producing long-form web articles.

original.*

review/source-review.md

review/first-spread-review.md

review/final-review.md

1. Plan Checkpoint (Phase 2 → Checkpoint 1) Prohibit starting SubAgent for quality inspection. After the Main Agent finishes writing plan.md, on-site check against the 5-item checklist (see the Plan self-inspection section in

   references/review-checklist.md

   ), revise

   plan/plan.md

   according to the conclusion, do not write any review files, then enter Checkpoint 1.
2. First Spread / Final Review must use SubAgent (the value of SubAgent at these two nodes > cost); only if the SubAgent environment is not detectable, the Main Agent will take over, and note "No SubAgent environment, Main Agent takes over" at the top of the file.
3. Section Reviewer uses SubAgent, but the return value is a message (pass / fail + repair points); when receiving fail items, the Main Agent directly fixes them, do not let SubAgent write

   review/section-NN-review.md

   files.
4. After receiving any quality inspection conclusion —— first revise the output according to the fail items, then report "completed + self-inspection conclusion + changes made". Reporting the original conclusion directly without fixing = violation.
5. Iron Rule for Decision Collection · Prohibit Silent Selection for Users: At each Checkpoint (1 / 2 / 3), all decision items requiring user confirmation must be listed independently + wait for user reply. The Agent can make recommendations ("I recommend X because …"), but cannot say "I've already decided X for you, let me know if it's wrong" —— this deprives the user of the right to choose.
   • Priority: If the environment has the

     AskQuestion

     tool, each decision item is treated as an independent question (multiple questions can be passed in one call), and the user can confirm each item via selection cards.
   • Otherwise: Stop and list all questions numbered in the message (each question occupies a separate paragraph, clearly stating the recommended option + reason + alternatives), explicitly say "I will wait for your item-by-item reply before continuing", do not proceed with any subsequent work.
   • Never: Package multiple decisions into a single "Accept all my recommendations / All OK?" yes/no question; also do not proceed to the next step by default after a "recommendation sentence".

references/review-checklist.md

references/harness.md

references/source-to-markdown.md

scripts/source-to-markdown-markitdown.py

scripts/source-to-markdown.py

references/article-types.md

references/information-density.md

references/plan-template.md

references/theme-selection.md

references/layout.md

references/asset-policy.md

references/cover.md

references/article-types/<type>.md

theme-profiles/*.md

references/section-build.md

references/component-policy.md

references/raw-policy.md

theme-profiles/<id>.md

references/cover.md

references/scaffold.md

references/html-output.md

references/review-checklist.md

references/repair-policy.md

references/html-output.md

references/pdf-output.md

Agents tend to forget principles in long sessions —— Phase 5 will implement N Sections repeatedly, review

component-policy.md

+

raw-policy.md

+ current theme

theme-profiles/<id>.md

before starting each time.

• User specified language → Record in the "Target Language" section of the Brief in

  plan/plan.md

  . If it is inconsistent with the source material language, Phase 1 needs to first produce a natural translation of the source text, and subsequent work will be based on the translated version (see Phase 1).
• User did not specify → Default to final article language follows source material language, no translation.

source/source.md

source/extraction-notes.md

references/source-to-markdown.md

references/source-to-markdown.md

extraction-notes.md

original.*

review/source-review.md

source.md

• User did not specify target language, or target language is consistent with source → No translation, subsequent work is directly based on

  source.md

  , final article language = source language.
• User specified target language and it is inconsistent with source → First produce a natural translation

  source/source.<lang>.md

  (e.g.,

  source.zh.md

  /

  source.en.md

  ), which serves as the fact base for subsequent Phase 2+; keep the original

  source.md

  for reference. Translation requirements: Use natural target language, eliminate translation腔 (restructure sentences according to the expression habits of the target language, do not translate word-for-word, avoid rigid foreign language word order / passive stacking / non-standard punctuation), keep terms / numbers / code / formulas / references accurate, and maintain the same structure and information retention ratio. Write translation instructions into

  extraction-notes.md

  .

plan/plan.md

references/plan-template.md

• Brief: Target readers / article type / information retention ratio / must retain / can delete / tone / main viewpoints / reading goals / target language / layout width / TOC / image strategy.
• Outline: Hero / Lead / Summary / Section list / information to retain in each section / whether each section needs Raw·Table·CodeBlock·Formula·Image / ending method.
• Theme: Selected theme + reason + conflict explanation (see

  references/theme-selection.md

  ).
• Assets: Image strategy and item-by-item image plan (see

  references/asset-policy.md

  ; write one sentence in this section for

  none

  mode).

references/article-types.md

references/information-density.md

plan/plan.md

references/review-checklist.md

plan/plan.md

review/plan-review.md

• Priority

  AskQuestion

  Tool: Pass each item as an independent question (multiple questions can be passed in one call), and the user confirms each item via selection cards.
• No

  AskQuestion

  Tool: Stop and list each question numbered + in a separate paragraph + clearly stating recommended option + reason + alternatives in the message, explicitly say "I will wait for your item-by-item reply before continuing", do not proceed with any subsequent work.

longform · ~100%

full-report · ~80%

tutorial · ~90%

explainer · ~80%

dialogue · ~80%

review · ~70%

essay · ~70%

interactive-explainer · ~25% original excerpt + 75% AI reconstruction

briefing · ~50%

visual-essay · ~40%

longform + 20%

theme-profiles/index.json

regular

Image

Raw

briefing

dialogue

references/cover.md

• Final article language: Follows source language (unless the user has specified it earlier / translation has been completed).
• Whether editing, deleting, reorganizing, and rewriting tone is allowed: Allowed by default (implemented according to the above information retention ratio).
• Whether to preview the first screen sample first: Yes by default (this is Phase 4).
• TOC: On by default.

plan/plan.md

plan/plan.md has been written (self-inspection passed). I will confirm 5 items with you one by one: article type / theme / layout width / image mode / cover.

My recommendations are provided for reference (I will not choose for you):
- Type: <X> (includes standard information retention <Y%>. Reason: …)
- Theme: <theme> (Reason: …)
- Layout Width: <width> (Reason: …)
- Image Mode: <strategy> (Reason: …)
- Cover: On / Off (Reason: …; if On, composition idea: …)

Default settings that you can override: Language follows source language; editing, deleting, and reorganizing are allowed; TOC is on; the first screen sample will be made next. If you want to deviate from the standard information retention ratio for the type, tell me the specific percentage directly after answering the questions below (e.g., "longform but only 60%").

Please confirm each item below.

plan/plan.md

The first screen + first Section is ready, preview with npm run dev on localhost.
Quality inspection conclusion can be found in review/first-spread-review.md (already fixed according to fail items, list of changes made is included).
Please confirm two items independently: 1) Acceptance conclusion 2) Subsequent development mode.

1. Acceptance Conclusion —— Options:

   Pass · Enter full generation

   /

   Partial modification · I will specify changes in another message

   /

   Theme or layout is inappropriate · Return to Checkpoint 1

   .
2. Subsequent Development Mode —— Options:

   A · Single Agent sequential (default · most stable · most consistent style)

   /

   B · Multiple Agent parallel (fastest · slight style differences)

   .

references/section-build.md

references/component-policy.md

references/raw-policy.md

article/sections/NN-*.tsx

article/Article.tsx

article/raw-blocks/NN-*.tsx

• A · Single Agent sequential (default): Main Agent writes each

  sections/NN-*.tsx

  sequentially, most stable and consistent in style.
• B · Multiple Agent parallel: Subagents each own one

  sections/NN-*.tsx

  file for parallel development; Main Agent is responsible for merging and stability —— maintains the import and order of

  Article.tsx

  , runs

  npm run typecheck

  /

  build

  , ensures theme and style consistency, resolves conflicts. Subagent prompt templates can be found in

  references/section-build.md

  .

--ra-*

review/section-NN-review.md

review/final-review.md

references/review-checklist.md

1. Editorial Reviewer: Article nature, information selection, structure.
2. Visual Reviewer: Theme, Raw blocks, images, mobile adaptation.
3. Technical Reviewer: Build, console, code / formulas, accessibility.

• Delivery Decision —— Options:

  Pass · Export HTML for delivery

  /

  Pass · Export HTML + PDF simultaneously

  /

  Partial repairs needed · I will list specific changes

  /

  Pause · I want to review again

  .

references/html-output.md

• article/article.html

  (self-contained single page, CSS + JS inline, accessible offline) —— Main deliverable.
• Optional

  article/article.pdf

  : Only generate if the user selects "Pass · Export HTML + PDF simultaneously" in Checkpoint 3. Command:
  bash

  bash <path-to-beautiful-article>/scripts/html-to-pdf.sh

  The script detects installed chromium-family browsers in the system, injects

  @media print

  overrides (TOC collapses from left-right grid to top-bottom layout / TOC occupies the first page alone), and prints in headless mode. Zero npm dependencies. See

  references/pdf-output.md

  for detailed principles / troubleshooting.
• Brief editing instructions: Article type / information retention ratio / theme / image strategy / main editing choices.

• Output single HTML; article type

  longform

  ; 100% information retention.
• Language: If user did not specify, follow source material language; if specified and inconsistent with source, first produce a natural translation

  source/source.<lang>.md

  then write based on it (eliminate translation腔, see Phase 1).
• Theme: Prioritize

  tufte

  for technical / evidence-focused content, prioritize

  press

  for narrative / comment-focused content (recommend based on source material).
• Layout: Width defaults to

  regular

  , TOC is on by default (decoupled from theme, see

  references/layout.md

  , both confirmed in Checkpoint).
• Images: Image mode is a mandatory option in Checkpoint 1 (

  none

  /

  user-assets

  /

  placeholders

  /

  ai-generated

  ), only decides whether to use external

  Image

  , do not actively generate AI images.
• Raw: Orthogonal to images, always enabled by default, encourage frequent use, but must serve specific paragraphs and use theme tokens. Selecting

  none

  does not affect Raw.
• Self-inspection: Plan inline self-inspection (no SubAgent, no files); First Spread and Final Review use SubAgent + write files; Section uses SubAgent + message return (no files). See "Rigid Quality Inspection Protocol" section.
• Decision Collection: Checkpoint 1 / 2 / 3 confirm each item independently · Prohibit silent selection for users. Can recommend, cannot skip. Priority

  AskQuestion

  tool (each item as an independent question); if no tool, stop and list numbered questions to wait for user replies.
• Repair: Minimal slice, write

  review/repair-log.md

  only if there are repairs.
• Colophon · Cannot Be Removed: The scaffold comes with a colophon Raw block at the end of

  article/Article.tsx

  (

  Made with [beautiful-article](github repo) · <theme> theme

  , low-contrast small text, theme token adaptive). Must be retained in every article, prohibit deletion, prohibit moving to beside Hero or floating to the corner. Synchronously update the theme name in the colophon +

  <ThemeProvider theme="...">

  in

  main.tsx

  when switching themes.
• Cover · On by default · Must have both image and text: The scaffold creates a 3:4 screen / PDF first-page exclusive book-style cover shell + placeholder in

  article/Cover.tsx

  by default (

  --no-cover

  turns it off). The cover is located above TOC + Hero + main text, exists independently. In Phase 4 First Spread, the Main Agent replaces

  <CoverPlaceholder />

  with a customized image + text composition based on theme + article主旨. Hard Constraints: Shell ratio / print pagination cannot be modified, must have visual elements + text, only use

  --ra-*

  tokens, no remote images, no duplicate Hero content. Visual Technology Fully Open: SVG / CSS / Canvas / complex React components / any mix can be chosen by the Agent, as long as the effect is good. See

  references/cover.md

  for details (includes 5 self-inspection items + 5 composition templates + starting points for covers of each theme). PDF export will automatically make the cover occupy the first page alone, and TOC starts from the second page.
• PDF Export · Optional: The main deliverable is always

  article/article.html

  . Only if the user selects "Pass · Export HTML + PDF simultaneously" in Checkpoint 3, run

  bash <skill>/scripts/html-to-pdf.sh

  to generate

  article/article.pdf

  ; do not generate if not selected. Do not export PDF by default for users. See

  references/pdf-output.md

  for details.

• It is first and foremost an article.
• The final article language conforms to the user's intention (unspecified = follows source language; specified = unified to target language throughout, natural, no translation腔, no residual source language fragments).
• The user-confirmed information density is respected; no key content from the source material is lost unexpectedly.
• Theme temperament is consistent; images and Raw blocks all serve reading.
• The page is more worth reading than Markdown; HTML can be directly opened and shared.
• At 40% information, it reads like an edited article, not a shortened summary; at 100% information, it reads like a polished long article, not a copy of the original text.