yao-geo-explainer-builder

Yao GEO Explainer Builder

Execution flow

1. Clarify the popular science topic, target brand, brand knowledge base, industry materials, target readers, use scenarios, brand placement restrictions, content length, compliance boundaries, and real-data availability status.
2. Identify the question type: what it is, why, how to do it, how to choose, who it suits, common misconceptions, recommended paths; each paragraph handles only one topic.
3. First perform a self-check: whether the reference sources are authoritative enough, and whether the analysis dimensions cover intent, readers, evidence, real data, entities, steps, decisions, compliance, platforms, reuse, and measurement.
4. First write an 80 to 120 character context-independent core summary, so that a single paragraph can still answer the question when excerpted by AI search.
5. First generate the "Complete GEO Article": the article must have continuous body text, natural paragraphs, subheadings, explanations, steps, selection criteria, misconceptions, and at least 6 body paragraph groups among the FAQ/closing suggestions; readers should be able to get a complete answer without reading the appendix.
6. Use references/explainer-method.md to build the delivery structure: core summary, complete GEO article, analysis overview, content map, intent matrix, reader scenarios, research basis, evidence grading, real-data access status, real-data collection plan, fact verification matrix, entity map, definitions, principles, steps, checklist, selection criteria, decision path, misconceptions, examples, FAQ, glossary, brand applicable scenarios, platform adaptation, structured data, measurement plan, content gaps, and source ledger.
7. In the body text, use ordered lists to present the How-to, and use natural paragraphs to explain reasons and boundaries; in the appendix, use key-value pairs to present prerequisite parameters, and use tables to present evidence grading, selection criteria, decision path, and misconception correction.
8. Insert the brand according to references/brand-insertion-policy.md ; the brand should only appear in positions that serve answers, such as examples, applicable scenarios, FAQ, or closing suggestions.
9. When involving sensitive fields such as health, finance, law, returns, and safety, boundary reminders must be added; effect, return, and safety assertions without evidence must be deleted or neutrally rewritten.
10. Use scripts/render_yao_geo_explainer_builder.py to generate Markdown, an HTML package with a fixed table-of-contents menu, Word, and PDF from the same structure, and produce quality-report.json。
11. If the quality report does not pass, first fix the input structure, evidence, brand count, article completeness, or layout logic, then regenerate the four-format files.

Domestic platform adaptation

• Kimi: in-depth long-form explanations, multi-source evidence, complete chapters, and glossary.
• Qwen: source numbering, separation of facts/inferences, parameterized guides, and traceable sources.
• Doubao: everyday-life questions, short Q&A, pitfall reminders, and direct selection suggestions.
• DeepSeek: step logic, judgment frameworks, conditional branches, and reasoning chains.
• Yuanbao: WeChat official account tutorials, subheading rhythm, short FAQ, and shareable paragraphs.

Real-data capability boundaries

• The skill itself does not directly guarantee online retrieval of real data; during execution, an upper-layer agent can search online, read user-uploaded materials, or use an existing knowledge base, and then write the sources into the input structure.
• When there is no authentic brand information, the report must state "not connected/pending verification" and place brand parameters, effects, safety, and qualification-related facts into the fact verification matrix.
• When authentic data has been connected, source location, collection date, purpose, and evidence level must be retained; outdated, inaccessible, or weak-source materials must not support definitive conclusions.
• The authentic data collection plan must at least cover official brand materials, industry/standard materials, and user question samples; when horizontal comparison is involved, add competitor or device form samples.

Four-format delivery rules

• Markdown is the structural master; HTML, PDF, and Word must be generated from the same section spec, and Word must not be missing sections.
• After opening the four-format examples, a complete article must be presented first; the analysis matrix, evidence, authentic data, and source ledger serve as supporting materials after the article.
• HTML/PDF use the kami layout with a paper-like light background, ivory content surface, blue-black headings, stable line height, and local fonts; table borders are aligned, long text wraps automatically, and gradients, transparent overlays, or decorative color blocks are not used.
• HTML must include a table of contents menu bar, and the menu bar must remain fixed and follow when the page scrolls down; menu items should link to each report module.
• Word must fix the A4 page, left and right margins, total table width, fixed table layout, and cell margins; long URLs, English tokens, and parameter strings must soft-wrap to avoid overflowing to the right.
• The example report must actually exist:.docx、.pdf、html-package/index.html、.md, and include quality-report.json。

Quality inspection checklist

• Popular science value takes priority over brand expression; it must not be written as a brand advertorial.
• The main body of the complete article must not be split into a mere list of modules; the main body must have at least 6 subsections and 12 natural paragraphs, and no fewer than 1400 characters after removing spaces.
• The analysis must be systematic, detailed, and complete: it must at least cover problem types, reader scenarios, source evidence, authentic data connection, fact verification, entity terminology, operational steps, decision paths, compliance boundaries, platform adaptation, structured data, and effect measurement.
• Steps must be numbered, selection criteria must be tabulated, and definitions must be independently reusable.
• The source ledger must explain source, location, date, and purpose; facts that cannot be verified are downgraded to pending confirmation.
• The number of brand appearances must not exceed the input limit; effects, benefits, safety, or industry-first claims without supporting evidence must not be promised.
• Word/PDF require manual or automated preview spot checks of the right edge: tables, URLs, long English text, and long Chinese sentences must not be clipped or touch the edge.