Improve infographic prompt guidance

SKILL.md

@@ -96,8 +96,8 @@ Generate images after the Markdown is ready, so prompts come from the title, art
   - Match the article's overall visual tone.
   - Constraints: no text unless explicitly requested, no logos, no watermark, central-safe visual focus.
 - Ordinary inline illustrations:
-  - Default count: 3
-  - Use 2 for short articles and 4 for long or highly structured articles.
+  - There is no default image count. Choose zero or more images from the article's actual structure.
+  - Decide section by section whether the content needs a supporting art image, an infographic, or no image.
   - Paths: `imgs/illustration-1.png`, `imgs/illustration-2.png`, etc.
   - Place each image after the most relevant major section.
   - Match the meaning and emotional style of the nearby paragraphs.
@@ -107,6 +107,24 @@ Generate images after the Markdown is ready, so prompts come from the title, art
   - An infographic's purpose is to replace the corresponding text, not repeat it.
   - Remove only the text whose complete meaning is accurately represented by the infographic. Preserve surrounding analysis, caveats, and transitions.
   - Treat this as an allowed image-backed structural edit, while preserving the article's meaning.
+  - Prefer an infographic for dense comparisons, model-selection matrices, timelines, or decision trees when exact values can be represented clearly.
+  - Do not force every table into an image: retain an HTML-rendered table when exact numbers, multiple columns, or mobile readability make it safer than generated artwork.
+  - When generating an infographic, paste the complete source block being visualized directly into the image-generation prompt, including its original headings, labels, values, caveats, and explanatory text. Then describe the visual tone, readability, audience, and output format.
+  - Do not pre-impose a card grid, table, node layout, column structure, or other composition unless the user requests one. Let the image model choose the information architecture from the source content.
+  - Ask the model to make the source text part of the visual composition, not an afterthought. Inspect every generated label, number, unit, and line break; regenerate with a targeted correction if any content is wrong or unreadable.
+  - Only use local text overlay as a fallback when the generated text remains unreliable after a targeted regeneration or when the user explicitly asks for deterministic typesetting.
+
+### Image and infographic selection gate
+
+Before generating any inline image, make a short section map:
+
+| Content shape | Preferred treatment |
+| --- | --- |
+| Atmosphere, concept, narrative transition | Ordinary editorial art |
+| Exact comparison, pricing, benchmark, timeline, process, or decision tree | Infographic or HTML table |
+| Short list with no visual relationship | No image unless it materially improves comprehension |
+
+The final image count must be justified by the section map. Never add images merely to reach a quota.
 
 Keep all final images under the article's `imgs/` directory. Inspect generated images before use and regenerate any image that does not match the content or shared visual style.
 
@@ -125,6 +143,26 @@ After generating images:
 - Do not insert the cover image into the body unless the user asks for it.
 - For an infographic, replace the corresponding information block with the image link instead of leaving duplicate text.
 
+## Table Rendering Quality Gate
+
+Markdown tables are fragile in the WeChat pipeline. Before rendering:
+
+1. Keep every table row adjacent. Do not leave blank lines between the header, separator, and body rows.
+2. Use a valid separator row with the same column count as the header.
+3. Do not generate a table by joining every Markdown line with a blank-line separator.
+4. Preserve exact numeric values; if a table is converted to an infographic, remove only the fully represented table text.
+
+After preparing the Markdown, run `md-to-wechat.ts` or the API dry-run and inspect the generated HTML:
+
+```bash
+$BUN_X "{skillDir}/scripts/md-to-wechat.ts" "post-to-wechat/yyyy-MM-dd/slug/slug.md" \
+  --theme default --color blue > /tmp/mp-render.json
+HTML_PATH=$(node -e 'process.stdout.write(JSON.parse(require("fs").readFileSync("/tmp/mp-render.json","utf8")).htmlPath)')
+test "$(rg -o '<table' "$HTML_PATH" | wc -l | tr -d " ")" -eq "<expected table count>"
+```
+
+The expected table count is the number of Markdown tables intentionally retained after any infographic replacements. Do not publish if a retained table is rendered as plain pipe-delimited paragraphs or if the count is wrong.
+
 ## Publishing Workflow
 
 Load config from these paths in priority order:
@@ -167,4 +205,11 @@ $BUN_X "{skillDir}/scripts/wechat-api.ts" "post-to-wechat/yyyy-MM-dd/slug.md" \
   --dry-run
 ```
 
-After successful publishing, report the title, account, image count, and that the article was saved as a draft and not mass-sent.
+The dry-run must confirm all of the following before publishing:
+
+- The intended number of retained tables appears as HTML `<table>` elements.
+- The intended number of inline image placeholders is present.
+- The cover path exists and is horizontal.
+- No Markdown table contains blank lines between rows.
+
+After successful publishing, report the title, account, image count, retained table count, and that the article was saved as a draft and not mass-sent.