content-evolution

name	content-evolution
description	Codex-native workflow for turning local long-form WeChat/Markdown/HTML articles into short social-card packages. Use when the user asks to select a high-potential article, generate low-cognition/emotional/opinionated card copy with Codex, create or register AI cover/background images, render Guizang social cards, or evolve content direction from market feedback in this repository.

Content Evolution

Use this skill inside content-evolution-workbench to run the content workflow through Codex, not through a separate model API. Codex writes structured JSON and may use the app's image generation capability; local scripts handle article selection, context building, validation, SQLite writes, asset registration, and rendering.

Workflow

1. Select candidate articles:

   npm run select-topic -- --limit 10 --tag AI --year 2026
   

   Omit filters when the user asks for the strongest current candidate.

2. Build generation context for one article:

   npm run context -- --article-id <articleId> --platform xhs > work/context.json
   

   Read work/context.json before writing the brief. It includes article excerpts, score reasons, market memory, and the required output contract.

3. Generate structured brief JSON with Codex:

   • Return JSON only: { "briefs": [...] }.
   • Produce exactly three variants unless the user requests otherwise: calm, conflict, confidence.
   • Preserve the source article's core judgment; do not invent data, results, citations, or revenue.
   • Keep anxiety tied to a real pressure and include an action/confidence outlet.
   • Use the schema in references/brief-schema.md.

4. Save the Codex-generated brief:

   npm run save-ai-brief -- --article-id <articleId> --platform xhs --file work/briefs.json --context-file work/context.json
   

   Use the returned briefs[].id for image and rendering steps.

5. Generate or register a cover image:

   • Prefer Codex App image generation when available.
   • Generate a cover/background image only; do not ask the image model to render Chinese title text.
   • Save the image under generated/<article-or-brief>/assets/.
   • Register it:

     npm run save-image-asset -- --brief-id <briefId> --file <imagePath> --prompt "<original prompt>"
     

6. Render the final card package:

   npm run render-with-asset -- --brief-id <briefId> --image-asset-id <assetId> --platform xhs --style editorial
   

   Use swiss only when the user wants a more restrained layout.

7. Build reusable content assets:

   • Use the workbench API or UI to extract reusable atoms from articles and briefs.
   • Atoms are judgments, hooks, reframes, actions, discussion prompts, and risk boundaries.
   • Prefer atom reuse when the user asks for daily candidates or long-term content evolution.

8. Generate daily candidates:

   • Use POST /api/content-plan/generate for a bounded daily plan.
   • Default shape is 1-3 candidates: main, reuse, experiment.
   • Treat the output as a human-review queue, not an auto-publishing queue.

9. Learn from feedback:

   • Feedback CSV/Excel should include articleId or briefId when possible.
   • Importing feedback updates strategy memory, content atom performance, and learning logs.
   • Optimize for interaction quality, not raw clickbait.

Guardrails

• Treat the original article directory as read-only.
• Keep generated state in the configured storage directory. By default this repository uses data/, generated/, and outputs/; set CONTENT_EVOLUTION_HOME or the per-directory environment variables when the plugin should persist state outside the source checkout.
• Never write directly to SQLite by hand; use the scripts.
• Keep final text in HTML/CSS-rendered cards, not inside AI images.
• Before bulk generation, ask for a cost/volume confirmation unless the user already specified a bounded count.
• If image generation is unavailable in the current Codex session, save the image prompt and stop at a registered text brief rather than faking an image.
• Do not auto-publish. Status changes in content_plans are planning metadata only.
• Do not reuse a high-performing atom verbatim when creating new content; adapt the angle and wording.

References

• references/brief-schema.md: JSON shape and copy requirements.