card-rewrite

# Card Rewrite — Teaching-Style Paper Cards ## Agent（claude / codex） 兩個 agent 皆可駕駛（Codex 端：`research-cards@private-plugins` plugin 的 card-rewrite skill）。alphaXiv MCP 的 grounding 僅 Claude 有；Codex 駕駛時以卡片現有內容＋arxiv HTML 抓取為準。 ## Backend `heptabase`/`both`：原路徑不變。`obsidian`：`read_card`/`save_card`/`list_todo` 走 vault （PM doc 為共用記憶體模型，markdown 方言承載顏色/toggle，重寫與上色流程不變）。 Turns a terse "translated report" card into one that **teaches an unfamiliar reader**. The single source of truth for the style + the easy-to-drop mechanics (preserve figures, re-colorize, fit 100k). `scholar-inbox-clip` and the overview skills reference this style; this skill owns it and drives batch retrofits. **Core principle:** the card's #1 job is to let someone who does *not* know the subfield absorb the paper fast. Simplifying reduces load, but **over-simplifying makes it unreadable** — explain the jargon, teach, don't just summarize. **MANDATORY grounding:** ALWAYS fetch the original English report (or full text) via the alphaXiv MCP **every time**, before authoring — not only when a term looks unclear. The existing card is a lossy translation; rewriting from it alone risks amplifying any gap into hallucination. The MCP original is the source of truth you write *from*; the existing card is only for diffing what's missing. ```python import sys, os; sys.path.insert(0, os.environ.get("CLAUDE_PLUGIN_ROOT", os.path.expanduser("~/.claude/skills/research-cards")) + "/skills/card-rewrite") import rewrite_lib ...