Chat Share Skill — AI skill for Claude Code

Paste a `chat.deepseek.com/share/...` or `kimi.com/share/...` link. You get the full conversation, and every inline citation resolved to the page the model actually read. The skill then checks each claim against its primary source before writing anything down.

The repo has two parts, and each works on its own:

• scripts/fetch_share.py is a plain Python CLI. It uses only the standard library and needs no browser, cookies, or account.
• SKILL.md is the workflow: fetch, list the claims, verify each one against its primary source, then write up only what checks out, with a corrections table for the rest.

Why this exists

Share pages and the answers in them both resist the naive approach:

• A share page is an empty app shell. Fetching it returns
  . The conversation comes from an undocumented API (a REST call for DeepSeek, a Connect RPC for Kimi), and some Kimi shares answer "not found" until you get a guest token.
• Kimi citations are pointers, not links. An answer cites tools://web_search:3#16, an index into that tool call's results. The index is 0-based, while result ids start at 1. The script resolves each pointer to its URL and keeps the quoted text and Kimi's cached copy of the page. That cached copy is often the only way to read an OpenReview PDF that returns 403.
• Answers drift from their sources. Checking two real shares against the papers turned up a figure whose values were invented, a table read with its columns shifted, two ablation variants swapped, a real-