菜单
Render a ChatGPT conversation as a styled PDF using a Typst template, with colour-coded user and assistant turns, a metadata cover block, artifact callouts and a sources page. Use when the user wants a PDF, a document to read or share, or to be identified by their own name.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/danielrosehill/Claude-ChatGPT-Importer-Plugin --skill render-typst-pdf命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Capture a ChatGPT conversation as raw JSON by calling the authenticated backend API from inside the browser tab. Use when importing from a chatgpt.com/c/ URL or a /share/ link. This is the primary, complete capture route.
Fallback capture of a ChatGPT conversation by scraping the rendered DOM with scrolling. Use only when capture-via-api fails. Lossy - no timestamps, branches, tool messages or artifact payloads.
Import a ChatGPT conversation into the current repository as agent context. Use when the user wants to bring a ChatGPT thread into a project, pull a chat into context, or save a conversation for Claude to read later. Writes a Markdown transcript to context/ with the human labelled User.
正在显示 SKILL.md
| name | render-typst-pdf |
| description | Render a ChatGPT conversation as a styled PDF using a Typst template, with colour-coded user and assistant turns, a metadata cover block, artifact callouts and a sources page. Use when the user wants a PDF, a document to read or share, or to be identified by their own name. |
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/render_pdf.py" <normalized.json> \
-o <out.pdf> --user-label "User" --timestamps
Requires pandoc (>= 3.0, for the typst writer) and typst on PATH. The script
checks both and tells you what is missing.
This is the renderer where identity is a real choice. --user-label sets what
appears on every human turn and in the cover block.
User — neutral, right for anything machine-facing or shared onward.Ask which one the user wants if they have not said, rather than defaulting
to their name. A transcript that leaves this machine with a real name on it is
harder to walk back than one that does not. --assistant-label works the same
way and defaults to ChatGPT.
REDACTED badge if the conversation went through
redact-conversation.Layout lives in ${CLAUDE_PLUGIN_ROOT}/templates/transcript.typ. Edit colours
and fonts there. Pass --template <path> to use a different one, and
--paper us-letter for US paper.
Each turn's Markdown body is converted to Typst by pandoc, then wrapped in a
#turn(...) call. Turns are split by the script, not recovered afterwards with
a show-rule that parses heading text — so styling never depends on guessing
which speaker a heading belonged to.
All bodies go through pandoc in one call separated by a sentinel, because one invocation per turn costs roughly twenty times as much. The script asserts that the split count matches the turn count and aborts if it does not, rather than emitting a transcript with two people's words in one block.
Use --keep-typst <path> to inspect the generated Typst when a compile fails.
A ChatGPT asset pointer (file-service://file-...) is an internal reference,
not a path. It is deliberately not rendered as an image — doing so makes
Typst try to open it as a file and abort the whole document. Images appear as
artifact callouts naming the pointer. To embed them for real, the asset has to
be downloaded first and local_path set on the artifact record.