| name | conversation-to-context |
| description | 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. |
Import a conversation into the repo as context
The common workflow. A ChatGPT thread holds decisions and background that the
repo does not, and the point is to make it readable by whatever agent opens the
repo next.
Conventions
- Destination is
context/ in the repository being worked on — the current
working directory's repo root, not the plugin's directory. Create it if it
does not exist.
- The human is labelled
User, not by name. This output is for an agent, and
User is unambiguous in a way a first name is not. Use a real name only if
the user explicitly asks — that is what render-typst-pdf is for.
- Filename:
context/chatgpt-<YYYY-MM-DD>-<slug>.md, date being when the
conversation was held, not when it was imported.
- YAML front matter stays. It carries the date, model, turn counts and source
URL, and an agent reading the file cold needs all of it.
Steps
-
Capture and normalize — see import-chatgpt-conversation.
-
Find the repo root and make the folder:
ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
mkdir -p "$ROOT/context"
-
Render:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/render_conversation.py" \
<normalized.json> \
--user-label "User" \
--assistant-label "ChatGPT" \
--sources section \
-o "$ROOT/context/chatgpt-<date>-<slug>.md"
-
Check whether context/ is git-ignored:
git check-ignore -q "$ROOT/context" && echo IGNORED || echo TRACKED
If it is tracked, the transcript will be committed. Say so before staging
it and confirm the user wants a ChatGPT conversation in the repo history.
If the thread is personal, offer redact-conversation first, or suggest
adding context/ to .gitignore.
-
If the repo has a CLAUDE.md, consider adding a line pointing at the new
file, so the next agent finds it without being told. Ask first — do not
silently edit CLAUDE.md.
Keeping it useful to read
Default rendering keeps every turn. For a long thread imported purely as
background, --no-summary and --heading-level 2 keep it tight. Do not
summarize the conversation into the file instead of transcribing it: the value
is the actual exchange, and a summary is something the next agent can produce
for itself but cannot reverse.
If the conversation produced canvas documents, run extract-artifacts and save
them next to the transcript as real files — a document quoted inside a
transcript is much harder to use than one sitting on disk.
