Capturing chats

/capture-this-chat reads your current AI session and turns it into a draft spell. This page covers per-tool setup, troubleshooting, and the privacy guarantees.

Quick reference

Tool	Setup needed	Status
Claude Code	Nothing — works automatically	v1.0
OpenCode	Nothing — works via the bundled plugin	v1.0
Cursor	Nothing — works automatically	v1.0
Codex CLI	Will work in v1.0.1	Planned
Gemini CLI	Will work in v1.0.1	Planned
GitHub Copilot CLI	Will work in v1.0.2	Planned

For tools not yet supported, save your chat to a file and use /build-spell --from-transcript <path>.

Privacy

• All capture is local. No network call is made by any wrapper.
• Tool calls in the transcript are summarized, not sent raw. (E.g. "ran web_search" is preserved; the search arguments and full results are not.)
• The captured transcript is shown to you before inference runs. You can View · Continue · Cancel.
• Nothing is written to disk except the final saved skill. The transcript stays in memory.

Secrets warning. If your transcript contains API keys, tokens, or other secrets in tool outputs, they may appear in the transcript. Review the captured transcript at the consent moment before continuing. Automated secret-scrubbing is planned for v1.1.

Per-tool setup

Claude Code

The CLAUDE_TRANSCRIPT_PATH environment variable is set automatically by Claude's session-start hook (configured in hooks/hooks.json). No extra setup.

OpenCode

The bundled plugin at .opencode/plugins/wizard.js handles capture. Make sure the plugin is loaded in your opencode.json.

Cursor

The wrapper auto-detects the most-recently-modified transcript in ~/.cursor/projects/<workspace-id>/agent-transcripts/. If you have multiple Cursor windows open, capture grabs the most-recent — verify with the consent-moment header.

Codex CLI / Gemini CLI / Copilot CLI

Not yet shipped. For these tools, use /build-spell --from-transcript <path>.

Troubleshooting

"Couldn't capture chat from automatically"

Three causes, in order of likelihood:

1. Env var not set. Some tools require explicit env vars. See per-tool setup above.
2. Wrong session captured. If you have multiple sessions running in the same tool, the wrapper may grab the wrong one. Use /capture-this-chat --tool=<tool> to override detection, or use --from-transcript <path> with an explicit file.
3. Tool version unsupported. Tested versions are documented per wrapper. Newer versions may change session storage; the fallback is always --from-transcript.

"TOO-THIN: not enough structure to infer a skill"

Your session is fewer than 5 substantive turns or shows no repeated pattern. Either keep working in the session and try again later, or run the regular interview: /build-spell.

"TOO-BROAD: I see two tasks here"

Your session contains 2+ distinct tasks. Pick one with a follow-up message, or capture them separately.

Override flags

/capture-this-chat --last 20            # capture last 20 turns instead of 50
/capture-this-chat --last all           # capture entire session
/capture-this-chat --tool=cursor        # override detection