The problem
He wanted posts to come out of the work he was already doing. He also writes that Substack’s MCP tool couldn’t render images: they showed up as raw text.
The blueprint
- A hook fires at the end of every Claude Code session. It reads the transcript, tags tools, files and topics against a taxonomy, and appends to a daily log.
- When enough sessions pile up on one topic, the queue tells him he has material for a post.
- Sub-skills cover content generation, headline optimization, image generation, session capture and Slack recaps.
- A 458-entry knowledge corpus sits behind “Edge Copilot.”
- A custom publisher converts markdown into Substack’s ProseMirror JSON, uploads images to Substack’s CDN and PUTs the post to the API. Auth comes from a browser session cookie that a wrapper script refreshes.
Stated results
40+ scripts, 5 sub-skills, 458 wiki entries, $0.04 per headline research run. The per-post cost figure did not come through in retrieval: Unverified.
Caveats
- X handle Unverified. Substack: @jordancrawfordgtm.
- The publisher uses an undocumented API and browser cookies. This is his setup, not a recommendation.
Sources
Get one teardown a week.
Same format, a new operator every issue. Double opt-in.
SPONSORYour tool here. One line, one link.
Sponsor this slot