Written each week by your child's mentor.
Larry Pang
July 5, 2026
Quality week. We blind-tested Michael's drafts against the client's own copywriter on the same three news events — the marketing lead picked Threadline's draft twice out of three, and couldn't tell which was which on the third. The remaining gap is consistency at volume; the voice-match scorer will lock that down.
Michael should be proud of this stretch. Two weeks from now, his system is writing scheduled content for a real company, in public, under his name.
Larry Pang
June 28, 2026
A milestone worth celebrating: the first complete Threadline run end-to-end — news event in at 6:00am, a publish-ready announcement thread out by 6:04, in the client's voice, with the compliance rules respected (no price talk, nothing near financial advice). Michael presented it like an agency lead defending creative, and when I pushed on one off-brand line he traced it to the voice-sample selection and fixed it the same day.
We're now in the quality phase: making every draft something the client would ship unedited. Then the approval dashboard goes live.
Larry Pang
June 21, 2026
This week was about prompt engineering — the craft of getting an AI model to write like a specific brand instead of like an AI. Michael iterated through eleven versions of his voice prompt, keeping a lab notebook of what each change did. Version 11 produced a draft the client's marketing lead said she "would have posted without edits."
He's starting to think like a builder: not "does it write one good tweet" but "does it stay on-voice across fifty, unattended."
Larry Pang
June 14, 2026
Michael had a breakthrough week. The ingestion pipeline now pulls the client's full tweet history plus live crypto news feeds — real data, not a toy dataset. What impressed me most: when the X API rate-limited his backfill halfway through, he didn't wait for our session. He read the rate-limit docs, added batching with retries, and had it running clean by morning. That instinct — debug first, ask second — is exactly what we try to build.
Next week we start on the AI layer: teaching the model to write in the client's voice, not a generic one.
