Every weekly note on this site is the exhaust of a system. This page is the system, written down. It exists for two audiences: readers who want to know how these notes get made, and the AI copilot I work with — when a future session needs to know what the plan is, it gets pointed at this URL.
The premise
I’m learning AI agent engineering in public, and the method is Feynman’s: you don’t know a thing until you can explain it with the source material closed. So every work session ends with a blind restatement — materials shut, ten minutes, my own words. The stumbles in that restatement are, by definition, the parts I haven’t learned. They become the opening agenda of the next session.
One principle above everything: learning is the product; content is exhaust. If producing videos or articles ever starts eating the learning itself, the content loses. The rules below exist to make that trade impossible.
Two work lines, nothing else
Main line — the SWE-bench harness. Two to three sessions a week, 60–120 minutes each, continuing the work from the first field note: variance reduction first, then the nine unresolved tasks. This is the primary recording material.
Sideline — Karpathy’s Zero to Hero. At most one session a week, hard cap. It fills in model internals while the main line builds harness engineering. Skipping a sideline week is normal; exceeding the cap is a rule violation.
No third track. TryHackMe, CodeCrafters, kernel courses — all explicitly parked. New holes get dug next quarter, if at all.
The session ritual
Every session, both lines, no exceptions:
- Start the screen recording (OBS, screen + mic, no camera).
- Work 60–120 minutes.
- Stop. Start a second recording. Blind restatement, ~10 minutes, answering three fixed questions: What did I get done? Where did I get stuck, and how did I get around it? Explain one concept from today in my own words.
- Stop. Five-minute wrap-up: the restatement’s key points go into the progress log’s “what I learned” field — a field only I am allowed to write.
One action, three outputs
That single ritual feeds everything:
- Video (Chinese, Bilibili, weekly). The work session plus the restatement, uploaded as a two-part video. Titles carry real progress numbers.
- Progress log. The permanent, auditable record. Every number that later appears in a video or article must exist here first.
- Article (English, here, weekly). A Sunday digest built from the week’s “what I learned” entries. Not a translation of the video — same source, different audience.
The pipeline, and who does what
The division of labor is a hard boundary. I record and restate. My AI copilot does everything after the recording stops: an automated cut (silent stretches play at 8× — this is a typing-mostly, talking-sometimes format), Whisper transcription of the restatement into burned-in subtitles, the upload, and the first draft of the weekly article.
What the copilot may never do: write the “what I learned” entries, or originate a claim or number. Drafting structure and prose from my log is fine; the substance stays mine. That’s not a sentimental rule — the entire value of an auditor writing field notes is that the scoreboard is real.
Six hard rules
- Human editing time ≤ 30 minutes per video. Automated processing is unlimited. Transitions, effects, and polished covers are banned no matter who makes them.
- No restatement → the session doesn’t count and must not be published.
- The “what I learned” field is written by me only.
- Skip protocol: in an overloaded week, the minimum publishable unit is the ten-minute restatement clip plus a three-sentence post. A fully missed week is skipped, never back-filled.
- Numbers must be auditable: any figure in a video or article must exist in the progress log first.
- Sideline cap: one Zero to Hero session per week. No third track.
The scoreboard
Four-week checkpoint (early August 2026): at least 8 recorded sessions, 4 published videos, 3 articles here, and measurable SWE-bench progress — either the variance-reduction mechanism landed or at least two previously-unresolved tasks flipped green. Hit it: continue. Miss it: the sideline dies, the main line and the ritual survive.
That last clause is the design philosophy in miniature. When something has to give, it’s never the learning and never the honesty — it’s the volume.