Hands-on Track · Build Log

Your Agent Build Log

This page is the journal for the whole hands-on spine. After each chapter, come back and fill the matching milestone. Filling it in is the acceptance check: if you can write it, you really did it; if you can't, that square isn't done yet. Fill all six, and what you export is a complete Agent build dossier.

Lit up 0 / 6 milestones. Content lives only in this device's browser — nothing is uploaded.

Know what you want it to do for you

Talks like a human, reliably

Actually gets work done

Better or worse, measurably

Runs long without losing memory

The process is reproducible

Six milestones · Come back after the matching Part

Know what you want it to do for you

LLM Fundamentals · One accept-able requirement + one plan to hold down hallucination

The four-line requirement for that job Which plan did you pick to hold down hallucination, and why

Acceptance: read the four lines to a non-expert — they can restate what the AI has to deliver.

Talks like a human, reliably

Harness Core · Same input, five runs in a row — every result usable

The few lines in the prompt that actually did the work Five-run score and why the fails happened

Two off-track runs out of five means you're not at M1 yet. Classify first (format drift / made-up content / length runaway), then fix the matching cause.

Actually gets work done

Practice · Wire the first real tool and run the full loop

First tool and its three-line description First clean run and first stuck loop

Stuck-loop notes are worth more than clean-run notes — they're the first tricky case in your M3 eval set.

Better or worse, measurably

Advanced · A 10-case eval set + one baseline score

Where the eval set lives, and how the cases are built Baseline score and the first before/after takeaway

If you can say “score went from X to Y, and the cost was which case regressed,” M3 is done.

Runs long without losing memory

Self-Improvement · Three endurance rules + a write-to-disk-then-read-back run

Your three endurance rules “Write to disk, then read back” — verification result

One acceptance check: a brand-new session can pick up the last job seamlessly.

The process is reproducible

Collaboration Methodology · Your first Rule + a spec someone else can follow

Your first Rule (and the fail that produced it) Who saw spec v1, and what feedback you got

After all six squares are filled, hit Export below — that Markdown is the physical artifact this course leaves in your hands.

Saved on this device