The Body
Part 3 Quiz · The Body
Five questions on isolation, taught moves, hands, and honest grading. The bar for moving on is 4 of 5.
Five questions on isolation, taught moves, hands, and honest grading. The bar for moving on is 4 of 5.
Question 1
Two makers wrote into one repo all day, without a worktree, and never collided. What made that safe — and what one change would have made a worktree mandatory (Step 8)?
Show answer
Their files were owned disjointly, enforced by rule — neither maker could touch the other's paths, so a shared tree had nothing to collide over. The instant both need the same file, path ownership can't slice it in two: worktree, merged only when green.
Question 2
Your loop prompt has swollen to 600 words of blow-by-blow procedure. Name the fix and its two payoffs (Step 9).
Show answer
Move the procedure into a skill (a SKILL.md with a when-to-use description); the prompt
falls back to intent. Payoffs: consistency (every beat runs the same recorded move) and
maintainability (correct the procedure once, and every later beat inherits the fix).
Question 3
Recite the three connector rules, and for each, the disaster it heads off (Step 10).
Show answer
Few, focused tools — stops a confused beat from grabbing a delete_* that never should
have been exposed. Idempotent writes — stops retries and double-fires from posting
twice. Actionable errors — stops 3 am stalls on opaque failures the loop can't route
around.
Question 4
A fused maker-checker "self-reviews before committing" and has passed 100% for a month. Why is that statistic a red flag rather than a gold star (Step 11)?
Show answer
A self-grader only ever measures how much it agrees with itself: the blind spot that ships a
bug is the same one approving the review. 100% means defects are being co-signed, not caught.
Health actually looks like a separate checker now and then logging items_found > 0, actions_taken: 0.
Question 5
Rank these checkers by cost, and say when the cheapest is also the best: a read-only LLM with a rubric · a human reviewer · a script.
Show answer
Script < read-only LLM < human. The script wins outright whenever the check is a fact (links resolve, tests pass, schema validates) — free, opinionless, unarguable. Keep the LLM for judgment calls and the human for whatever truly needs someone accountable.
Next: Part 4 · The Spine · flashcards
Sources: this quiz tests Part 3, drawn from Panaversity's Loop Engineering: A Crash Course (S1) and Agentic Coding Crash Course (S2). Full attribution: resources/sources.md.