Brandon wrote a review with a robot, and every time he ships it he finds another mistake: a made-up fact, a boring robot phrase, a part of the show it forgot to talk about. So he built a checklist machine. You run it once and it checks all the mistakes at the same time and tells you exactly what to fix, before anyone else reads it. This page shows the machine and the five kinds of mistakes it now catches by itself.
The mistakes stop when each one becomes a gate. This runs every gate at once and tells you exactly what to fix, so the check catches the mistake instead of the reader.
It came out of one long correction cycle. A review shipped, got caught, got fixed, got caught again, over and over. Each catch was a different class of failure. Instead of fixing them one at a time forever, each became an executable check, and this command fires all of them before the next ship.
# run the whole battery on a review python3 ~/.claude/skills/review-update/review_update.py review.md [PASS] DESLOP SHAPE hits: 0 (density 0.00/1000w) | severity-3: 0 [PASS] REVIEW texture + fact F1/F2 + slop + coverage [PASS] BEATS beats present: 46/46 (100%) [PASS] TIGHT 4403 words REVIEW-UPDATE: PASS — all gates green, safe to ship
| Gate | Catches | The mistake it exists for |
|---|---|---|
| DESLOP | AI-tell shapes: quip-rider, name/lands as metaphor verbs, absurdist-attachment, essay filler | "a coping mechanism with a record light on" shipping green |
| REVIEW · fact | every concrete claim must appear verbatim in a cited source | "viewfinder" and "the ladders guy" asserted as fact |
| REVIEW · attribution | who-did-what taken from a recap, never guessed from an unlabeled transcript | the Dean guessed as the officiant when it was Todd |
| BEATS | real coverage against a canonical checklist built from the recap | token-recall said 40% "pass" while whole beats were missing |
| TIGHT | word band ~4300-4600, plus a thesis-circling detector | the review re-explaining its own thesis five times |
Six classes of mistake, six gates, one command. The result review passes all of them: the corrected read.
A claims.md (source paths plus claim/evidence pairs, every evidence line verbatim in a source) and a beats.tsv (one beat per line, keywords plus a description, built from the recap not from memory). Both live next to the review and are the ground truth the gates check against.