git:20261012.42b420c to git:20261015.07749c7

- description: Review a BashCut edit before export as a capped loop — set the review profile from sourced ranges, measure, look in batches, judge findings by what a viewer perceives, fix one labelled edit per finding, prove each fix with review verify, and report what was and was not checked; an independent critic (or a self-audit from the packet) is always part of it, and the skill is also the critic's brief for a review packet. Use at the review stage, before a draft or export, or when asked to check a video. Triggers: "review", "kiểm tra video", "soát lỗi", "xem lại bản dựng", "chấm bản dựng", "critic", "QC", "trước khi xuất".

- Reply in the user's language. Core measures, renders evidence, tracks issue identity and rounds, and states what it

- did not check. Which findings matter and how many rounds to run are decided here, by you, for this video.

-

- Passing checks is not proof of a good video. One tool's edit had all 29 actions on the beat and every cut on a bar

- You are a fresh critic: you got a folder from `review packet --point strategy|draft|process` and this skill, nothing

- else. You did not make the edit and you have none of the maker's reasons, on purpose (T16 §4). `README` says which

- point you audit. Judge what is there against the brief and the plan, never against what the maker meant.

- | `strategy` (before the rough cut) | brief (with `inferred` fields), plan (options, `promise`, sections, `checks`), a story sheet | one message; the payoff answers the hook; fits the brief; each `inferred` field is a fair guess; the recipe checks that apply to a plan |

- | `draft` (before G5) | the files below + `plan.checks` | the generic checks and `plan.checks` as a viewer; the measured issues |

- **Generic checks** (every video, whatever the recipe): one message; `promise.payoff` answers `promise.hook`; text

- and speech in the opening (`hookSeconds`) say the same thing; understandable with the sound off; captions never repeat text that

- is on screen; nothing on screen the brief did not ask for: no credits, no "AI voice", "minh hoạ", "stock" or effect labels.

- | File | What it holds | Use it to |

- |---|---|---|

- | `README` | what the packet is, the project, revision and round | know which round you judge |

- | `plan.json` | the brief, the edit plan (sections with length ranges, shots, beats, decisions), the review profile, the outputs | know what the edit promised and for which platforms |

- | `digest.json` | what changed since the last review round | in round 2+, check those changes first |

- | `issues.json` | the measured issues with severity, frame, fix, and the round diff (fixed, new, persisting) | start from facts; a measured issue is a number, not yet a finding |

- | `shots.json` | every shot on Main with the cut into it (kind, framing before and after), rhythm, shares | find jump cuts, same-framing cuts, unintended gaps, monotony |

- | `word-landing.json` | words against cuts and titles: offsets, cuts inside words | find clipped words, titles early or late on their word |

- | `coverage.json` | the described shot each clip plays; the plan's beats against the words heard | find promised shots or lines that are missing (join the clips with `plan.json` shots) |

- | `measured.json` | what was measured and what was not (stale, notChecked, failed, unreliable) | never pass what nobody measured |

- | `sheet-*.png` | a contact sheet of every cut and title, labelled `<cell> <m:ss.s>` | see framing, text size and placement, repeated shots |

- How to work: read `README`, `plan.json` and `measured.json` first; then `issues.json`, `shots.json`,

- `word-landing.json`, `coverage.json`; then every sheet; judge the opening from the first shots, words and cells. Do not edit the project. If you also have the

- `bashcut` CLI, read-only commands are fine for a doubt (`ui frame F --phone`, `review window F`), nothing else.

- 1. A verdict: `pass` (no blocker), `changes` (blockers the maker can fix) or `fail` (the strategy or the cut is

- wrong). A check you could not judge is not a pass: name it under 5.

- 2. Findings, blockers first: time (m:ss.s) or frame, severity (below), what the viewer perceives, the evidence (file

- and field, or sheet cell), and which stage owns the fix (story, rough cut, rhythm, sound, text, colour, effects).

- 3. The 3–5 fixes to do first (T16 §7).

- 4. `needs_user`: taste choices and missing material, kept apart from defects.

- 5. What you could not judge: sound by ear, motion between sheet cells, anything `measured.json` lists as not measured.

- Record the verdict (it is a log entry, not an edit): `run append audit --point draft --verdict changes --findings 4

- --by critic`. Without the CLI, return it and the maker records exactly your verdict and count with `--by critic`.

- Read the checklist first, then the log and the changes. Flag: a required stage `skipped` or missing; `done` without

- evidence or with evidence that is not in the log; a skill marked unread for a stage that ran; a missing or `self`

- audit; a gate request without the user's answer; claims in the summary (`end`, gate summaries) that no entry or

- change backs ("fixed" without `review verify`, "exported" without an export). Return the verdict as above and the

- lessons, each as symptom → cause → what to do, for `bc:self-learn`.

- Define severity by the viewer at normal speed and size on a phone, not by the metric (T16 §4, reelmimic, talkcraft):

- - **blocker**: a viewer will notice it at normal speed and it hurts understanding or trust: a black or frozen

- frame, a clipped word, unreadable or covered text, speech drowned by music, a missing promised moment, a hook that

- does not show what the video is about. Core `error` issues (gaps, black picture on Main, a cut inside a word,

- missing glyphs, wrong length or shape for an output) are blockers unless you show they are not real.

- - **polish**: visible only when paused, zoomed or compared side by side. Fix when cheap; never a reason for another round.

- - **taste**: a choice a viewer could like either way (pace, a look, a transition). Leave it to the user (`needs_user`).

- - **needs user**: something only the user can supply or decide (a missing shot, music choice, a name to spell).

- A measured warning is not automatically a blocker: a locked-off interview trips "static shot"; a deliberate long

- title trips "long still". Decide, then fix, adjust the profile, or accept with a reason.

- Core has no editorial defaults. The project's `review` object holds the limits (`minShotSeconds`, `maxShotSeconds`,

- `maxStillSeconds`, `stillMotion`, `jumpCutChange`, `blackMinSeconds`, `hookSeconds`, `maxSilenceSeconds`,

- `maxMusicGapSeconds`, `voiceoverMarginSeconds`, `captionLineChars`, `captionMaxLines`, `minTextSize`,

- `minTextContrast`, `minSpeechCoverage`, `loudnessToleranceLU`, `severities`, `platform`). A limit left unset reports only the measured

- value, as info, and `review run --summary` lists it under `checks.unsetLimits` (only the limits a check of this

- timeline reads: no music, no `maxMusicGapSeconds`). `plan check` lists the same set before the build; with the

- quick path write them in `plan.cut.review`. A vlog recipe sets them; otherwise

- --merge` (or `plan.cut.review`; a `setProjectProperties` patch replaces the whole review object and drops the other

- limits), record each in the plan's `ranges` with its source and reason, and tell the user.

- | `maxStillSeconds` (exact freeze) | 0.8–1 s | a frozen picture on video media is usually a decode or render fault |

- | `maxShotSeconds` with `stillMotion` (static shot) | 3–6 s | genre: a calm daily vlog keeps long stills, a montage does not |

- | `maxSilenceSeconds` (dead air, whole mix) | 2–5 s | ASMR or a calm vlog tolerates more; a gap inside speech is another measure (~0.8 s) |

- | `minTextContrast` (text against the picture behind it, measured by `review measure`) | 3–4.5 : 1 | WCAG 2.x: 3:1 for large text such as titles, 4.5:1 for body-size text such as captions |

- The platform's own numbers (length, loudness target, safe zones) come from `platforms get <id>`, never from here.

- bashcut review measure # a job: picture checks, plugin checks

- bashcut jobs wait <job> --done --timeout 100 # returns when it ends (repeat on timedOut)

- 1. **Measure, then read.** `review run --summary` returns the issues and `checks`: `measured`, `stale` (an older

- revision: measure again), `notChecked` (with how to measure it), `failed` (plugin checks that timed out),

- `unreliable` (picture that barely changes: not a pass) and `unsetLimits`, plus `status`: `pass`, `fail` (errors)

- or `incomplete` (no errors, but `unsetLimits` or `notChecked` is not empty: the editorial checks did not run;

- `atExport`, loudness until the normalized draft, does not count).

- `incomplete` is never reported as passing: set the profile or say what was not checked. Errors come first. Info issues

- `voice-text-changed-*`, `captions-source-changed-*` and `beats-source-changed-*` mean a result was made from an

- older text or file: run their fix (speak, transcribe or detect again) before judging that voice, those captions

- or that beat grid.

- 2. **Look in batches.** `timeline sheet --cuts --text` (and `--outputs all` for safe zones), then

- `review window F --span S --step K` only at cuts in doubt: ±0.4–1.5 s at 10–12 fps, wider for dialogue cuts,

- at the width a viewer sees (360–420 px, T16 §3). `review coverage` (which described shot each clip plays) and `script check` against the plan.

- 3. **Critic, always.** Run `review packet --point draft` and give the folder to the `bc:critic` agent (or your own

- sub-agent with `<skill_dir>/CRITIC.md` as its brief): it has none of your reasons, so it sees what a viewer

- sees. No sub-agent: audit yourself by that brief from the packet only and record `--by self`. Either way record

- the verdict with `run append audit`; a skipped G5 needs it. The critic gives its fixes as one `plan.cut` patch:

- apply it (`plan apply --patch`) without re-reading the packet. A second look only for a blocker no measurement

- proves; `review verify` proves the rest.

- 4. **Decide each finding** by severity: fix it (one labelled edit; the issue's `fix` when it has one), change the

- profile when the limit is wrong for this video (say so and why), accept it with a reason, or list it for the user.

- 5. **Verify each fix.** `review verify <id>` re-measures the issue's range and returns before and after numbers, the

- status (fixed or persisting) and a strip; `review verify all` (or `ID1,ID2`) proves a whole round in one call,

- with a strip only for what persists. Say "fixed" only with that result: "looks better" does not count (T16 §2).

- After a fix, look again only at what changed (`timeline sheet --entries c2,t1`), not the whole timeline.

- Applying a `fix`: `timeline.apply` → write `fix.arguments.ops` to a file and run `timeline apply ops.json --base-rev

- N --label "<the issue title>"`; `timeline.close-gap` → `timeline close-gap --at-frame <atFrame> --base-rev N`;

- `review.measure` → measure, then review again; `export.start` → loudness comes from a normalized draft, last;

- `selects.place` → place it or unmark it with a reason (`bc:rough-cut`); a `hint` only → do it with that area's skill,

- or ask the user when it is taste. Never hide a failure with a transition or a flash (T16 §4).

- | jump cut, repeated framing, gap, cut in a word | `review window F`, `review shots` (`cut`), `timeline apply --dry-run` (`cutsInsideWord`) | `bc:rough-cut`, `bc:beat-cut` |

- | the picture pulses on the beat, too many flashes | `review sync --events impacts` (`summary.perMinute`, `minSpacingBeats`) | `bc:effects`, `bc:beat-cut` |

- | one sound effect repeated like a machine gun | `review sync --events sfx` (`sfxRuns`: `constantGain`, `constantSpacing`) | `bc:audio-mix` |

- | no hook, slow opening | `review shots --to F`, `review layout --to F --ink`, `transcript words --to F`, the first cells of the sheet | story (`bc:edit-workflow`), `bc:captions-text` |

- | dead air, music gaps, voice drowned | `review window F`, `audio mix-measure` (a job) | `bc:audio-mix` |

- | a `plan.visuals` row not built, off its word by more than 0.3 s, or over the face | `timeline get` (items at each row's `at`), `review window F`, `media subjects` | the row's building skill (`bc:visual-plan` §2) |

- - Cap: 1–3 rounds (T00 §3, T16 §3), never more than `workflow gates` › `maxReviewRounds` (the user's setting).

- - From round 2, run `review run --summary --since-rev R` (R = the revision of the previous round): it returns

- `diff {fixed, new, persisting}` and the round number. First confirm the previous fixes; then **only a new blocker

- fails the round**. New polish and taste findings are listed, not chased (T16 §2, reelmimic): rubric creep is how

- loops never end.

- - Stop early when a round fixes nothing. Whatever is left goes to the user with reasons, never silently shipped.

- A warning or info that is right for this video is kept on purpose:

- `review accept <id> --reason "locked-off interview, chosen" --base-rev N`. Later runs show it with the reason, leave

- it out of the counts, and the export report lists it. Accept an error only when no fix exists (the summary counts

- it as `acceptedErrors`) and tell the user; to change a whole check for this video set `review.severities.<check>` to

- `{"severity": "warning", "reason": "…"}`. `review accept <id> --remove --base-rev N` counts it again.

- Loudness and the delivered file are measured from a normalized export of this revision, so make it last, then

- review once more and request G5 (`bc:edit-workflow`):

- bashcut export status # also the delivered-file checks: stream starts, drift, fps, size

- Report the loop in a few lines, from `run log` and the last `review run --summary`: its `status`; the audit verdicts

- (and whether self-audited); rounds; what was fixed

- (issue → edit → `review verify` numbers); profile changes with the reason; every accepted issue with its reason;

- `needs_user`; and coverage: copy `checks.notChecked`, `failed` and `unreliable`, and say "not listened to" for sound

- you could not hear. Example: "Review: 2 rounds. Closed a 1.2 s black gap at 0:41 (verified: 0 black frames), raised

- the caption out of the Reels bar. Kept: the 9 s shot at 0:12 (locked-off interview). Not checked: the music under

- your voice by ear; the face-overlap check (no provider)." Log it: `run append measured --measured … --not-measured …`.