Skip to content

dongnguyenvie/bashcut-agent-kit / skills/color-grade

color-grade

Verified

Colour grade

Grade after picture lock, before captions (T12 §7). The loop: probe → measure → candidates → check → back off. Numbers are readings with their source, not targets: what to change and how much follows the footage and intent.

BashCut's colour tools

Whole video: library place LOOK --at-frame 0 --duration TOTAL_FRAMES. A range: adjustment add --look matte-cinematic --at-frame F --duration D (grades every layer below). One clip: its color. A LUT: luts import /abs/look.cube --name "Matte" --base-rev N (ID in timeline get luts). Built-in looks: original, vivid, muted-film, black-white, bright-airy, moody. Grade picture layers only; adjustments never touch text layers above them. White balance and black/white points need a generated LUT (grade.py).

1. Probe

media analyze (job), then media analysis --media ID: a tech.transferKind other than sdr (iPhone shoots HLG by default) needs handling first; the looks were fitted on Rec709. No tone-mapping control yet: tell the user which clips are HLG/PQ/log instead of grading over them (T12 §4, §5).

2. Measure

color measure, --by clip (outliers), --samples N for long or changing clips. Find the brightest, darkest and skin-heavy clips: check them later. Never judge from one frame (T12 §4).

ReadingRange seenWhat moves itSource
Mid luma, daylight vlog~30–50lower on purpose for night, low-key, silhouettes: not "too dark"T12 §3, §7
Clipped highlightsshare near 0deliberate sun or sky may clipT12 §7
Matte blacks / rolled whitesblack > 3 reads as matte; white < 90 as rolledexamples, not targetsstyle-study (T15 §3)
Saturation< 30 reads muted, > 45 punchygenre: food punchy, cinematic mutedstyle-study (T15 §3), T12 §3
Split toneshadow R−B < 0 with highlight R−B > 0 = teal-orange—style-study (T15 §3)
Shot matcheach clip's difference from the median clipyou choose how far to pull an outlierdrama-skills (T12 §3)

A recipe's intent (recipe.skill in project get), e.g. "bright, warm, skies kept": turn it into which readings matter, never into fixed numbers.

3. Candidates

Default: one sheet, one pick. Turn the intent into a target, then let the scripts rank, check and draw:

uv run "<skill_dir>/fetch_luts.py"      # once per Mac (setup, or a new --folder of your LUTs): prints one line
uv run "<skill_dir>/grade.py" candidates --target warm,film [/abs/clip.mp4 | MEDIA_ID ...]   # sheet path + 1 line each
uv run "<skill_dir>/grade.py" candidates --pick 3 --strength 0.7   # luts import + adjustment add over the edit

--target: tags (warm, cool, teal-orange, matte, film, bw, vivid, muted, punchy, soft, bright, dark, moody, clean, green), a look name below, or a bc:style-study profile JSON. It ranks the LUT library (log LUTs only with --log), tries the best on 3 frames (brightest, darkest, most skin), drops damage past --max-clip (default 1 % of detailed pixels; 0.5–2 seen), --max-crush (2 %; 1–3) and --max-skin (8°; 5–12), and adds --gen 3 looks below fitted to the footage. Look at the sheet once, pick a number (N@frame:duration for a range). Picks and skips, remembered per recipe (--for), move the ranking. Never read a LUT list or .cube.

LookUse forFitted reference (looks.json)
matte-cinematicdefault cinematic: street, daytime, travelblack ~7, white ~84, sat ~28; teal shadows, peach highlights, greens to olive, skin protected
warm-filmwarm indoor light, late afternoon, personal storiesdeep blacks (~2.6), highlights ~66–76, sat ~29, warm mids
faded-memoryhook, flashback or emotional low point, then back to colournear black-and-white, lifted blacks
teal-orange, clean-bright, moody-green, bw, film-printblockbuster/travel; food, product, tutorial; forest, rain, night; drama; narrativeset by hand, not fitted (grade.py looks)

Also library list --kind look.

  • Strength 0.5–1.0; 0.6–0.8 is typical (openmontage). Lower for log-like or already saturated footage, skin-heavy shots and mixed light (T12 §3).
  • One look for the video; fix outliers with per-clip color or a second adjustment over them (T12 §4).
  • A LUT is half the look, light when shooting the other half. Harsh noon sun stays harsh: say so instead of pushing the grade until skin breaks.
  • A matte look lifts pure black backgrounds to dark grey; use another look over those ranges if true black matters.

Generated looks by hand (grade.py, when the sheet has nothing close)

G="<skill_dir>/grade.py"           # the folder this SKILL.md is in; run with uv (installs numpy + Pillow)
uv run "$G" looks
uv run "$G" match /abs/clip.mp4 matte-cinematic --exposure S --wb R,G,B   # JSON report, writes nothing
uv run "$G" lut matte-cinematic --exposure S --wb R,G,B -o /abs/project/luts-src/matte.cube
uv run "$G" preview /abs/clip.mp4 /abs/project/luts-src/matte.cube -o /tmp/prev.jpg --n 6

match reports the footage through the look against the look's ref and range in looks.json (outside: 0 inside, − below, + above, null with no range; each range has its why); it corrects nothing. You decide the exposure (stops added to the look's) and white balance (R,G,B multipliers on the look's), try them with match again, then write the LUT with lut. Leave the white balance alone on scenes meant to be coloured (sunset, neon, screen recordings, graphics): the reference tint is the look's, not the scene's (T12 §4). match runs on the source file only; the real check is BashCut's composite in step 4. Never read a .cube body; judge a LUT by its frames (T12 §4). Your own look: a params JSON (keys in grade.py -h) in place of a look name. Then luts import the .cube and adjustment add --lut LUT_ID --lut-strength 0.8 over the edit.

4. Check

color measure --compare source, ui frames --compare graded --frames F1,F2,F3 (brightest, darkest, skin-heavy), ui frame F --phone.

Read the damage facts in change relative to the source, then look at the frames. One worked example (DaVinci Resolve's grade checker, thresholds are parameters): clippedGrowth 0.5 % of pixels, washed-out chromaRatio 0.60, milky blackLift L* 8, and a grade that moved meanDeltaE 6 without any flag "is worth a second opinion": undamaged is not the same as good (T12 §2, §3). A matte look lifts blacks on purpose; a warm look moves the mid tint on purpose: compare with the intent. Check every sampled clip, not the average; name the clip and frame that fails.

5. Back off

On damage, lower before adding: lutStrength first, then the exposure or white balance you passed. Resolve backs off ×0.8 per try down to 0.5 (T12 §2). Change the adjustment's color (send every key) or timeline undo and add it again. At most 3 attempts. Then stop: show the user ui frames --compare graded of the gentlest attempt with its numbers and ask (ui view --compare on toggles before/after in the viewer for them).

Keep what worked

  • library place ID (an adjustment) or library apply ID --item ITEM (one clip's grade) copies its LUT in.
  • Once the user approves: library save-selection --kind look --item ADJUSTMENT, or from a LUT: library add --kind look --file /abs/matte.cube. A correction: library update ID, or --as NEW_ID for a built-in or plugin look.
  • Matching a reference channel: measure several of its frames with grade.py measure --json (videos the user has the right to use), compare with your footage, and move a look's params toward them; see bc:style-study.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall