Brand Setup Skill for Claude

Brand Setup

Run this once, before your first video. This repo ships with a house brand — a calm, premium

indigo look. It's a real, working brand, not a placeholder, and you're welcome to keep it. But if you

never touch it, every video you make here looks like the channel it came from. This skill fixes that

in one pass.

What this skill owns

Three files that must stay in sync, and this skill is the only thing that should write all three:

| File | What it holds | Who reads it |

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

| brand.md | the contract, in prose | every step-2+ skill (/make-tsx, /suggest-sfx, /packaging) |

| remotion/src/brand.ts | the same tokens as code — COLORS, GRADIENT, RADIUS, SHADOW, EASINGS, BRAND | every shot |

| remotion/src/fonts.ts | the font families, loaded from @remotion/google-fonts | every shot |

They drift silently. brand.md says indigo, brand.ts says teal, and nothing errors — the docs

just quietly stop describing the videos. That's why one skill writes all three in one pass.

Every shot imports from brand.ts / fonts.ts, so changing a value here re-brands **every video in

the repo, retroactively.** That's the whole point of a contract.

The flow

Stage 1 — Interview (ONE message, not twenty questions)

Ask it all at once, and say up front that "keep the default" is a valid answer to any of it —

most people care about two or three of these and have no opinion on the rest.

• Channel identity — the wordmark text, and what the channel is about in a sentence. The

wordmark renders in three parts with the middle part in the accent color

(['Learn','With','Hasan'] → the "With" is indigo). Two-part marks use an empty third

(['Acme','Labs','']). Also: the sign-off line for the end card.

• The feeling — calm/premium, loud/energetic, playful, editorial, technical? This drives motion

and shape more than color does.

• Palette — do they have brand colors already (paste hex), or should you propose from the

feeling? Light base or dark base? The accent is the single most important choice: it's the key

words, the highlights, the CTAs, the wordmark's middle.

• Type — existing brand fonts, or propose to fit the feeling? Must be Google Fonts (see the

hard gate in Stage 3).

• Delivery specs — what does your camera actually shoot? Resolution and frame rate. This is

not taste; it's the one value that must match reality (see the trap below).

• SFX taste — subtle accents under the voice, or louder/punchier?

Stage 2 — Propose before you write

Show the full proposal in the chat first — palette as a table (role · hex · what it's for), the

font names, the motion feel, the delivery specs. Get a yes.

Reading hex codes in a table is cheap. Rewriting three files and re-rendering is not. Do not skip to

writing because the answers "seem clear."

Stage 3 — Hard gate: the fonts must actually exist

Fonts are the #1 way this breaks the build. fonts.ts imports from @remotion/google-fonts, so

a family that isn't in that package fails to resolve. A local/purchased/system font will not work —

if they name one, tell them plainly and pick the closest Google family.

Verify before writing. Both checks, every font:

cd remotion

# 1. Does the family exist? (PascalCase, no spaces: "Space Grotesk" -> SpaceGrotesk)
ls node_modules/@remotion/google-fonts/dist/esm/ | grep -i "^Poppins\."

# 2. Does it HAVE the weights you're about to request?
grep -oE '"[0-9]{3}":' node_modules/@remotion/google-fonts/dist/esm/Poppins.mjs | sort -u

~1,800 families are available, so there is almost always a good match.

The weight trap is real, not theoretical. Space Grotesk stops at 700 — no 800/900. Oswald stops

at 700. Requesting a weight the family doesn't ship gives you a silent fallback or a broken render,

and you won't notice until a headline looks wrong. Check, then request only what exists.

Stage 4 — Contrast gate

Text has to be readable, and a beautiful palette that fails here produces unreadable videos. Compute

it, don't eyeball it (chroma-js is already a dependency):

cd remotion && node -e "
const chroma = require('chroma-js');
const C = { ink:'#1a1a2e', muted:'#6b6b7b', paper:'#fffef7', accent:'#6366F1' };
const p = [['ink','paper'],['muted','paper'],['accent','paper'],['paper','accent']];
for (const [a,b] of p) console.log(a.padEnd(7),'on',b.padEnd(7), chroma.contrast(C[a],C[b]).toFixed(2)+':1');
"

Judge each pair by how it's actually used, not one blanket number:

| Pair | Used for | Gate |

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

| ink on paper | body text, everywhere | ≥ 7:1 |

| muted on paper | captions, secondary labels | ≥ 4.5:1 |

| accent on paper | large key words only (34px+) | ≥ 3:1 |

| paper on accent | pill/chip text, small | ≥ 4:1 |

For reference, the house default measures 16.87 / 5.17 / 4.42 / 4.42. Note the accent sits at 4.42 —

comfortable for big words, tight for small ones. That's the normal shape of this tradeoff: a

saturated accent that pops on a light base rarely also clears 7:1. Fine. Just don't put small text on

it.

Below 3:1 on any pair, do not ship it — darken the shade and re-check. Between 3:1 and the gate,

say so out loud and let them choose. Never silently accept a palette that can't be read.

Stage 5 — Write all three files, in one pass

remotion/src/brand.ts — keep every exported name exactly as-is. 29 shots import COLORS,

EASINGS, GRADIENT, RADIUS, SHADOW, BRAND; renaming or dropping a role breaks all of them.

Change the values, never the keys. Fill every role — a palette that omits warn or danger

compiles and then explodes at render.

remotion/src/fonts.ts — swap the families and weights (verified in Stage 3). Keep

FONT_DISPLAY / FONT_BODY / FONT_MONO exported. FONT_SERIF exists only for the Claude Code

wordmark clone — it isn't a brand font, so leave it alone unless asked.

brand.md — rewrite §1–§7 and §9–§10 to match. Do not touch §8 (Asset & source locations) —

that's pipeline structure, not brand, and the other skills depend on it being accurate.

Keep brand.md's *shape*: the other skills read it by section, and /suggest-sfx reads §10 for its

SFX taste and function table specifically.

The delivery-specs trap (§7). Canvas + fps are the one place where taste is wrong and reality is

right: the canvas fps must match the master cut's frame rate. Shoot 1080p30 and it should say

1080p30. A mismatch here doesn't error — it desyncs every shot from the narration, and you'll blame

the timing before you blame the brand file. Ask what the camera shoots; never assume.

Stage 6 — Prove it (do not skip)

Render the brand back at them and read the frame yourself before showing it:

cd remotion && npm run gen
npx tsc --noEmit
npx remotion still src/index.ts BrandProof out/brand-proof.png --frame=95

BrandProof (remotion/src/shots/brand/BrandProof.tsx) is a utility shot, not a video beat. It

hardcodes nothing — it reads brand.ts and fonts.ts live, so what renders is genuinely what the

shots will use. It shows the wordmark, the full palette with hex labels, the gradient, all three

fonts, the card depth, and the accent-legibility line.

Look at it and check: **is the wordmark right? Is the accent word legible? Do the fonts look like

what you proposed?** Then show them the image.

Finally, render one real example shot (EndCard is the fastest) to confirm nothing broke downstream.

Hard rules

• All three files or none. Never update brand.ts without brand.md. Silent drift is the

failure mode this skill exists to prevent.

• Change values, never keys. Every shot imports the token names. Renaming a color role breaks 29

files at once.

• Fonts must be verified in @remotion/google-fonts before writing — family AND weights.
• The palette must clear the contrast gate, judged by usage. Below 3:1 is not shippable.
• Delivery fps must match the camera, not the brand's preference.
• Don't touch brand.md §8. It's structure, not style.
• Never re-tune a thumbnail to match the brand. Thumbnails are deliberately louder — they live on

the browse wall and play by CTR rules. See /packaging. Two systems, on purpose.

• The house default is a real brand, not a placeholder. "Keep it" is a legitimate answer; the only

thing that's genuinely wrong is shipping someone else's *wordmark* on your channel. If they keep

everything else, still set BRAND.wordmark.