Back to Livepeer Agent

How-to Guide

Exact Fonts And Logos In AI Media

Use the Livepeer Agent `font` and `logo` skills to capture brand facts, reuse exact assets, generate only where it helps, and validate outputs before they are called production-ready.

1. Capture

Find exact SVG logos, typography hints, palette tokens, and capture limitations from public brand pages.

2. Compose

Generate clean visuals first, then apply deterministic text or logo overlays with HyperFrames/ffmpeg tools.

3. Assess

Score exact copy, logo reuse, video consistency, and brand readiness before shipping.

Quick Start

Exact words and exact brand marks should not be left to a generative model. Ask the agent to use the `font` and `logo` skills, capture the source facts, then finish with deterministic overlays.

Use the font and logo skills.
Capture the logo and typography from https://getjoggy.com/.
Create a product video, but overlay this headline deterministically:
The world's best Organic energy drink
Then assess the result and give me the scorecard.

Use The `font` Skill

  • Exact headlines, captions, legal copy, prices, and offers.
  • Website typography matching and headline lockups.
  • Video overlays rendered after generation so spelling stays exact.

Use The `logo` Skill

  • Capture and reuse SVG/transparent logo assets.
  • Generate logo variants only when the user asks for variants.
  • Place brand bugs, watermarks, and end cards deterministically.
  • Embed the exact logo onto a product with `logo_composite` (pixel-exact, never redrawn by a model).

What To Say To The Agent

You just talk to the agent in plain language — no commands or code. Capture the brand facts, save them, then ask for the visual with the exact text and logo locked on top.

"Look at getjoggy.com and capture the exact logo, the fonts, and the brand
colors. Remember them as my Joggy brand: keep headline casing exactly as
written, and always use the real captured logo instead of drawing one."
"Make a 9:16 product video, but don't let the model spell the words. Put
this exact headline on top so the spelling stays perfect:
The world's best Organic energy drink
Then double-check the text is exactly right."

To put the exact logo ON a product (no model redraw), just ask:

"Take my Joggy logo and place it on the front of the can, wrapped to
follow the curve. Use the real logo — don't redraw it."

Under the hood: logo_capture / font_capture grab the facts, a brand kit stores them, deterministic overlays (hyperframes-render) place exact text, logo_composite places the exact logo, and font_assess checks spelling. You never type any of those — the agent picks them from your words.

The Same Request, Anywhere

The exact same natural-language request works in Claude Desktop, theLivepeer Agent webapp chat, and the CLI chat:

"Capture the logo and font style from getjoggy.com, then make a launch
card with the exact headline: The world's best Organic energy drink.
Use the real captured logo as the brand mark, and keep the text exact."

Joggy Acceptance Workflow

The Joggy validation target is overall 4/5 or better, no dimension below 3/5, `logo_capture = 5/5`, and `typography_text_accuracy = 5/5` for this exact phrase:

The world's best Organic energy drink

  1. Capture the Joggy logo and font tokens from the public site.
  2. Create or update a Joggy BrandKit with palette, logo URL, font profile, and text rules.
  3. Use the exact SVG for production logo placement.
  4. Generate a clean product visual or product video without relying on model-rendered copy.
  5. Apply the headline as a deterministic overlay.
  6. Run the rubric and publish a scorecard with references, outputs, scores, costs, and limitations.

Example scorecard path: /launches/joggy-font-logo-scorecard.html

What To Say When Generation Drifts

The generated image is visually close, but the logo/text is not exact enough for production. I kept the concept image and switched the final text/logo to deterministic overlay so the deliverable is accurate.
The can shot is attractive but the package label drifted. For a 5/5 package match, train a product model (see the LoRA for marketers guide) and still finish with deterministic logo/text overlay — never rely on the model to spell your wordmark.