Hundrads Pfp Banner

Page identity assets: pfp + banner (Nano Banana)

A page that runs ads needs to look like a real brand. The two assets every profile click hits first: the profile picture and the cover/banner. This skill generates both with Nano Banana (Gemini's image model) via POST /v1/media/poster and hands the user finished PNG files.

Safety model. This skill only generates images. There is no endpoint

that sets a profile picture or cover photo anywhere — the user downloads

the files and uploads them in the platform's own UI. Don't ask for

platform credentials.

Setup

HUNDRADS_API_KEY env var. API base URL: https://hundrads.com. Endpoint reference: https://hundrads.com/llms.txt.

Step 1 — Quick intake

Ask only what you can't infer:

• Which brand / page? Brands come from GET /v1/accounts.
• Which platform(s)? Default Facebook page (pfp + cover). Same pfp

works for IG/Threads/X; banners differ per platform (see specs below).

• Text on the assets? Brand name on the pfp? Tagline on the banner?

Get the exact spelling — generated text must match character-for-character.

• Style direction, if the brief doesn't already carry one (colors,

mood, logo-like vs photo-like).

Then read the brand brief — it carries voice, colors, and positioning:

curl -s "https://hundrads.com/v1/brand/brief?brand=<brand>" \
  -H "Authorization: Bearer $HUNDRADS_API_KEY"

Scan the working folder for an existing logo / brand reference

Before writing any prompt, look for a real logo or brand asset already sitting in the folder this skill runs in — most brands have one. Don't invent a logo when one exists.

find . -maxdepth 3 \
  \( -iname '*logo*' -o -iname '*brand*' -o -iname '*icon*' -o -iname '*mark*' \
     -o -iname '*wordmark*' -o -iname '*favicon*' \) \
  \( -iname '*.png' -o -iname '*.jpg' -o -iname '*.jpeg' -o -iname '*.svg' -o -iname '*.webp' \) \
  -not -path '*/node_modules/*' 2>/dev/null

Also glance at the brand-assets/ output dir and any assets/, public/, static/, or img/ folders for unobviously-named brand imagery.

If a logo is found, pass it to the model as a reference image — the poster endpoint accepts up to 4 via the reference_images field (base64 PNG/JPEG). The model conditions on it, so the generated pfp/banner carries the real mark instead of an invented one. Base64-encode the file:

LOGO_B64=$(base64 -i path/to/logo.png | tr -d '\n')

(SVG isn't a raster format — rasterize to PNG first, e.g. rsvg-convert logo.svg -o /tmp/logo.png or magick logo.svg /tmp/logo.png, then base64 that.)

Still describe the logo in the prompt too (shape, hex colors, where it sits in the composition) — the reference image steers, the prompt directs. If the logo carries text you must reproduce exactly, quote it character-for-character.

If nothing is found, don't fabricate one — ask the user for the real logo, or generate logo-free and let them composite (see Etiquette).

Step 2 — Design for the crop, not the canvas

Platform display realities the prompt must account for:

Prompt construction:

• Pfp: flat, logo-like, high contrast, single subject or monogram,

solid/simple background from the brand palette. Avoid busy scenes — at 40px they turn to mud.

• Banner: wide composition, brand colors, tagline (if any) centered in

the safe band; explicitly tell the model to keep generous empty margin at top and bottom edges.

• Include exact brand colors (hex if known), the exact text in quotes and

its language, and "no watermark, no lorem ipsum, no extra text".

Step 3 — Generate

complexity: "simple" routes to Nano Banana (the Gemini image model in the registry) — cheap, good for clean graphic compositions. Every call is logged to the workspace Logs tab automatically.

Needs your own AI key (BYOK). Hundrads doesn't supply AI keys. Image

generation uses your workspace's stored provider key — complexity: "simple"

needs a Gemini key, "complex" an OpenAI key. Add it once at

https://hundrads.com/providers. Without it the call returns 400 (not

502): No <provider> API key configured for this workspace…. If you'd rather

not store a key, call the provider directly and skip Hundrads.

Image gen is slow — raise your Bash timeout. A single /v1/media/poster

call can take up to 5 minutes (model queue + generation). The default

Bash tool timeout (2 min) will kill the request mid-flight. Set the call's

timeout to 600000 (10 min) so it has room to finish, and don't treat a

long-running call as hung. Generating a pfp + cover = two calls, so budget

~10 min total.

curl -s -X POST "https://hundrads.com/v1/media/poster" \
  -H "Authorization: Bearer $HUNDRADS_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "prompt": "<full visual description, exact text in quotes>",
    "complexity": "simple",
    "aspect": "1:1",
    "include_b64": true
  }' > /tmp/pfp.json

python3 -c 'import json,base64; d=json.load(open("/tmp/pfp.json")); open("brand-assets/<brand>-pfp.png","wb").write(base64.b64decode(d["png_b64"]))'

If you found a brand logo (Step 1), include it so the asset carries the real mark. Build the payload with jq to embed the base64 cleanly:

LOGO_B64=$(base64 -i path/to/logo.png | tr -d '\n')
jq -n --arg p "<full visual description, exact text in quotes>" --arg logo "$LOGO_B64" \
  '{prompt:$p, complexity:"simple", aspect:"1:1", include_b64:true, reference_images:[$logo]}' \
  | curl -s -X POST "https://hundrads.com/v1/media/poster" \
      -H "Authorization: Bearer $HUNDRADS_API_KEY" -H "Content-Type: application/json" \
      -d @- > /tmp/pfp.json

Repeat with "aspect": "16:9" for the banner → brand-assets/<brand>-cover.png.

Rules:

• Never print `png_b64` to chat or read the JSON into context — it's

hundreds of KB. Always pipe the response to a file and decode from there.

• 400 `No <provider> API key configured`? The workspace has no stored

Gemini/OpenAI key — tell the user to add one at https://hundrads.com/providers, then retry. Don't treat it as a transient error or retry blindly.

• Leave brand empty — that field uploads to a Meta ad account

(image_hash for ad drafts), which is not what a pfp/cover needs.

• Generated text wrong/misspelled? Retry once with the text quoted and

emphasized in the prompt. Still wrong after 2 tries → switch that asset to "complexity": "complex" (the accurate-text model; costs more — say so), or generate it text-free and tell the user to overlay text themselves.

Step 4 — Deliver + iterate

1. Show the user both images (send the files, not base64). 2. Iterate on feedback — vary one thing per regeneration (color, subject, text placement) so the user can steer. 3. When approved, deliver the files with upload instructions:

• FB page: Page → Edit → profile picture / cover photo. FB will

circle-crop the pfp and crop the cover top/bottom — the safe-band design above means nothing important gets cut.

• Same pfp file works on IG, Threads, X. For non-FB banners (X 3:1,

LinkedIn 4:1), the 16:9 cover usually crops fine if the safe band held; offer a re-generate per platform if it didn't.

Etiquette

• Never claim anything was uploaded — the user does every upload.
• Never invent a logo for a brand that already has one. Scan the folder

(Step 1); if a logo exists, pass it as a reference_images input so the model reproduces it. If none is found, ask for the real one or generate logo-free and let the user composite.

• Pfp + cover should read as one identity: same palette, same mood.

Generate the pfp first, then describe it in the banner prompt.