APIPod logoAPIPod

a set09 / 14

Ad variations

A matrix against a fixed base, changing exactly one axis per round.

apipod-ad-variationsSKILL.md on GitHub

When to use it, and when not to

when to use it

  • The user wants options to test rather than one final image.
  • One asset needs versions for different channels — feed 4:5, story 9:16, banner 16:9.
  • The base is approved and the question is which hook or angle wins.

not for

  • apipod-generate-image — There is no base yet — the first image has to exist and be approved.
  • apipod-ugc-ad — The ad needs a person holding the product, not a still-life variant.

procedure

6 skills

What the skill does, in order.

  1. 01

    Do not start from an unrendered prompt

    The base is one approved image plus one written angle. Generate the base first if it does not exist — a variant matrix over an unapproved base multiplies the wrong thing.

  2. 02

    Write the base spec in one sentence

    Subject, surface, light, colour grading, framing, and any on-image text quoted verbatim. This sentence goes into every cell unchanged.

  3. 03

    Settle one axis

    Angle (front / three-quarter / top-down), hook (the line of copy or none), crop (4:5, 9:16, 16:9), or context (studio / lifestyle / season). Exactly one per matrix.

  4. 04

    Say the cell count out loud

    Three variants × two crops is six paid generations. estimate_cost per cell, show the total, then proceed.

  5. 05

    Vary only that axis

    Everything from the base spec stays; the approved base goes in image_urls. Changing angle and hook and crop together produces images that cannot be compared — which defeats the point of a matrix.

  6. 06

    Grid the results and report the cost of testing

    One row per variant, one column per channel, each cell labelled with its axis value. Report total and per-variant cost, then say which looks strongest for the stated channel and why — in one sentence.

try it

Send it from here

example request

Use APIPod to make three variants of this approved packshot — front, three-quarter, top-down — for a feed test. Only the camera angle changes.

The buttons open a new chat with this prompt prefilled. If APIPod is not connected yet, the prompt tells your assistant to stop and help you connect first. In Claude Code or a terminal, copy the prompt instead.

Not set up yet? Wiring guides for Claude, ChatGPT, Codex and more →

call trace

  1. estimate_cost (3 cells)

    → $— · total stated up front

  2. generate_image (cell 2 · top-down ← image_urls[base])

    → base spec unchanged, one axis moved

  3. check_job (per cell)

    → 3 · completed · compared like for like

Illustrative: the tools this skill runs, in order, for the request on the left. Prices appear the way estimate_cost returns them live — this page itself quotes nothing.

what it will not do

The rules it holds itself to.

  • 01

    Do not silently fill a slot with a different model when one fails. Report the failure and its refund.

  • 02

    If you cannot see the images, say that instead of picking a winner.

  • 03

    Three to six cells is a matrix. Twenty is an invoice nobody approved.

Install once, then just ask.

Connect the MCP server to your client. The skills load with the plugin — there is nothing else to configure.