image-ad-clone
Skill1.6k starsChanged 8 months agoMITView on GitHub
Use when the user wants to reverse-engineer an existing image ad into a reusable prompt template. Validates via Arcads — picks gpt-image-2 or Nano Banana at Phase 1. Triggers on "clone this ad as a template", "reverse engineer this ad", "turn this ad into a prompt", "extract a template", "make this ad reusable", "add to my prompt library", "study this ad and make a template". Input is an EXISTING ad image; does NOT trigger for fresh generation (use chatgpt-image-ad or nano-banana-image-ad).
- Reads credentials
1 other project uses a version of this.
Adopted by
Most recently taken up by
- krusemediallc/arcads-claude-codeword for word, first seen Apr 2, 2026
Other versions1 project uses a version of this.
Word for word · 1
- krusemediallc/arcads-claude-code★ 1.6k100% same
Edited · 0
None found.
Wrote their own version · 0
None found.
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository8 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- shared/CLAUDE.mdCLAUDE.md
- .cursor/rules/project-context.mdcCursor rule
- arcads-external-apiSkill
- chatgpt-image-adSkill
- generate-youtube-thumbnailSkill
- nano-banana-image-adSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 4 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 23
Backend is one of: ChatGPT Image 2 OR Nano Banana on Arcads: Never uni-1. The script choice happens in Phase 1 once the user picks (or the …
- Distinctive: line 42
If the user explicitly says "clone this with gpt-image-2" or "with Nano Banana", honor that.
- Distinctive: line 59
When the shared guide Phase 1 tells you to locate the companion generator, look here in order based on the model choice:
- Very specific: line 64
If neither: stop and ask the user to install chatgpt-image-ad first.
- Very specific: line 79
Document the ratio fallback in the template's Aspect ratio: field so future users know they're rendering at a mapped ratio, not the …
- Very specific: line 92
Phase 8: Cross-model validation (recommended): Run the same template through the OTHER backend in this repo.
- Very specific: line 93
Phase 9: Document the template: Use the format in template-format.md.
- Distinctive: line 116
If you only validated against one model in Phase 8, say so explicitly:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
image-ad-clone (Arcads)
Take an existing image ad and turn it into a reusable, parameterizable prompt template that gets appended to the shared 37-template image-ad library. The template is validated by round-tripping through one of the Arcads image-ad generators — ChatGPT Image 2 (typography / UI-mimicry templates) or Nano Banana (photoreal / lifestyle / multi-reference templates).
This skill replaces the older Uni1-locked image-ad-clone (which only worked with Luma uni-1). It's backend-agnostic: at Phase 1 the agent asks you (or auto-detects from the reference) whether to validate against gpt-image-2 or Nano Banana, then routes through the matching generator script in this repo.
Read order
- This file — Arcads-specific generator paths, model-choice decision, what's locked at the per-repo layer.
- shared/skills/image-ad-clone/prompting/guide.md — the full model-agnostic 10-phase workflow (visual analysis → draft prompt → generate-with-reference → iterate → generalize → test → cross-model validate → document → save).
- shared/skills/image-ad-prompting/prompting/template-format.md — entry skeleton.
- shared/skills/image-ad-prompting/prompting/prompt-library.md — destination for the new entry. 37 validated templates already there; new entries go at T40+.
Hard rules
Inherits all 6 hard rules from the shared guide (strip platform chrome, validate by generating, test the generalized version, no brand-specific text in the final template, never silently overwrite, document model notes for both backends). Plus per-repo:
- Backend is one of: ChatGPT Image 2 OR Nano Banana on Arcads. Never uni-1. The script choice happens in Phase 1 once the user picks (or the agent auto-detects).
Picking the right backend in Phase 1
Pick by what the reference ad is showing — most templates fall into one clear bucket.
Use chatgpt-image-ad (gpt-image-2) when the reference is:
- Typography-heavy / UI mimicry (Apple Notes lists, fake Google search, fake Slack threads, ChatGPT-conversation ads, iMessage screenshots, comparison tables, fake AirDrop dialogs, Hinge-style cards, calendar UI, weather forecast UI, magazine masthead)
- Brutalist / editorial typography heros (huge type makes the joke)
- Dense small text inside UI elements
Use nano-banana-image-ad (Nano Banana family) when the reference is:
- Photoreal handheld objects (whiteboards, napkins, sticky notes, letter boards, scratch-off tickets)
- Aspirational lifestyle photography (sunset, kitchen at golden hour, OOH / transit)
- Multi-image reference blending (logo + product + style + character all in one)
- Clay / claymation / Pixar-adjacent textures
If the reference straddles both (e.g. a UGC-style photo with rendered text overlays), the safer default is to clone twice — once per backend — and ship the template with Model notes saying which renders cleaner. The agent will offer this in Phase 8.
If the user explicitly says "clone this with gpt-image-2" or "with Nano Banana", honor that.
Dependencies
This skill uses the matching generator script in the SAME repo:
- For gpt-image-2 validation:
skills/chatgpt-image-ad/scripts/generate_image.py(locked tomodel: gpt-image-2) - For Nano Banana validation:
skills/nano-banana-image-ad/scripts/generate_image.py(nano-banana-2default;--model nano-banana-proornano-banana-editopt-in)
Fail Phase 1 with a fix-it message if neither generator is installed in this repo.
Also required:
.envwithARCADS_BASIC_AUTHorARCADS_API_KEY- (Optional)
PRODUCT_IDin.env— if not set, the generator auto-fetches the first Arcads product - Python 3.12+
Where this skill's generator lives
When the shared guide Phase 1 tells you to locate the companion generator, look here in order based on the model choice:
For gpt-image-2:
~/.claude/skills/chatgpt-image-ad/scripts/generate_image.py<repo>/skills/chatgpt-image-ad/scripts/generate_image.py- If neither: stop and ask the user to install
chatgpt-image-adfirst.
For Nano Banana:
~/.claude/skills/nano-banana-image-ad/scripts/generate_image.py<repo>/skills/nano-banana-image-ad/scripts/generate_image.py- If neither: stop and ask the user to install
nano-banana-image-adfirst.
Aspect ratio mapping
The Arcads image endpoint (/v2/images/generate) accepts only 1:1, 16:9, 9:16 — regardless of which model (gpt-image-2 or nano-banana) you're hitting. When measuring the original ad's aspect (Phase 2):
4:5,2:3,5:4ads → render at1:1and post-crop in your downstream ad-builder skill1.91:1ads → render at16:9and post-crop9:16ads → native, no change
Document the ratio fallback in the template's Aspect ratio: field so future users know they're rendering at a mapped ratio, not the original.
(The KIE per-API repo's image-ad-clone skill supports a broader native ratio set — 4:5, 2:3, 3:2, etc. — because KIE's /jobs/createTask Nano Banana endpoint accepts them. If aspect-ratio fidelity matters more than Arcads-specific control, consider the KIE repo for that template.)
Workflow phases (model-agnostic, see shared guide for full detail)
- Phase 1: Preflight + model choice. Reference image file resolves;
.envhas Arcads creds; both generators detected. Ask the user which model to validate against (or auto-detect from the reference's typography-vs-photo balance). - Phase 2: Visual analysis. Describe the reference structurally — aspect ratio, format type, layout, typography, color palette, photography style, every text string verbatim, decorative elements, chrome to strip.
- Phase 3: Draft v1 prompt (brand-specifics intact).
- Phase 4: Generate with reference using the matching generator script. Pass
--image-ref <reference_path>and the matched aspect ratio. - Phase 5: Compare and iterate. Refine prompt based on deltas. Cap 4 iterations.
- Phase 6: Generalize into placeholders (
{brand.name},{brand.color_primary}, etc.). - Phase 7: Test the generalized template against a DIFFERENT brand. Regenerate. If structure breaks, refine placeholder set.
- Phase 8: Cross-model validation (recommended). Run the same template through the OTHER backend in this repo. Document deltas in the
Model notesblock. - Phase 9: Document the template. Use the format in
template-format.md. Required fields: tag, when-to-use, aspect ratio, reference image guidance, variable schema, template prompt (full validated), example fill, Model notes for both backends, validated example path. - Phase 10: Save and confirm. Append to
shared/skills/image-ad-prompting/prompting/prompt-library.md. Print path. Move PNGs to permanent iteration dir.
Iteration directory layout
<cwd>/iterations/clone-2026-05-26/
T40-lifestyle-hero/
prompt.txt
v1.png, v2.png, … # against the source ref (chosen backend)
test-fill-v1.png, … # Phase 7 generalization test against a different brand
cross-{other-backend}/v1.png # Phase 8 cross-model validation (optional)
notes.mdCommon Model notes patterns to write in Phase 9
**Model notes:**
- **gpt-image-2:** {observed behavior — e.g. "clean — strong on UI mimicry and table text", "tends to add a 4th Slack message — keep prompt explicit about exactly N", "small chart axis labels blur — bump font size feel"}
- **nano-banana:** {observed behavior — e.g. "strong — preferred backend for handheld board photos", "character identity drifts across variants on -2; use -pro to lock", "weak on dense table text — keep rows to 4 max"}If you only validated against one model in Phase 8, say so explicitly:
**Model notes:**
- **gpt-image-2:** validated clean (see iteration path)
- **nano-banana:** untested — validate before using on nano-banana-image-ad backendThe diff between a uni-1-era library and this one is this Model notes block. Don't skip it — it's the difference between a portable template and one nobody knows how to use.
See also
- shared/skills/image-ad-clone/prompting/guide.md — full 10-phase workflow
- shared/skills/image-ad-prompting/OVERVIEW.md — ecosystem context
- chatgpt-image-ad skill — gpt-image-2 generator (used for validation when you pick that backend)
- nano-banana-image-ad skill — Nano Banana generator (used for validation when you pick that backend)
- arcads-external-api skill — Arcads conventions (session folder, credit cost confirmation, presigned-upload flow)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

