Recipe Pin Image Stack
Enter hub password to continue.
That's not right. Try again.
How to make recipe pins with crisp text overlays, the 2026 playbook
You showed me two pins: a Keto Cheese Crisps pin (one food photo, then a title, an ingredient list, and little nutrition badges laid on top) and a Sweet Potato & Chicken Bowls pin (two stacked photos with a bold text band between them). Both are the same recipe: a clean food photo plus sharp, correctly-spelled text on top. That's exactly the approach you picked, and it's also what the research says is right.
The one-line answer: generate the food photo with an AI image tool (or use a real photo), then add the title, ingredients and badges as a separate text layer using real fonts, so the words are always spelled right and look designed. Do not ask one AI model to draw the photo and write all the text; that's where you get gibberish and misspellings, especially on a dense ingredient list.
1. The base food photo
For the photo underneath, you have three good routes, and one of them is already free on your machine:
| Tool | Cost | Best for | Note |
|---|---|---|---|
| Your local Qwen-Image | Free | The everyday food shot | Food has no faces, so your never-use-Qwen-for-people rule doesn't apply here. This is your cheapest, most autonomous option for the dish photo. |
| Ideogram 4.0 (your recipe-photo skill) | Paid API / plan | When you want the photo and some baked-in text | Top open-weight model for readable in-image text (0.97 English OCR, best in its class). You already have a recipe-photo skill wired to it. It currently needs a quick re-authorization. |
| Nano Banana Pro (Google/Gemini) | Paid API | Richest scenes, densest text | Roughly tied with Ideogram on short headlines, slightly ahead on busy captions and scene realism. API only. |
Confidence: HIGH on the model ranking, sourced to Ideogram's own spec page plus hands-on comparisons at Segmind and Imagine.art. Two popular claims got killed here: "GPT Image 2 hits 99% text accuracy" (refuted), and the idea that Midjourney/DALL-E render usable text (they produce misspelled gibberish). AI model rankings move monthly, so re-check before hardcoding one.
Takeaway: use whichever gives the prettiest dish. Don't lean on it for the title and ingredient text. That's the next layer's job.
2. The text-overlay layer, three tiers
This is where the words go on. Pick the tier that matches how hands-on you want to be.
| Tier | Tool | Cost | What you do |
|---|---|---|---|
| No-code | Canva Bulk Create | Free / Pro | Design one pin template, connect a spreadsheet of recipes (one row per pin), Canva spits out one pin per row. A month of pins is doable in an afternoon once the template exists. |
| Managed API | Bannerbear + Make.com | Paid | Design a template once, then a link with your photo and title in it renders a finished pin. Make.com has a ready-made recipe that reads a table, builds the pin, and posts it to Pinterest, all unattended. |
| Fully coded | HTML/CSS template → PNG | Free | Build the pin as an editable HTML/CSS layout (fonts, badges, exactly like your examples) and render it to a 1000x1500 image in a loop. This is the one a Claude Code skill automates end to end. Details in the pick below. |
Confidence: HIGH on all three mechanics (Canva's own docs, Bannerbear's own guide, the live Make.com template, and Vercel's Satori repo). One important limitation to know: Canva Bulk Create can swap the text per row, but it can not reliably swap a different photo per row (that claim was refuted 0-3). So if every pin needs its own AI food photo, Canva means placing photos by hand, and the automated photo-swapping belongs in Bannerbear or the coded route.
3. The pick: your repeatable recipe-pin stack
AI food photo → HTML/CSS pin template → render to a 1000x1500 PNG, batched from your recipe data, with Claude drafting titles you approve.
- Photo: Qwen-Image (free) for the dish, or your Ideogram recipe-photo skill when you want more polish.
- Template: one HTML/CSS pin layout, designed once, holding the title, ingredient list, badges, and your brand look. Editable forever, no design tool needed.
- Render: Playwright (already on your machine, Python) loads the template with each recipe's data and screenshots it at exactly 1000x1500. Real browser fonts, so the text is always crisp and spelled right.
- Copy: Claude drafts up to five title options (10 words or fewer) and a few short call-to-action lines. You approve them before they get placed. This human-approval gate is the one part every good creator workflow keeps.
- Batch: feed it a folder or sheet of recipes, get a folder of finished pins.
Why this one for you: it stays in Python (your stack), the design lives in HTML/CSS you can tweak like any of your sites, the text is browser-rendered so it never misspells, and the whole loop is scriptable as a Claude Code skill. Canva Bulk Create is the no-code fallback if you'd rather design by hand; Bannerbear is the paid middle ground if you want it posting to Pinterest on its own.
The coded routes the panel confirmed 3-0: Satori (JSX plus a font file to pixel-exact text, the engine behind Vercel's social images) and the Node sharp library. Both are excellent but Node-based; Playwright keeps it in your Python world with the same crisp result, which is why it's the recommendation. If we ever want the lighter, no-browser option, Satori is the fallback.
4. Pinterest rules the template must hit
- Size: 1000 x 1500 pixels, a strict 2:3 vertical. This is Pinterest's own recommended standard and the shape the feed favors; other ratios get cropped. (Any 2:3 multiple like 1200 x 1800 also works.)
- Legibility: pins show at roughly 236 pixels wide in the feed, about a quarter of full size, so the title has to survive heavy shrinking. Big, bold headline; test by zooming your pin out to 25%.
- Your examples already do this right: huge title, short word count, high contrast, badges for the skimmer. Keep that.
Confidence: HIGH on size and legibility (Pinterest's own best-practices page, echoed by Tailwind and others). One thing I'm not repeating as gospel: the tidy "4 to 8 words at 60px, only 3 to 4 colors" CTR rule got refuted (1-2). It's reasonable design advice, just not a verified performance law, so treat it as taste, not a rule.
5. Is there a ready-made Claude skill for this? Not yet.
There's a documented creator workflow that pairs Claude with Canva: Claude drafts pin titles and call-to-action lines, the human approves them, and they get rotated across template pages. That part is real (bypwdesigns). But the claims that it's a packaged, installable Claude Skill, and that Claude edits the Canva file itself through a connector, both got refuted. So nobody's shipping a drop-in skill for this. It has to be built, which is exactly what the pick above is: a small Claude Code skill we own.
6. Honest caveats
- AI image rankings change fast. Ideogram 4.0 and Nano Banana Pro are the 2026 leaders on text, but re-verify before we lock a model into a skill.
- The "make a month of pins in 90 minutes" figure is one creator's best-case, self-timed run that skips setup and keyword research. Real first-time setup is longer; the speed comes on repeat runs.
- Whether Pinterest lets a pipeline auto-post pins (versus needing a scheduler like Tailwind, or you posting) is worth confirming before we automate the publishing end. The image-making end is fully solvable today.
- Killed and not reused anywhere above: GPT Image 2's "99% text accuracy," the Canva image-URL photo swap, the "Puppeteer is the one dominant method" claim, and the 4-to-8-words CTR rule.
Want me to build it?
Say the word and I'll stand up the recipe-pin skill: an HTML/CSS pin template in your brand look, a Playwright renderer that outputs 1000x1500 PNGs, and the recipe-data-to-batch loop, wired so Claude drafts titles you approve. First target could be ProteinFirstRecipes or SmartStrongAlive. I'd hand you a few sample pins to react to before we run a batch.