How AI recipe import works, and when to trust it

Cookfolio’s AI import runs the same four steps whatever the source: the file or link is checked, its content extracted, a model structures it into title, ingredients, steps and times, then the result opens in an editable review screen. Nothing is saved until you press save. The AI transcribes, it never invents: a cooking time absent from the source stays empty.

From your file to your saved recipe

Verification. The file type is read from the file’s own signature, not its extension, and matched against a closed whitelist: JPEG, PNG, WEBP, PDF, Word (DOCX), plain text (TXT). Caps are 10 MB for images, 20 MB for PDF and Word, 5 MB for text. A file renamed to look like a photo is rejected here.

Preparation. Photos are re-encoded, re-oriented and reduced to 2048 px on the longest side. A PDF goes through two channels, its text layer extracted and its first page rendered as an image, which is why a scanned PDF with no text layer can still work. A Word file gets raw text extraction only, so pictures inside a .docx are ignored, and TXT passes through as is.

Structuring. The model receives that content plus fixed instructions and returns a fixed shape: title, description, servings, prep and cook times, calories, ingredients split into quantity, unit and name, ordered steps with section titles, tags, and the position of a photo of the finished dish.

Review. Every field on the “Review recipe” screen is editable, and the recipe reaches your book only when you tap save. The AI proposes, you validate.

One pipeline, several front doors

Five entry points: photo, file, pasted text, link, and, on Pro, bulk import of a collection. The link alone branches several ways.

• A recipe site. The app first looks for the structured recipe data most cooking sites publish. Given a title, one ingredient and one step, the recipe is built straight from it, no AI involved. Only when that data is missing does the cleaned page go to the model, as importing from a website explains.
• YouTube. Analysed by image and by sound, not just the description, up to 30 minutes, public videos only: importing from YouTube.
• TikTok, Instagram, Facebook. The caption is read first, since ingredients and quantities are often typed there; the video is analysed only when the caption falls short, capped at 10 minutes: TikTok, Instagram and Facebook.
• Pinterest. No AI touches the pin: the app follows its outbound link to the source page and imports that like any web page, as importing from Pinterest describes. A pin with no outbound link fails with a clear message, not an invented recipe.

X (Twitter) and Threads are not supported: paste the text or a screenshot instead.

Transcribe, never invent

One instruction governs every extraction: the model is a transcriber, not a cook, and when in doubt it leaves a field empty. A prep time absent from the source stays empty instead of being estimated, and so does a missing serving count. Calories are never computed, only copied when the source states them. A “preheat the oven” step is never added, and no description is invented.

Units are normalised onto a closed list of 18 codes: g, kg, ml, cl, dl, l, tsp, tbsp, cup, fl oz, oz, lb, piece, pinch, clove, slice, sprig, handful. Non-English abbreviations map onto it: “c. à s.” and “EL” become tbsp, “cucchiaino” becomes tsp, a garlic “gousse” becomes clove. The unit is always separated out of the ingredient name.

The recipe is never translated: a German recipe imported inside a French interface stays in German. Translating means rewriting, and a rewritten quantity is one nobody double-checked. Better empty fields you fill in yourself than plausible values you would never think to question.

Where extraction stumbles

Faint or cramped handwriting. Photos are reduced to 2048 px, so pale pencil on squared paper loses the definition the model needs. Fill the frame, use flat light with no reflection, and get ingredients and steps both in shot.

Implicit quantities. “A glass of flour”, “a knob of butter”, “salt to taste” carry no number and no unit from the list, so the line is kept and the quantity stays empty. That is the rule working, not a defect.

Columns and tables. Ingredients in columns or a table can scramble the reading order and attach a quantity to the wrong ingredient: the most common problem with magazine scans, and the reason the ingredient list deserves the first read-through.

One import, one recipe. A page with two variants, a PDF of thirty recipes, a post stacking three: nothing gets split for you.

Silent videos. A montage set to music, without caption or on-screen text, gives the model nothing to transcribe: the app returns a clear “this does not look like a recipe”, not a fabrication.

Walled pages. Behind a login, a paywall or an anti-bot wall, nothing opens: paste the text. A file with a real text layer always beats a photograph of a screen.

What to reread first

1. Quantities and units, ingredient by ingredient. Mistakes cost most here, since changing the serving count multiplies every quantity. Watch for a unit slipped into the ingredient name, decimals, and empty quantities.
2. Servings. Everything scales from this number. If the source said “for 4”, check it says 4, and fill it in when it came back empty.
3. Step order. Check that two steps were not merged, and that section titles such as “the dough” landed on the right step.
4. Prep and cook times. Often empty, because the source never wrote them. Fill them if you know them.
5. The photo. When the source image showed the finished dish, the app crops it out and pre-fills it; you can recrop, replace or remove it.
6. Title and tags. Fastest to fix, and worth the ten seconds: searches lean on the tags.

Nothing is locked: description, calories, notes and the source URL stay editable. You can add or delete ingredients and steps, and define a custom unit when none of the 18 codes fits.

Quotas and failed imports

The free tier gives 5 recipes and 5 AI imports for life, all sources included: a lifetime allowance, not one that refills weekly or monthly. A blog page or a Pinterest pin consumes a credit exactly like a video, even though those paths often use no AI.

A failed import consumes nothing: the counter moves only on a successful extraction. Two guardrails remain: five seconds between imports and 100 a day, which normal use never meets.

Pro, at 4.99 euros a month or 39.99 a year, lifts the recipe and import limits and unlocks bulk import of a Paprika (.paprikarecipes) or Mela (.melarecipes) collection, up to 600 entries, photos included, in a file of 150 MB. Family, at 6.99 a month or 59.99 a year, covers up to 4 accounts sharing recipes, meal plan and shopping list in real time, each member keeping their own account and login. Two caveats: an internet connection is required, there is no offline mode; and these amounts are for web and Android, an iPhone purchase following the local App Store price.

Getting a clean result on the first try

One recipe per file. On a link, copy the direct URL of the recipe page, not a home page or a category listing, and check it opens without a login. On a social post, expand the truncated caption with the “more” control, and check the account and post are public. When pasting text, take the whole recipe, title, ingredients and steps included: plain text is enough, but avoid a whole web page with its menus and comments. If a video import fails, paste its description instead.

And if one recipe seems hopeless, spend a free import on it: a failed attempt costs nothing.