Draft:Recipe markup language

• Comment: In preparing this draft, I disclose that AI assistance was used as follows: https://claude.ai/code/session_01RCAae2kx9Y1iPGqhtVEeCw Adubovskoy (talk) 21:26, 1 October 2026 (UTC)

A recipe markup language is a machine-readable notation or file format for cooking recipes. Such formats encode the parts of a recipe that software can act on, such as ingredients, quantities, units, cookware, timings, yield and nutrition, so that recipes can be searched, scaled, exchanged between programs, converted between measurement systems and used to generate shopping lists. They range from fixed-column plain-text layouts used by 1990s desktop software, through XML schemas of the 2000s and the structured-data vocabularies embedded in web pages, to lightweight plain-text languages of the 2010s and 2020s.

The earliest documented notations appeared in 1985: David A. Mundie's RxOL, a postfix formal language for recipes, and Brian Reid's USENET Cookbook, which distributed recipes in a rigid troff macro format designed for later database processing. During the 1990s, text export formats of recipe database programs, in particular Meal-Master's "MMMMM" layout and MasterCook's MXP, became the common currency of recipe exchange on bulletin board systems, online services and Usenet. Around 2000 several XML schemas were proposed, including RecipeML, CookML and REML, none of which gained wide adoption.

From 2008 the microformat hRecipe and, from 2011, the schema.org Recipe type embedded recipe data directly in web pages. Google's recipe rich snippets (2010) and Recipe View (2011) made this markup commercially important for recipe publishers, and by 2014 schema.org Recipe appeared on more than 1.6 million web pages in the Common Crawl corpus. Schema.org JSON-LD became the de facto web interchange format, consumed by search engines, Pinterest, recipe-import applications and scrapers.

In parallel, developers created plain-text formats intended to be written by hand and kept in version control, including Open Recipe Format (2013), Pesto (2015), RecipeMD (2018) and Cooklang (2021). Academic work in natural language processing and knowledge representation has separately modelled recipes as flow graphs, dependency trees, formal programs and ontologies.

Background

Written recipes acquired their modern shape in the nineteenth century. Eliza Acton's Modern Cookery for Private Families (1845) is credited with introducing the practice of listing ingredients and cooking times with each recipe, and Fannie Farmer's The Boston Cooking-School Cook Book (1896) was the first cookbook to specify level measurements, fixing cups, tablespoons and teaspoons as the standard units of American recipe writing. In institutional food service a "standardized recipe" is one that has been tested repeatedly and produces the same result and yield when the same procedures, equipment and ingredients are used. The ingredient list followed by a prose method is thus itself a notation convention, and most computer formats preserve it.

Computers were marketed for recipe storage before any recipe format existed. The Honeywell Kitchen Computer, a pedestal version of the Honeywell 316 offered in the 1969 Neiman Marcus Christmas catalog, was advertised for storing recipes but had to be programmed through binary front-panel switches, and there is no record of one being sold. Home recipe programs for microcomputers followed in the early 1980s: Virtual Combinatics' Micro Cookbook had sold 15,000 copies by March 1984 and was the first such program to enter the home-software best-seller lists, offering searching, scaling by number of diners and grocery lists, with additional recipes sold on diskette. A 1984 book of BASIC programs for the kitchen described a "culinary database" and menu planning. Later commentators have argued that these early tools failed because they required re-typing existing recipes and did not match the provisional way people cook.

The motivation for a formal notation was stated explicitly in 1985 by David A. Mundie, who argued that "a notation must be adopted which captures the essence of recipes more clearly, more cleanly, and more succinctly than the prosaic recipes that are the norm today", and that a formal language "readily lends itself to computer analysis". In computing research the same idea recurs: a 2019 survey of food computing notes that recipe retrieval generally begins by converting cooking instructions into a structured representation such as a graph or tree.

History

Early computer formats (1980s–1990s)

RxOL was designed by David A. Mundie at Culinary Software Systems in Pittsburgh and described in his self-published 1985 book Computerized Cooking, together with Cocina, a set of Macintosh programs that used it. RxOL uses postfix ("Reverse Polish") notation: ingredients are marked with an asterisk, unary operations with an equals sign and binary operations with a slash, so that a recipe forms a tree whose operations are only partially ordered. Variants are written in square brackets separated by vertical bars, modifiers follow a comma, and only metric units are permitted. Cocina could generate the ingredient list and shopping list from the notation and convert recipes back into English. A sample from the book:

< *[PORK WITH PARSLEY #1 | BEEF WITH MINT #2].
[ *pork #1 | *beef #2] *butter /sauté in
[ *parsley #1 | *mint #2] /sprinkle with >

Mundie later said he distributed the text by e-mail to a few friends and that he remained the only person writing recipes in RxOL, although he continued to maintain his own collection in it. The format is cited as prior art by the 2015 Pesto specification, which describes RxOL as constructing a graph from recipes written with a few operators and postfix notation.

The USENET Cookbook was begun in December 1985 by Brian K. Reid at Digital Equipment Corporation's Western Research Laboratory as an experiment in electronic publishing. Recipes were submitted by e-mail, edited, and distributed weekly through the moderated newsgroup mod.recipes and later alt.gourmand; by December 1987 it had about 13,000 subscribers, some 300 contributors and 442 articles in 99 issues. Recipes were encoded in a macro package layered on the Unix manual-page macros of troff, and Reid wrote that they were "structured in a very primitive, very rigid line-oriented style" with every ingredient on a separate line and fixed quoting conventions, so that "it does not take much programming skill to write simple database programs to do interesting things with a database full of published recipes". Each recipe began with a .RH line carrying a database key and dates, followed by .RZ (title and one-line description), .IH (yield), one .IG line per ingredient with quantity, name and optional metric equivalent, .PH and numbered .SK steps, a rating section and a .WR attribution; an "international desk" added metric or customary equivalents so that every recipe carried both. The collection was later re-encoded in Meal-Master format and circulated with other recipe archives.

Meal-Master, a shareware recipe database for MS-DOS by Scott Welliver's Episoft Systems, produced the most widely circulated interchange format of the 1990s. Its import specification requires a first line beginning with at least five hyphens and containing the words "Meal-Master", followed by Title:, Categories: and Yield: lines, then ingredient lines in fixed columns: the quantity in positions 1–7, a two-letter unit code in positions 9–10 and the ingredient name from position 12, with a maximum of 28 characters; directions follow, and a line of hyphens ends the recipe. Exported files usually open with the header MMMMM----- Recipe via Meal-Master (tm) and close with MMMMM, and the .mmf extension became attached to them. The 1994 README for version 7.07 notes that many users posted exported recipes on Prodigy's Food and Wine forum for others to capture and import "as the electronic transfer requires no re-typing", that the program was distributed through bulletin board systems and shareware distributors, and that a converter was supplied for CompuChef, Micro Cookbook and Micro Kitchen Companion files. The same file dates version 6.41 of the program to January 1992. Version 8.06, dated 24 October 1998, was released as freeware for personal use, and the program became unusable as later versions of Windows dropped DOS support, prompting migrations to newer applications. Archives of more than 158,000 recipes in Meal-Master format remain available, and the format is still imported by many recipe managers. In German-speaking countries the Rezkonv format, developed by Ilka Spieß and Ulli Fetzer, succeeded Meal-Master with longer fields, range quantities and German units.

MasterCook, developed by Arion Software of Austin, Texas, and acquired by Sierra On-Line in 1995, exported recipes in the MXP text format, which begins with the line * Exported from MasterCook *, uses spaces and keywords rather than tabs, separates an ingredient's preparation note with -- , and was recommended by the vendor as the format for exchanging recipes over the Internet because every version of the program could import it. MasterCook 5 (1999) introduced MX2, an XML-like export with an external DTD; a 2017 analysis found that MX2 files are not well-formed XML because of an invalid declaration, unescaped characters and control characters. Both MXP and MX2 remained import and export targets for later software such as Living Cookbook and Now You're Cooking!. Other programs of the period, such as Compu-Chef (.ccf text export) and Key Home Gourmet (an export identical to MasterCook's), are known mainly through the import lists of their competitors. A 2004 Linux Journal column observed that thousands of recipes were available on the Internet in Meal-Master and RecipeML format and that the KDE program Krecipes could import RecipeML, MasterCook and Meal-Master files.

XML formats (2000s)

RecipeML was published by the American company FormatData, first under the name DESSERT, with drafts dated 1 April and 1 August 2000; version 0.5, dated 4 November 2000, adopted the name RecipeML. A document consists of a recipeml root containing one or more recipe elements, each with a head (title, categories, preparation time, yield), ingredients made of ing elements whose amt separates quantity and unit from the item, and directions divided into step elements; equipment, nutrition and notes are optional. An incomplete working draft by James Saiya dated 8 April 2001 extended the vocabulary to menus, alternative ingredients, tool and ingredient references within steps, and nutrient groups. FormatData's licence allowed anyone to use the DTD provided it remained unchanged, required users to mention RecipeML by name and required software developers to obtain permission; the specification was never declared final. Four development partners were listed in 2001, including the appliance maker TurboChef and the website Recipezaar, but the site's gallery of applications, last updated in 2002, listed a single personal website, and the site was last updated in 2006. Support was confined to a few desktop recipe managers. The author of a later format described RecipeML as "mired in licensing problems".

CookML is a German XML format created to transfer the full data of the recipe program Kalorio, developed by Jochen Herz through discussion with other recipe-software developers. It supports images, menus and the assignment of codes from the German national nutrient database (Bundeslebensmittelschlüssel) to ingredients, uses the .cml extension, and was at version 1.1.3 in September 2018. A 2015 survey of recipe formats recorded version 1.0.4, dated 2006, and the format was in use for exchanging recipes among Kalorio users by August 2007. REML (Recipe Exchange Markup Language), an XML schema by Gary Gocek released under the BSD licence with version 0.5 dated 2005, was written to support commercial uses such as restaurant menus and cookbooks; it defines menu, vendor and mealItem elements covering ingredients, nutrition, costs and procedures, with a C# reference implementation. Aaron Straup Cope's Eatdrinkfeelgood Markup Language, whose version 1.1 dates from 2002, is an XML application for recipes, courses and menus designed for reuse through XSLT, automatic unit conversion and editing in a plain text editor; in a 2005 article he wrote that when he reviewed existing XML recipe work "most applications simply lacked the features I was looking for" and the rest "seemed only to exist as an effort to 'monetize the recipe space'". Application-native XML exports of the period, such as MasterCook's MX2 and Living Cookbook's FDX, which embeds recipe XML together with encoded photographs, were never published as open standards but survive as import targets.

Web structured data (2008–present)

hRecipe is a microformat for marking up recipes in HTML using class attributes. Drafted within the microformats community from 2007, it was placed in the public domain on 14 November 2008 by its authors Frances Berriman, Ben Ward and Toby Inkster, with Thomas Lörtsch as editor. Its only required properties are fn (the recipe name) and one or more ingredient; optional properties include yield, instructions, duration, photo, summary, author, published and nutrition. Food Network marked up all its recipes with hRecipe in October 2009, and Epicurious, Allrecipes, Williams-Sonoma and Plan to Eat followed by April 2010. Its microformats2 successor, h-recipe, uses prefixed class names such as p-ingredient and e-instructions.

Google introduced rich snippets, search-result summaries built from structured data in web pages, on 12 May 2009 for reviews and people, and on 14 April 2010 added recipes as its fifth rich-snippet type, showing ratings, preparation time and a picture in results for pages marked up with microdata, RDFa or hRecipe. On 24 February 2011 Google launched Recipe View in the United States and Japan, a filter restricting results to recipes and allowing them to be narrowed by ingredient, cooking time and calories; the company described it as the first search filter built entirely on webmasters' structured-data markup. Critics argued that the requirement to embed markup disadvantaged small food bloggers and favoured large sites with search-optimisation staff, that default filters for short cooking times and low calories promoted a particular cooking culture, and that recipe search engines surfaced recipes whose metadata had been coded to boost rankings even when inaccurate.

On 2 June 2011 Google, Bing and Yahoo! announced schema.org, a shared vocabulary for structured data that ported the existing rich-snippet types, including Recipe, into one hierarchy. The schema.org Recipe type is a subtype of HowTo and defines properties including recipeIngredient (which superseded the earlier ingredients), recipeInstructions, recipeYield, prepTime, cookTime and totalTime as ISO 8601 durations, nutrition, recipeCategory, recipeCuisine, suitableForDiet and cookingMethod, and can be expressed in Microdata, RDFa or JSON-LD. Google's documentation requires name and image for recipe rich results and recommends times, yield, ingredients, instructions as HowToStep items, nutrition and ratings, with examples in JSON-LD. In May 2016 Google introduced Rich Cards, a mobile-oriented result format launched first for recipes and movies and based on schema.org markup, for which it recommended JSON-LD, and in January 2020 it announced that its older data-vocabulary.org markup would cease to qualify for rich results from 6 April 2020 because of the popularity of schema.org.

Measurements of the Common Crawl by the Web Data Commons project document the transition. In the November 2013 crawl hRecipe appeared on 3,530 domains and 814,793 URLs, the largest publishers being Epicurious, GroupRecipes and Chefkoch. In the December 2014 crawl hRecipe had fallen to 3,476 domains and 630,402 URLs, while schema.org Recipe in Microdata appeared on 12,593 hosts and 1,627,496 URLs, most often together with AggregateRating and Review.

Other platforms adopted the same markup. Pinterest introduced Rich Pins on 20 May 2013 with recipe pins showing ingredients and cooking times, with launch partners including Bon Appétit, Epicurious, Real Simple and Whole Foods Market; recipe Rich Pins require schema.org or h-recipe markup. Facebook's Open Graph protocol defines no recipe type. Recipe-import tools such as the Python library recipe-scrapers parse schema.org JSON-LD, Microdata and RDFa or Open Graph metadata from recipe pages, and the self-hosted manager Nextcloud Cookbook stores each recipe natively as a schema.org JSON-LD file. Because the machine-readable recipe card is separable from surrounding prose, food bloggers have been able to write long introductions for search-engine and advertising reasons while offering "jump to recipe" links, a practice that has drawn recurring criticism, and in 2021 a tool that stripped blog prose to show only the structured recipe prompted debate over the ethics of separating recipes from their authors' writing.

Plain-text and developer-oriented formats (2010s–2020s)

A separate line of formats was designed to be written by hand in a text editor and kept in version-control systems, in contrast to markup emitted by publishing software. In 2004 Michael Chu's website Cooking for Engineers began presenting recipes as summary tables, with ingredients down the left and operations as columns whose cells merge as components are combined.

The Open Recipe Format, created by Joseph Hall in 2013 under the Apache License 2.0, uses YAML on the grounds that it is more readable than JSON, lighter than XML and easily managed with Git; ingredients are mappings with amounts, units, processing tags, substitutions and USDA database numbers, and the schema also covers steps with HACCP control points, yields, oven settings and source metadata. Chowdown (January 2015) stores recipes as Markdown files with YAML front matter listing ingredients and directions, rendered as a static website with Jekyll.

Pesto, first published on 14 June 2015 by Lars-Dominik Braun under the CC0 licence, is a "text-based, human-editable, and machine-transformable" interchange format implemented as a literate Haskell program that is simultaneously specification and parser. A recipe begins with %pesto and is a sequence of operators, + for ingredients, & for tools, square brackets for actions and > for results, that build a directed graph from ingredients through actions to the finished dish; the design goal is that "a human editor must be able to remember the entire syntax". Its specification surveys earlier formats and credits RxOL for the graph-construction idea.

RecipeMD, begun by Tilman Stehr in August 2018 and released under the LGPL, constrains CommonMark Markdown: the title is a level-one heading, tags are a paragraph in italics, yields a paragraph in bold, ingredients a list whose amounts are italicised and may link to sub-recipes, and instructions follow a second horizontal rule. A Python reference implementation provides scaling and extraction from web pages.

Cooklang was created by Alexey Dubovskoy; its repositories date from early 2021 and it was released under the MIT licence. A recipe is written as prose steps in which @ marks an ingredient, with an optional quantity and unit in braces, # cookware and ~ a timer, with metadata in YAML front matter; the ingredient list is derived from the annotations rather than written separately. A command-line tool converts recipes to JSON, YAML, Markdown, LaTeX and Typst. A 2025 study in IEEE Access evaluated large language models converting unstructured recipe text into the format.

Large language models have also influenced the field. YumML, a YAML recipe format first drafted by Paul Jenkins in September 2011, was revived in January 2026 by Miguel García, who argued that YAML consumes fewer tokens than JSON or XML when processed by large language models, and such models are used to convert unstructured recipe text into structured formats, a feature offered by Mealie and by Cooklang's import tooling.

Alongside these languages, consumer recipe applications exchange data in their own JSON-based files. Paprika's .paprikarecipes export is a ZIP archive of gzip-compressed JSON objects, and Mela's .melarecipe is a JSON document with base64-encoded images and limited Markdown in ingredients and instructions. Self-hosted recipe managers act as conversion hubs: Tandoor Recipes documents import from more than twenty applications and formats, including Meal-Master, Rezkonv, Paprika, Mela, Chowdown, Cooklang and the JSON-LD export of RecipeSage.

Design approaches

Recipe formats differ mainly in how they relate the ingredient list to the procedure. Ingredient list plus prose formats, from the USENET Cookbook and Meal-Master through RecipeML, schema.org Recipe, Open Recipe Format and RecipeMD, keep the traditional two-part layout and structure the ingredient lines (quantity, unit, item, preparation) while leaving the method as text or as an ordered list of steps. This is the easiest to produce from existing recipes, but the method remains opaque to software; schema.org's recipeIngredient is a plain string, and Google's recipe rich results do not require quantities to be separated from items. Inline-annotated prose formats, of which Cooklang is the main example, mark ingredients, cookware and timers where they occur in the instructions and derive the ingredient list from them, so that quantities are stated once and steps can be linked to what they consume. Graph or postfix notations, RxOL and Pesto, abandon prose and describe a recipe as operations applied to ingredients, yielding a tree or directed graph whose partial ordering exposes which steps can run in parallel; Mundie noted that in such a notation "the ingredient list falls out naturally as a by-product". Schema-based structured data, hRecipe and schema.org, is not an authoring format at all but a vocabulary layered onto an HTML page for consumption by search engines and aggregators.

What a format encodes also varies. Quantities and units appear in nearly every format, but only some handle fractions, ranges and metric–customary pairs: the USENET Cookbook carried both systems on each ingredient line, Meal-Master used two-letter unit codes within a 28-character field, Rezkonv was created partly to allow ranges such as "1–2 cloves", and RxOL permitted only metric units. Cookware is explicit in RxOL, RecipeML's equipment, Pesto's & operator, Cooklang's # and schema.org's inherited tool property. Timers and durations are first-class in Cooklang and REML and expressed as ISO 8601 durations in schema.org. Nutrition is modelled in RecipeML, REML, CookML (through national nutrient-database codes) and schema.org, and deliberately excluded from Pesto. Yield and scaling are supported by most formats; Meal-Master and MasterCook scaled recipes to a number of servings in the 1990s, and RecipeMD and Cooklang provide scaling in their tools.

Research in natural language processing has pursued the graph view independently of these interchange formats. Hamada and colleagues performed structural analysis of Japanese preparation steps in 2000, and Tasse and Smith defined MILK, a first-order-logic "Minimal Instruction Language for the Kitchen", with an annotated corpus of recipes. Mori and colleagues released a corpus of recipe flow graphs, directed acyclic graphs whose root is the finished dish and whose vertices are foods, tools and actions, later extended to English. Kiddon and colleagues interpreted recipes as action graphs without supervision, and Jermsurawong and Habash proposed SIMMR, a tree in which instructions merge ingredients, with a parser reaching 93.5% edge accuracy. Tree-structured representations have been used to compare hundreds of recipes for the same dish, and computer-vision work has represented recipes and food images as executable "cooking programs".

Ontologies and knowledge graphs

Several ontologies model recipes for the Semantic Web. The BBC Food Ontology, version 1.1 of 18 March 2014 and released under a Creative Commons Attribution licence, is a lightweight RDF vocabulary for publishing recipe data with classes for recipes, ingredients, methods, stages, diets, courses, cuisines, seasons and occasions, and presents itself as complementary to Google rich snippets and schema.org. An ontology design pattern for cooking recipes, developed at Wright State University in 2014, aimed to integrate recipe information from heterogeneous websites, and Chatterjee and colleagues proposed a recipe model in 2016 supporting queries not possible with schema.org or the BBC Food Ontology. Broader food ontologies and knowledge graphs incorporate recipes: FoodOn (2018) harmonises food vocabulary for traceability and data integration, FoodKG (2019) links recipes, nutrition data and FoodOn for recommendation with allergy constraints, and a 2022 review surveys knowledge graphs for food science, including those built on the BBC Food Ontology.

Comparison of formats

Format	Year	Creator or maintainer	Syntax	Licence or status
RxOL	1985	David A. Mundie (Culinary Software Systems)	Postfix notation with *, = and / operators	Self-published; single user
USENET Cookbook	1985–1987	Brian K. Reid (DEC WRL)	troff macros (.RH, .IG, .SK ...)	Copyright USENET Community Trust; discontinued
Meal-Master (.mmf)	by 1992	Episoft Systems (Scott Welliver)	Fixed-column plain text with MMMMM delimiters	Shareware, freeware from 1998; program discontinued
MasterCook MX2	1999	Sierra On-Line	XML-like with DTD	Proprietary export format
RecipeML (DESSERT)	2000	FormatData (James Saiya)	XML (DTD and XSD)	Free use, no modification; dormant since 2006
Eatdrinkfeelgood ML	by 2002	Aaron Straup Cope	XML	Version 1.1 dated 2002; described on XML.com in 2005
REML	2005	Gary Gocek	XML Schema	BSD licence
CookML (.cml)	by 2006	Jochen Herz (Kalorio)	XML	Free of charge; v1.1.3 in 2018
hRecipe / h-recipe	2008	Microformats community (Berriman, Ward, Inkster; Lörtsch)	HTML class attributes	Public domain; draft
schema.org Recipe	2011	Schema.org (Google, Bing, Yahoo!)	Microdata, RDFa or JSON-LD	Open vocabulary; in active use
YumML	2011	Paul Jenkins; revived by Miguel García (2026)	YAML	Draft specification
Open Recipe Format	2013	Joseph Hall	YAML	Apache License 2.0
BBC Food Ontology	2014	BBC	RDF/OWL	CC BY 4.0
Chowdown	2015	clarklab (GitHub)	Markdown with YAML front matter	Unlicense
Pesto	2015	Lars-Dominik Braun	Operator-based graph notation	CC0
RecipeMD	2018	Tilman Stehr	Constrained CommonMark Markdown	LGPL-3.0
Cooklang (.cook)	2021	Alexey Dubovskoy	Inline-annotated plain text with YAML front matter	MIT

See also

• HRecipe
• Schema.org
• Microformat
• USENET Cookbook
• Recipe
• Esoteric programming language, for Chef, a language whose programs read as recipes
• Honeywell 316, for the Honeywell Kitchen Computer
• Lightweight markup language
• Computational gastronomy