Mustache is a very sweet template system inspired by ctemplate and el. Ever wondered how mustache is implemented or how to write your own template system? this post will teach you the basics.

NOTE: although this post discusses abstract syntax trees and lexical analysis, don’t be intimidated, no experience other than a basic knowledge of Clojure and regular expressions is needed to read, understand and be able to implement your own template language.

A simple example of how our templates will look like:

;; First we define the template
(def template
"<p> Hello {{name}}, your friends are: </p>
   <li>{{first-name}} {{last-name}} </li>

;; Now we render the template
(render template
        {:name "Bob"
         :friends [{:first-name "Bill" :last-name "Gates"
                    :first-name "Chuck" :last-name "Cheese"}]})

;; Output should be something like:
"<p> Hello Bob, your friends are: </p>
   <li>Bill Gates</li>
   <li>Chuck Cheese</li>

Get the code

This blog post is based on gabo, I suggest you take a look at the README, clone the project or at least play with it a little using the repl. Code is also available in clojars.


Rendering templates is a 3 step process:

Step 1: Lexical Analysis

The purpose of the lexical analysis step is to convert a template string into a list of tokens.

Step 2: Parsing

The second step, parsing, is in charge of converting the list of tokens produced by the lexical analysis step into an abstract syntax tree.

Step 3: Rendering

The final step takes as input an abstract syntax tree and some input and renders the given input using the template as structure.

Now that we have a high level understanding of the steps required to compile and render templates let’s get into detail.

Step 1: Lexical Analysis

The main purpose of the lexical analysis step is to implement the tokenize function which takes a template as input and converts it into a list of tokens. A token is a very simple data structure (usually a tuple) that splits the template into meaningful lexical elements. A typical programming language will have string tokens, reserved keyword tokens, number tokens, etc.

Let’s see with an example how the tokenize function works.

Example 1: tokenizing a simple template

user=> (use 'gabo.lexer)
user=> (tokenize "Hi {{name}}")
[[:literal "Hi "] [:symbol "name"]]

Example 2: tokenizing a more complex template

user=> (tokenize "{{#people}}Name: {{name}}, age: {{age}}{{/people}}")
[[:iter-init "people" :default]
 [:literal "Name: "]
 [:symbol "name"]
 [:literal ", age: "]
 [:symbol "age"]
 [:iter-end "people"]]

Note: for now let us ignore the :default keyword present in the first token, we will get to that later.

Tokens and their syntax:

All tokens are vectors of 2 or more elements. The first one is called the token identifier, the second one corresponds to the token’s value.

There are 4 types of tokens:

  • :symbol: Symbols use the following syntax {{something}}.
  • :iter-init: Mark the beginning of an iteration, they have the following syntax: {{#something}}.
  • :iter-end: they mark the termination of an iteration as started by an :iter-init token. They look like {{/something}}
  • :literal: Literals match anything that is not a symbol, iter-init or iter-end


Tokenization is the process of converting a string into a list of tokens, the tokenization procedure is implemented using two functions: tokenize and tokenize-chunk:

The tokenize function
(defn tokenize
  (loop [remaining template
         tokens []]
    (if (empty? remaining)
      (let [[str-match token] (tokenize-chunk remaining)]
        (recur (.substring remaining (count str-match))
               (conj tokens token))))))

The tokenize function is a very simple one, it takes as input a template string and loops through the template until it reaches the end. On every iteration the tokenize-chunk function is called which tries to match the remaining template to a token, when a match is found, the tokens vector will conjoined with the new token and the remaining template will be substringed by the matched string.

As you probably guessed, the tokenize-chunk is the function that is doing the actual heavy work of matching strings to tokens.

(defn tokenize-chunk
  "Given a template substring returns a tuple of [match token]
  where match is the exact string that matched and token is
  the token that corresponds to the regex match.
  Example: (tokenize-chunk '{{bar}} foo') will return
  ['{{bar}}' [:symbol 'bar']]"
  (when-match template
    ;; match symbols e.g. {{foo}}
    [sym #"\A\{\{\s*([\w-\.]+)\s*\}\}"]
      [ (first sym) [:symbol (last sym) ] ]

    ;; match iter-end e.g. {{/foo}}
    [iter-end #"\A\{\{/\s*(\w+)\s*\}\}"]
      [ (first iter-end) [:iter-end (last iter-end) ]]

    ;; match iter-init with no args e.g. {{#foo}}
    [iter-init #"\A\{\{#\s*(\w+)\s*\}\}"]
      [ (first iter-init) [:iter-init (last iter-init) :default ]]

    ;; match iter-init with args e.g. {{#foo 'separator'}}
    [iter-init #"\A\{\{#\s*(\w+)\s+'([\s\S]*?)'\s*\}\}"]
      [(first iter-init) (cons :iter-init (rest iter-init))]

    ;; match literals
    [literal #"\A([\s\S][\s\S]*?)\{\{"]
      [(last literal) [:literal (last literal)]]

    ;; more literals
    [literal #"\A[\s\S]*"]
      [literal [:literal literal]]))

when-match, what is that?

You probably noticed the use of the when-match macro. The implementation for the when-match macro can be found here.

Example usage of when-match:
user=> (when-match "a long string"
  #_=>   [match #"long"] (keyword match)
  #_=>   [wont-match #"string"] (keyword wont-match))

when-match takes a string as its first argument followed by n pairs of forms. The macro will try to match the regular expression in the first form. If a match is found then the symbol preceding the regular expression will be let and the second form will be evaluated. Regular expression matches use re-find under the hood.

If there is no match then the same will be attempted with the next pair of forms. I suggest you play around with when-match in your REPL to get a solid understanding of how it works.

Going back to the tokenize-chunk function: Implementation is very simple once you understand when-match, each regex matches to one type of token, when a token is matched a pair of [string token] will be returned where string is the actual string that was matched and token is the corresponding token constructed from the matched string.

Example: tokenize-chunk at work.
user=> (tokenize-chunk "Hi {{name}}, how are you?")
["Hi " [:literal "Hi "]]
user=> (tokenize-chunk "{{name}}, how are you?")
["{{name}}" [:symbol "name"]]
user=> (tokenize-chunk ", how are you?")
[", how are you?" [:literal ", how are you?"]]

Notice how this example tokenizes the "Hi {{name}}, how are you?" by matching chunk by chunk to a token.

We are done with the lexical analysis step. Let’s go now to the parsing step.

Step 2: Parsing

Let us recall that the purpose of this step is to produce an abstract syntax tree or AST. Think of an AST as a tree like representation of our template which we can later use to actually render templates.

Consider a template which renders a country and every city in the country. We will first tokenize it and then manually convert the tokens into an AST.

(def template "Countries with their cities:
{{#countries '\n'}}
Country: {{name}}
{{#cities '\n'}}
 * {{name}}

Tokenization is quite simple using our tokenize function

user=> (tokenize template)
[[:literal "Countries with their cities:\n"]
 [:iter-init "countries" "\n"]
 [:literal "\nCountry: "]
 [:symbol "name"]
 [:literal "\nCities:\n"]
 [:iter-init "cities" "\n"]
 [:literal " * "]
 [:symbol "name"]
 [:iter-end "cities"]
 [:literal "\n"]
 [:iter-end "countries"]]

We will now convert those tokens into a tree. The conversion will use these rules:

  • :literal or :symbol tokens will be left untouched
  • :iter-init and :iter-end tokens will be grouped into an :iter node and all tokens between the init/end pair will be parsed recursively.
Example of an AST as produced by the parse function
user=> (parse template)
[[:literal "Countries with their cities:\n"]
 [:iter "countries" "\n" [[:literal "\nCountry: "]
                          [:symbol "name"]
                          [:literal "\nCities:\n"]
                          [:iter "cities" "\n" [[:literal "\n * "]
                                                [:symbol "name"]
                                                [:literal "\n"]]]
                          [:literal "\n"]]]]

Let’s now take a look at the parse function:

(defn parse
  (-> (tokenize string)

It doesn’t do much, let’s take a look at the build-ast which takes a list of tokens as input and outputs an AST.

(defn- build-ast
  "Builds an abstract syntax tree given a list of tokens as produced by tokenize"
  (loop [tokens tokens
         ast []]
    (if (empty? tokens)
      (let [token (first tokens)]
          (or (is-literal token) (is-symbol token))
            (recur (rest tokens)
                   (conj ast token))
          (is-iter-init token)
            (let [sub-list (find-iter-sub-list tokens)
                  [_ token-val separator] token]
              (recur (drop (+ 2 (count sub-list)) tokens)
                     (conj ast [:iter token-val
                                      (build-ast sub-list)])))
            (throw (unexpected-token-exception token)))))))

The build-ast function is a bit complex, but let’s break it down into a series of steps:

  • It loops through all the tokens; the ast which will be the result of the function.
  • The if in line 6 simply asserts that when there are no tokens left, we are done so just return the ast.
  • If there are tokens left then line 8 lets token to the first token in the iteration.
  • The cond has 2 cases and an else
    • Case 1: the “simple” case, if the token is a literal or symbol simply add token to the ast.
    • Case 2: the “hard” case is evaled when an :iter-init token is found.
      The find-iter-sub-list function will return all tokens inside the :iter-init and :iter-end pair. The recur call will drop all tokens found in the sub-list + 2 which corresponds to the :iter-init and :iter-end tokens. The ast is then conjoined to a recursive call of the build-ast procedure over the sub-list of tokens.
    • else case: if something other than an :iter-init, :symbol or :literal is found, then throw an exception because we found an unexpected token.

This finishes our explanation of the parsing step. I have tried to write the simplest possible parser, hopefully you think so too. A more complex one would, for example, include better error detection mechanisms.

Step 3: Rendering

We have arrived to the final step in our template rendering system. This system builds upon steps 1 and 2 and takes an AST and a context as arguments and returns a rendered template. A context is simply a map where every key is a keyword and vals are numbers, strings or collections of contexts.

;; The {{.}} refers to the item that is being iterated.
user=> (def tokens (parse "{{#numbers}}Number:{{.}}{{/numbers}}"))
user=> tokens
user=> [[:iter "numbers" :default [[:literal "Number:"]
                                   [:symbol "."]]]]
user=> (eval-tree tokens {:numbers [1 2 3 4 5]})

Let us now look at the implementation of eval-tree:

(defn eval-tree
  "Evaluates a compiled template as a tree with the given context"
  [tree ctx]
  (cond (is-literal tree)
          (second tree)
        (is-symbol tree)
          (let [[_ token-val] tree]
            (if (= token-val ".")
              (str ctx)
              (get ctx (keyword token-val) "")))
        (is-iter tree)
          (let [[_ token-val separator sub-tree] tree
                coll (get ctx (keyword token-val) [])]
            (->> (map #(eval-tree sub-tree %) coll)
                 (interpose (if (= :default separator) "," separator))
                 (apply str)))
        ;; Only executed once: the first time eval-tree is called, no subsequent
        ;; recursive call will go through this branch.
        (coll? tree)
          (apply str (map #(eval-tree % ctx) tree))))

eval-tree runs through each node in the AST recursively as follows:

  • If a :literal token is found, return the token’s value.
  • If a :symbol is found then check if its value is "."
    • Case token-val is ".": return the complete context as a string.
    • Case token-val is not ".": look for token-val as a keyword inside the context map and return that.
  • If an :iter tree is found then:
    • let coll to to the collection inside the context that matches the :iter’s token-val.
    • Recursively call eval-tree on every node in the sub-tree.
    • Interpose the previous result with the separator which is by default ','.
    • Finally concatenate all “sub-templates” and separators.
  • The final branch (coll? tree) is actually only evaled once, the first time that eval-tree is called as ASTs produced by parse are not actually trees, they are a list of trees. This step is equivalent to the previous branch sans interposing with separators.

Wrapping things up

Now that we have implemented tokenize, parse and eval-tree. The render function is very simple:

(defn render
  "Compiles and evaluates the template with the given context"
  [template ctx]
  (eval-tree (parse template) ctx))


I hope you now have a good understanding on how to write your own template system. If you have any questions, suggestions or just want to chat, you can e-mail me at fernandohur at

Criticism is always welcome but go easy on me, it’s my first post ;)