diff --git a/.gitignore b/.gitignore
index dd130e4f00..e8dfeddbe6 100644
--- a/.gitignore
+++ b/.gitignore
@@ -7,6 +7,7 @@
 /doc/java/
 /doc/rb/
 /doc/rust/
+/grammars/
 /pkg/
 /spec/reports/
 /top-100-gems/
diff --git a/Gemfile b/Gemfile
index dad7c06177..5fd9650c6b 100644
--- a/Gemfile
+++ b/Gemfile
@@ -26,3 +26,5 @@ gem "onigmo", platforms: :ruby
 if RUBY_PLATFORM =~ /aarch64.*mingw/
   gem "nokogiri", github: "sparklemotion/nokogiri", branch: "main"
 end
+
+gem "lrama"
diff --git a/Gemfile.lock b/Gemfile.lock
index 6c738e2e7e..3efb5000fe 100644
--- a/Gemfile.lock
+++ b/Gemfile.lock
@@ -27,6 +27,7 @@ GEM
       rdoc (>= 4.0.0)
       reline (>= 0.4.2)
     jar-dependencies (0.5.5)
+    lrama (0.8.0)
     mini_portile2 (2.8.9)
     nokogiri (1.19.1)
       mini_portile2 (~> 2.8.2)
@@ -96,6 +97,7 @@ DEPENDENCIES
   benchmark-ips
   ffi
   irb
+  lrama
   onigmo
   parser
   prism!
diff --git a/rakelib/parsey.rake b/rakelib/parsey.rake
new file mode 100644
index 0000000000..8a0dc21f39
--- /dev/null
+++ b/rakelib/parsey.rake
@@ -0,0 +1,172 @@
+# frozen_string_literal: true
+
+class ParseYGrammar
+  def initialize(source, filepath)
+    stdlib_path = Lrama::Command::STDLIB_FILE_PATH
+    stdlib = Lrama::Parser.new(File.read(stdlib_path), stdlib_path).parse
+
+    grammar = Lrama::Parser.new(source, filepath).parse
+    grammar.prepend_parameterized_rules(stdlib.parameterized_rules)
+    grammar.prepare
+    grammar.validate!
+
+    @grammar = grammar
+    @filepath = filepath
+  end
+
+  def strip
+    stripped = "#{declarations}\n%%\n\n#{rules}\n%%\n"
+    raise if shape != self.class.new(stripped, @filepath).shape
+    stripped
+  end
+
+  protected
+
+  # The shape of a parsed grammar used to validate that we have not accidentally
+  # modified the automaton in some way while stripping the actions.
+  def shape
+    renames = build_renames(@grammar)
+    rules =
+      @grammar.rules.filter_map do |rule|
+        lhs = rule.lhs.id.s_value
+        next if lhs == "$accept" || lhs.match?(/\A[@$]/)
+
+        [symbol_name(rule.lhs, renames), rule.rhs.map { |sym| symbol_name(sym, renames) }, rule.precedence_sym&.id&.s_value]
+      end
+
+    precedence =
+      @grammar.terms.filter_map do |term|
+        if (precedence = term.precedence)
+          [term.id.s_value, precedence.type, precedence.precedence]
+        end
+      end
+
+    [rules, precedence.sort, @grammar.expect]
+  end
+
+  private
+
+  def symbol_name(symbol, renames)
+    s_value = symbol.id.s_value
+    s_value.match?(/\A[@$]/) ? "{}" : renames.fetch(s_value, s_value)
+  end
+
+  def declarations
+    out = +""
+
+    @grammar.define.sort.each do |key, value|
+      out << "%define #{key}#{" #{value}" if value}\n"
+    end
+
+    out << "%expect #{@grammar.expect}\n"
+    out << "%parse-param {#{@grammar.parse_param}}\n"
+    out << "%lex-param {#{@grammar.lex_param}}\n"
+    out << "\n"
+
+    @grammar.terms.each do |term|
+      s_value = term.id.s_value
+
+      # Token names lrama defines internally; they are never declared by a
+      # grammar and must not be re-declared by the emission.
+      next if %w[error YYerror YYUNDEF].include?(s_value) || !s_value.match?(/\A[a-zA-Z_]/)
+
+      out << "%token #{s_value} #{term.token_id}"
+      out << " #{term.alias_name}" if term.alias_name
+      out << "\n"
+    end
+    out << "\n"
+
+    levels = @grammar.terms.select(&:precedence).group_by { |term| term.precedence.precedence }
+    levels.sort.each do |_level, terms|
+      out << "%#{terms.first.precedence.type} #{terms.map { |term| term.id.s_value }.join(" ")}\n"
+    end
+
+    start = @grammar.rules.first.rhs.first.id.s_value
+    out << "\n%start #{start}\n"
+    out
+  end
+
+  def rules
+    out = +""
+    ind = " " * 16
+
+    renames = build_renames(@grammar)
+    previous = nil
+
+    @grammar.rules.each do |rule|
+      lhs = rule.lhs.id.s_value
+      next if lhs == "$accept" || lhs.match?(/\A[@$]/)
+
+      alts = rule.rhs.map { |sym| symbol_name(sym, renames) }
+      alts = ["%empty"] if alts.empty?
+
+      # A lone {} in final position would be the rule action rather than a
+      # hidden nonterminal; a mid-rule action ending its rule needs a trailing
+      # empty action after it to keep its own reduction.
+      alts << "{}" if alts.last == "{}"
+
+      # The default rule precedence is the last terminal's; only explicit
+      # overrides need re-emitting.
+      last_term = rule.rhs.reverse.find(&:term?)
+      if rule.precedence_sym && rule.precedence_sym != last_term
+        alts << "%prec #{rule.precedence_sym.id.s_value}"
+      end
+
+      lhs = renames.fetch(lhs, lhs)
+      if lhs == previous
+        out << "#{ind}| #{alts.join(" ")}\n"
+      else
+        out << "#{ind};\n\n" unless previous.nil?
+        out << "#{lhs}: #{alts.join(" ")}\n"
+        previous = lhs
+      end
+    end
+
+    out << "#{ind};\n"
+    out
+  end
+
+  # Parameterized-rule instantiations get generated names that are not always
+  # valid grammar identifiers (e.g., option_'\n').
+  def build_renames(grammar)
+    @grammar.nterms.each_with_object({}) do |nterm, renames|
+      s_value = nterm.id.s_value
+      if !s_value.match?(/\A[@$]/) && !s_value.match?(/\A[a-zA-Z_][a-zA-Z0-9_]*\z/)
+        renames[s_value] = s_value.gsub(/[^a-zA-Z0-9_]/, "_")
+      end
+    end
+  end
+end
+
+directory "tmp"
+directory "tmp/ruby" => "tmp" do |task|
+  mkdir_p task.name
+  chdir task.name do
+    sh "git init"
+    sh "git remote add origin https://github.com/ruby/ruby"
+  end
+end
+
+directory "grammars"
+["3.3", "3.4", "4.0", "master"].each do |version|
+  desc "Fetch the #{version} parse.y grammar"
+  file "grammars/#{version}.y" => ["tmp/ruby", "grammars"] do |task|
+    chdir "tmp/ruby" do
+      ref = version
+
+      if ref != "master"
+        major, minor = ref.split(".")
+        refs = `git -c versionsort.suffix=_preview -c versionsort.suffix=_rc ls-remote --refs --sort=-version:refname origin "v#{major}_#{minor}_*" "v#{major}.#{minor}.*"`
+        ref = refs[%r{refs/tags/(\S+)}, 1] or raise
+      end
+
+      sh "git fetch --depth=1 origin #{ref}"
+      sh "git reset --hard FETCH_HEAD"
+    end
+
+    sh "ruby -p tmp/ruby/tool/id2token.rb < tmp/ruby/parse.y > #{task.name}"
+
+    require "lrama"
+    File.write(task.name, ParseYGrammar.new(File.read(task.name), task.name).strip)
+  end
+end