Rails Code Style

1. Method Ordering

Order methods in classes: class methods first, then public methods (initialize at the top), then private methods.

class SomeClass
  # 1. Class methods first
  class << self
    def deliver_all_later
      DeliverAllJob.perform_later
    end
  end

  # 2. Public methods (initialize first if present)
  def initialize(user)
    @user = user
  end

  def deliver
    # ...
  end

  # 3. Private methods
  private
    def deliverable?
      # ...
    end
end

2. Invocation Order

Within each visibility section, order methods vertically by invocation: callers before the methods they call.

class SomeClass
  def some_method
    method_1
    method_2
  end

  private
    def method_1
      method_1_1
      method_1_2
    end

    def method_1_1
      # ...
    end

    def method_1_2
      # ...
    end

    def method_2
      # ...
    end
end

3. Conditional Returns

Prefer expanded conditionals over guard clauses.

# Bad
def todos_for_new_group
  ids = params.require(:todolist)[:todo_ids]
  return [] unless ids
  @bucket.recordings.todos.find(ids.split(","))
end

# Good
def todos_for_new_group
  if ids = params.require(:todolist)[:todo_ids]
    @bucket.recordings.todos.find(ids.split(","))
  else
    []
  end
end

Exception: a guard clause is acceptable when the return is right at the beginning of the method and the main body is non-trivial:

def after_recorded_as_commit(recording)
  return if recording.parent.was_created?

  if recording.was_created?
    broadcast_new_column(recording)
  else
    broadcast_column_change(recording)
  end
end

4. Visibility Modifiers

No newline under private; indent content under it. For modules with only private methods: private at the top, extra newline after, no indent.

class SomeClass
  def some_method
    # ...
  end

  private
    def some_private_method
      # ...
    end
end

module SomeModule
  private

  def some_private_method
    # ...
  end
end

5. CRUD Controllers

Model web endpoints as CRUD on resources; when an action doesn't map to a CRUD verb, introduce a new resource instead of a custom action:

# Bad
resources :cards do
  post :close
  post :reopen
end

# Good — create = close, destroy = reopen
resources :cards do
  resource :closure
end

Full pattern (controllers, more resource examples): [[rails-controllers]] Pattern 3.

6. Controller-Model Interaction

Thin controllers directly invoke a rich domain model — plain ActiveRecord (@card.comments.create!(comment_params)) or intention-revealing model APIs (@card.gild). No services or other artifacts connecting the two; a form/service object like Signup.new(email_address:).create_identity is the rare justified exception. See [[rails-controllers]] Pattern 8 and [[rails-philosophy]].

7. Bang Method Conventions

Only use ! for methods that have a corresponding counterpart without !. Don't use ! to flag destructive actions.

record.save!     # Good: save! has a counterpart save
@card.gild!      # Bad: no gild counterpart exists
@card.gild       # Correct

8. Async Operations

Shallow job classes delegating to domain models, with _later for the enqueueing method and _now for the synchronous version when both exist:

module Event::Relaying
  extend ActiveSupport::Concern

  included do
    after_create_commit :relay_later
  end

  def relay_later
    Event::RelayJob.perform_later(self)
  end

  def relay_now
    # actual logic here
  end
end

class Event::RelayJob < ApplicationJob
  def perform(event)
    event.relay_now
  end
end

Full job patterns (recurring, queues, error handling): [[rails-jobs]].

9. ERB Comments & Template Linting

<% # comment %> looks harmless but trips ERB parsers and template tooling (Herb reports it as a parsing error). Always use the dedicated ERB comment tag <%#:

<%# Good: dedicated ERB comment tag %>
<% # Bad: Ruby comment inside an execution tag — causes parsing issues %>

If the project uses Herb (ERB language server / linter):

• bin/herb analyze app/views app/components checks all templates for parsing errors — run it after template changes and before committing.
• Configuration lives in .herb.yml (accessibility rules, HTML validity, ERB best practices).
• The VS Code Herb extension adds live linting and ERB format-on-save.

10. Enforcing Style on AI-Generated Code

Everything above this point is taste — and a skill or CLAUDE.md only asks the agent to follow it. A Claude Code hook guarantees the check runs. Deterministic beats hopeful: don't rely on the model remembering the style guide, run RuboCop and make a clean result a condition of finishing. (Technique adapted from thoughtbot's Enforcing Your Ruby Style Guide on AI-Generated Code.)

rails-toolkit ships a gentle version of this for free. When the plugin is installed, a PostToolUse hook (bin/rubocop-autocorrect-hook) runs after each .rb/.rake edit, safe-autocorrects that one file, and reports any remaining offenses back as context — but only in a project that opted into RuboCop (a .rubocop.yml) with a resolvable runner; it no-ops everywhere else. It never blocks. The Stop hook below is the stronger, blocking variant you add per-project when you want the agent's completion gated on a clean lint.

The hook

Drop this into your own Rails project (not into a plugin — a project-local hook only fires for that repo). A Stop hook runs RuboCop over the Ruby files in the diff once the agent thinks it is done, autocorrects what it can, and — if offenses remain — exits non-zero so Claude Code feeds the message back and the agent gets exactly one corrective pass before it is allowed to stop.

.claude/settings.json:

{
  "hooks": {
    "Stop": [
      { "hooks": [ { "type": "command", "command": ".claude/hooks/rubocop.sh" } ] }
    ]
  }
}

.claude/hooks/rubocop.sh:

#!/usr/bin/env sh
# Stop hook: lint the Ruby files the agent touched, autocorrect, then re-check.
# Exit 2 + stderr -> Claude Code blocks the stop and feeds the message back,
# giving the agent one chance to fix what --autocorrect could not.
set -e

files=$(git diff --name-only --diff-filter=d HEAD -- '*.rb' '*.rake' | tr '\n' ' ')
[ -z "$files" ] && exit 0

bundle exec rubocop --autocorrect $files >/dev/null 2>&1 || true

if ! out=$(bundle exec rubocop --format simple $files 2>&1); then
  printf 'RuboCop offenses remain — fix them (do not disable cops):\n%s\n' "$out" >&2
  exit 2
fi

A lighter alternative is a PostToolUse hook on Edit|Write that just runs bundle exec rubocop -a "$file" per edit — cheaper, but it only autocorrects and never blocks on what it can't fix. The Stop version above is what catches the offenses that need the agent's judgment.

What config to lint against — and what NOT to do

Use `rubocop-rails-omakase`: it is the Rails 8 default and the same 37signals house style the rest of this toolkit follows. Do not hand-author a sprawling custom `.rubocop.yml` — omakase's own README calls that bikeshedding, and a big bespoke cop set cuts against [[rails-philosophy]] §1 ("Vanilla Rails Is Plenty"). Two things to know about what omakase does and doesn't do:

• It only covers mechanical consistency — spacing, indentation, case/end alignment, double-quoted strings, trailing whitespace. It cannot express the taste rules in this skill (method ordering, caller-before-callee, expanded conditionals, CRUD-as-resources, bang / _later / _now conventions, thin controllers). So the hook and this skill are complementary, not redundant: RuboCop enforces the mechanical layer; the agent follows the taste layer because this skill is loaded.
• It disables the whole `Rails` and `Security` departments, so Rails/OutputSafety is off by default. To get the html_safe/raw XSS guard referenced below, re-enable a small targeted subset on top of omakase — not a full rule set:

# .rubocop.yml
inherit_gem:
  rubocop-rails-omakase: rubocop.yml

Rails/OutputSafety:   # html_safe / raw are XSS vectors — keep this on
  Enabled: true
Lint/Debugger:        # no stray binding.irb / debugger in committed code
  Enabled: true

The discipline (.claude/rules/rubocop.md)

A hook stops the agent from silently shipping offenses; a rules file stops it from cheating past them. Add .claude/rules/rubocop.md telling the agent:

• Fix offenses, don't hide them. Never reach for inline # rubocop:disable or # rubocop:todo to make a violation go away.
• If a cop genuinely doesn't fit this codebase, surface it in the final response and let the human decide (adjust the config vs. write a real fix) — don't silence it unilaterally.
• Never silence `Rails/OutputSafety`. html_safe and raw are XSS vectors; if a specific use is truly safe, surface it for the user to approve rather than disabling the cop. See [[rails-security]] for CSP/XSS context.

Beyond omakase: a custom cop for a rule it can't express

omakase covers mechanical consistency; it has no cop for a semantic rule like "never mutate params" (see [[rails-controllers]] Pattern 11). No upstream cop covers this either — rubocop-rails ships Rails/StrongParameters and Rails/StrongParametersExpect, but those are about permitting, not mutating. When you have a rule that specific and no cop exists, a ~40-line custom cop makes it enforceable. Drop this into your project:

lib/rubocop/cop/custom/params_mutation.rb:

# frozen_string_literal: true

module RuboCop
  module Cop
    module Custom
      # Flags in-place mutation of the request `params` object. `params` is
      # shared request state, not a scratchpad — build a separate hash instead.
      # Heuristic: flags a mutating call whose *immediate* receiver is a bare
      # `params`. It cannot follow aliases (`p = params; p.delete`) — static
      # analysis does not track that.
      class ParamsMutation < Base
        MSG = "Do not mutate `params`; build a separate hash instead."

        MUTATING_METHODS = %i[
          []= delete delete_if merge! update store deep_merge!
          except! extract! slice! compact! reject! select! keep_if clear
          transform_keys! transform_values! deep_transform_keys! deep_transform_values!
        ].freeze

        def_node_matcher :bare_params?, "(send nil? :params)"

        def on_send(node)
          return unless MUTATING_METHODS.include?(node.method_name)
          return unless bare_params?(node.receiver)

          add_offense(node)
        end
        alias on_csend on_send   # also catch `params&.delete(...)`
      end
    end
  end
end

Wire it up in .rubocop.yml (a project-local Custom/ department sidesteps any clash with the extracted Rails/ namespace):

require:
  - ./lib/rubocop/cop/custom/params_mutation.rb

Custom/ParamsMutation:
  Enabled: true
  Include:
    - app/controllers/**/*.rb

It flags params[:x] = …, params.delete, params.merge!, params.update, params&.delete, and the other bang mutators — while leaving reads, non-bang copies (params.except, params.permit(...).merge), and mutations of a plain hash or a .dup alone. Limitation: it only sees a bare params receiver; it can't catch mutation through an alias (p = params; p.delete), which needs flow analysis. Treat it as a high-signal tripwire, not a proof.

Share it across projects — a personal extension gem. A per-project Custom/ cop is fine for one repo, but copy-pasting it into every app rots fast. The durable home is a personal RuboCop extension gem — the same shape as rubocop-rails/rubocop-performance. This toolkit's owner keeps one at mickzijdel/rubocop-mick, where ParamsMutation lives as Mick/ParamsMutation. Consuming it is two lines — a github-sourced gem plus the plugins: key (the gem ships its cops' defaults, so no inherit_gem: is needed):

# Gemfile (dev group)
gem "rubocop-mick", github: "mickzijdel/rubocop-mick", require: false

# .rubocop.yml
plugins:
  - rubocop-mick

Update the cop once, bundle update rubocop-mick, and every project picks it up. In this owner's setup the dev-hooks dev-env-setup standard (v22+) wires those two lines into every Ruby repo automatically.

Upstreaming. Rails-specific cops live in rubocop-rails, not rubocop/rubocop (RuboCop core is pure-Ruby only), added under its single Rails department via bundle exec rake 'new_cop[Rails/ParamsMutation]'. It's a plausible contribution — the mutation methods are already deprecated on ActionController::Parameters — but maintainers weigh false-positive rate and general applicability, and the alias blind spot is a real weakness. Ship it in your own extension gem first; propose upstream once it's proven low-noise on a real codebase.

Summary Checklist

When writing or reviewing code, verify:

• [ ] Class methods come before public methods, which come before private methods
• [ ] Methods are ordered by invocation flow (caller before callee)
• [ ] Using expanded conditionals instead of guard clauses (unless at method start with complex body)
• [ ] No newline after private, content indented under it
• [ ] Non-CRUD actions are modeled as separate resources
• [ ] Controllers are thin, calling rich model APIs
• [ ] Bang methods only used when non-bang counterpart exists
• [ ] Async operations use _later/_now naming convention with shallow jobs
• [ ] ERB comments use <%#, never <% #
• [ ] Style is enforced on AI output via a RuboCop hook; offenses are fixed, never # rubocop:disable-d (and Rails/OutputSafety is never silenced)