RAILS · AI · 2026 — by Abraham Kuri (kurenn)

A workflow for Rails, built for the era of AI builders.

Fifteen years of Rails, distilled into a workflow. Six tools, one line, and Claude on the controls — the way I'd build today if I were starting from zero, and the way I build now after shipping for a decade and a half.

SCROLL

I · The Manifesto

Most AI workflows stop at suggestions. That's not enough anymore.

For fifteen years I've been writing Ruby on Rails. Long enough to remember when a senior engineer was someone who'd seen what happens to an app with no test suite, no migrations strategy, no idea what a callback chain costs. The trade was always the same: experience, paid in production scars, in exchange for the right to ship with confidence.

AI changes the trade. A solo developer in 2026 can move at the speed of a small team — but only if they stop using AI like a fancier autocomplete and start running it like a workforce. Models are not engineers. They're workers. They need scaffolding: a forge that lays the rails, an audit that calibrates risk, a pipeline that catches the obvious mistakes, a translator between vague intent and executable spec, a dispatcher that runs the team that builds it, and a loop that plans, reviews and grades the work before it ships.

That's what this toolkit is. Five Claude Code skills and one Ruby CLI — each the distilled version of a thing I'd previously do by hand, slowly, on every Rails project I touched. Plug them in and you stop doing AI by improvisation. You start running a crew that just works for you.

One developer. The leverage of a team. The same Rails you've been writing for years.

— Abraham Kuri (@kurenn) · 15 years of Rails

II · The Pipeline

Six stations. One direction.

Each tool does one thing well. The output of the upstream stage is the input to the next. Use them à la carte; or run the whole line and watch the work move down the track.

  1. 01 · Forge

    Railwyrm

    A Ruby CLI that bootstraps Rails apps with an opinionated default stack — PostgreSQL, Tailwind, RSpec, Devise, Brakeman, RuboCop, GitHub Actions — and an interactive wizard for magic-link, passkeys, CI and quality features.

  2. 02 · Assess

    Rails Audit

    An 18-dimension stability audit for the codebase you inherited (or the one you wrote and stopped trusting).

  3. 03 · Harden

    BooRails

    Script-first Rails skills that execute the workflow end-to-end and return evidence.

  4. 04 · Translate

    Prompt Refiner

    Casual intent in, executable spec out.

  5. 05 · Dispatch

    Roundhouse

    A Claude Code plugin that simulates a Rails dev team.

  6. 06 · Loop

    Dev Loop

    The virtuous loop that wraps the line — plan, build, adversarial review, an independent rating against the plan, gated fixes, then a PR.

III · The Toolkit

The six.

Each one is a Claude Code / Codex skill. Each one is on GitHub. Each one is open source. Scroll — each one gets its own scene.

01 · Forge

Railwyrm

A Ruby CLI that bootstraps Rails apps with an opinionated default stack — PostgreSQL, Tailwind, RSpec, Devise, Brakeman, RuboCop, GitHub Actions — and an interactive wizard for magic-link, passkeys, CI and quality features. New apps in minutes, not days.

github.com/kurenn/railwyrm railwyrm new

02 · Assess

Rails Audit

An 18-dimension stability audit for the codebase you inherited (or the one you wrote and stopped trusting). Severity-ranked. Evidence cited inline. A fix sequence ordered so each phase unblocks the next.

github.com/kurenn/rails-audit /rails-audit

03 · Harden

BooRails

Script-first Rails skills that execute the workflow end-to-end and return evidence. Diagnose → security → safety → quality gates. Not suggestions — runs.

github.com/kurenn/boorails /rails-framework

04 · Translate

Prompt Refiner

Casual intent in, executable spec out. Expands “add login” into auth flows, rate limits, validation rules and edge cases — so the agent that builds it isn't silently guessing.

github.com/kurenn/prompt-refiner-skill /prompt-refiner

05 · Dispatch

Roundhouse

A Claude Code plugin that simulates a Rails dev team. An Opus orchestrator triages each task into trivial / single-domain / cross-cutting, then dispatches Sonnet specialists with TDD and conditional gates. Benchmarked 7– 34× cheaper than the comparable swarm.

github.com/kurenn/roundhouse /rails-feature

06 · Loop

Dev Loop

The capstone that wraps the whole toolchain. An Opus orchestrator plans and writes the contract; a parallel Sonnet specialist army implements; Codex adversarially challenges the approach; a fresh independent Opus reviewer rates the work against the plan; Sonnet agents fix until the work clears an 8.5 quality gate. Learnings are captured, then — and only then — a PR is opened.

github.com/kurenn/dev-loop /dev-loop

IV · A morning, end-to-end

From rails new to a green PR — same morning, same Rails.

~/work · zshkurenn@studio
$ railwyrm new my_app --devise_magic_link  # Postgres · Tailwind · RSpec · Devise · CI  → my_app/  (Ruby 3.3 · Rails ~> 8.0 · magic links wired) $ cd my_app$ /rails-audit --quick                          # baseline, ~$0.10  → 0 blockers · 3 medium · ready to build on $ /prompt-refiner "add team accounts with invites + roles"  → spec.md  (Account · Membership · Invite · admin/member roles) $ /rails-feature spec.md                       # roundhouse dispatch  ▶ opus triage → cross-cutting · TDD · security gate ON  ▶ models · controllers · request specs · pundit policies  ✓ all gates green  · branch: team-accounts-2026-05-07 $ /rails-framework                               # boorails final pass  ✓ diagnose · ✓ security · ✓ safety · ✓ quality-gates $ gh pr create --fillhttps://github.com/kurenn/my_app/pull/1 $ 

V · Who built this

15 years on the rails.

Abraham Kuri (@kurenn) — Mexico City. Co-founder of Icalia Labs. I shipped my first Rails app in 2010 and have spent the years since auditing, refactoring and resurrecting Rails codebases for teams of two and teams of two hundred.

This toolkit is the residue of that work. Every dimension in rails-audit caught a real production fire on a real project. Every BooRails skill exists because I was tired of doing it by hand. The Prompt Refiner is what I wish I'd had every time a non-technical founder said “I want users to log in.”

VI · Get on the line

One marketplace. Six plugins.

claude code · plugin marketplace
$ claude plugin marketplace add kurenn/marketplace      # one-time per machine $ claude plugin install rails-audit@kurenn$ claude plugin install boorails@kurenn$ claude plugin install prompt-refiner@kurenn$ claude plugin install roundhouse@kurenn$ claude plugin install dev-loop@kurenn # Or install all five at once:$ claude plugin install roundhouse@kurenn boorails@kurenn rails-audit@kurenn prompt-refiner@kurenn dev-loop@kurenn # Railwyrm is a Ruby CLI — clone and run from the repo:$ git clone https://github.com/kurenn/railwyrm$ cd railwyrm && bundle install$ bundle exec ruby exe/railwyrm new my_app --devise_magic_link # Restart Claude Code. /rails-audit, /boo-framework, /prompt-refiner,# /rails-feature and /dev-loop now appear in the slash menu.

VII · Questions

FAQ.

Do I need all six?

No. Each tool stands alone. The most common starting point is rails-audit on a codebase you don't fully trust, then prompt-refiner the next time someone hands you a vague request. Roundhouse is what ties the implementation together once you have two or more installed, and dev-loop is the full virtuous cycle when you want the whole six working as one crew.

What models / clients does this support?

Claude Code only. The skills, plugins and CLI are built specifically against Claude's Sonnet/Opus models and Claude Code's plugin and slash-command primitives. No API key gymnastics — they use your existing Claude Code session.

Is this a replacement for Brakeman / RuboCop / etc.?

Emphatically no. Rails Audit invokes Brakeman, bundler-audit, RuboCop, reek, rails_best_practices, simplecov and friends. The point is that a 400-warning brakeman dump is not a report — it's evidence. The toolkit synthesizes evidence into something you'll actually act on.

What does it cost to run?

Sonnet/Opus token costs only — no subscription, no SaaS, no telemetry. A --quick rails-audit runs around $0.05–0.20. A full --standard pass on a 30–50 KLOC monolith is $0.20–1.50. Roundhouse benchmarks 7–34× cheaper than the comparable swarm — a trivial fix runs about $0.36, a cross-cutting feature about $2.65.

Why Rails specifically?

Because this is what I know cold. Fifteen years of Rails means fifteen years of knowing exactly what an N+1 looks like at 4am, where the Stripe-webhook idempotency bug always lives, why your Sidekiq retry policy is probably wrong. The toolkit encodes that, not generic LLM intuition.

Can I extend it?

Yes — every repo takes issues and PRs. The dimensions in rails-audit, the skill prompts in BooRails, the question batches in Prompt Refiner — all just markdown.