For people building a product with AI

You can build now. Next you need a way to run it.

Request, story, scenario in plain language, task. With that structure, AI implements without guessing and you focus on what to build. Traditional task trackers were not designed for building together with AI.

Start by writing down one feature you have in mind.

PM on Rails task detail. The parent story and the acceptance scenario it came from on one screen. Request to task, connected in a single structure.
terminal
npm install -g pm-on-rails
pm-on-rails login
pm-on-rails mcp install

Connect Claude Code

login opens your browser for approval and stores the token in your local keychain. You need a workspace account to sign in. Restart Claude Code and tools like these are available.

Tools you will use most

start_task
Start a task. Returns the branch name and the line to paste into the pull request
read_scenario
Read an acceptance scenario (Given / When / Then)
get_traceability
Trace a task back to its request, and forward to what a change would affect
See every tool
(01)

The problem

Building with AI, does this sound familiar?

What you want to build is scattered across chat logs

You can no longer tell which thread decided what.

You lose track of what you asked AI for and what is finished

Features that seem to work mix with features you have not touched.

It runs, but you have not checked that it is right

There are no acceptance criteria, so there is nothing to check against.

(02)

How it works

Even alone, it runs as if you had a PM.

Write what you want to build and AI splits it into requests and scenarios. The agent reads them, implements, and the results are recorded against the scenarios.

  1. 01
    PM on Rails user stories. Stories split from a request, each with its acceptance criteria. What was scattered across chat logs fits into this structure.

    Write what you want to build

    Write down an idea and AI splits it into a requirement card, user stories and acceptance scenarios (normal, error and boundary cases). Anything vague stays behind as a question.

    You: Write

    About stories and scenarios
  2. 02
    PM on Rails Claude Code and Codex integration. The coding agent reads the scenario and acceptance criteria before implementing. What you asked for stays recorded as a task.

    Claude Code reads the scenario and implements

    Connected over MCP, the agent reads the scenario and acceptance criteria before it starts writing. When it finishes, the pull request completes the task and it moves to the next one.

    You: Watch

    Developing with Claude Code / Codex
  3. 03
    PM on Rails test results. Pass or fail and evidence per scenario. “It works” is no longer verbal; it is a result against the acceptance criteria.

    Passes are recorded against scenarios

    CI results are kept per scenario. Instead of “it seems to work”, you can say “this scenario passed”.

    You: Check

    About test results
(03)

The loop

From request to passing test, the same loop every time.

You put in what to build and leave the implementation to the agent. The tool runs the steps in between.

  1. 01RequestNotes and asks
  2. 02ScenarioAcceptance criteria
  3. 03BuildClaude Code / Codex
  4. 04TestCI results recorded
  5. 05FeedbackQuestions and decisions
Where you are involved
Where AI and the tool take over
(04)

From early users

What surprised the people who tried it.

I had started building but had no idea how to manage it. Existing task trackers did not fit working with AI.
A solo product builder
Scenarios are in plain language, so I can check them myself. Having CI passes recorded automatically helps.
An engineer at a development agency

Summarized from interviews with early users.

Structure makes you faster

Give your vibe coding a structure.

Free to start. Write down one feature you have in mind and watch it become a scenario.