Exercises

Exercises for the workshop. We'll work through these together.

1. Trying HITL Ralph

The intent of this exercise is to give you a feel for how Ralph works.

• Go to the ralph-workshop-001 repo
• Check out the main branch

git checkout main

• Open the plans/prd.md file to look at the shape of the PRD (don't read the whole thing!)

• Run the ./plans/once-claude.sh script (you might need to make it executable)

chmod +x ./plans/once-claude.sh
./plans/once-claude.sh

• Observe the output and see what happens

2. Adding Tracer Bullets To Our Prompt

Here we're trying to steer Ralph towards better task prioritization using some old techniques.

• Open the ralph-workshop-001 repo
• Check out the main branch (or start from where you left off)

git checkout main

• Open the plans/prompt.md file

• Add a section to the 'Task Selection' section telling the AI in which order it should implement features in the PRD

• Consider using the 'tracer bullets' approach to prioritize getting feedback quickly

• If you get curious about my solution, check out the tracer-bullets branch

git checkout tracer-bullets

3. Using Skills To Steer Ralph

Let's see how good Ralph is at invoking its own skills and how useful skills are in steering Ralph passively.

• Open the ralph-workshop-001 repo
• Check out the skills branch

git checkout skills

• OR if you want to retain your code, merge the skills branch in:

git add . && git commit -m "My awesome work so far" && git merge skills

• Run the ./plans/once-claude.sh script

• While it's running, open the .agents/skills/writing-tests folder

• Check out the files in the folder and see what they contain

• Check the logs to see if the LLM has invoked the skill, and if it changed anything about its behavior

4. Ralph In A Frontend Feedback Loop

Let's check out a front-end feedback loop using the Playwright MCP server with Ralph.

• Open the ralph-workshop-002 repo
• Check out the main branch

git checkout main

• OR if you want to try it with a more realistic, half-finished app, try the run-through branch:

git checkout run-through

• Run npx playwright install chrome to install the Chromium browser

• Run ./plans/once-claude.sh to run Ralph in a frontend feedback loop

• Observe how connecting it to the Playwright MCP Server changes its behavior

5. Building A Red Green Refactor Loop

Let's try another old technique and use a red-green-refactor loop with Ralph.

• Open the ralph-workshop-001 repo
• Check out the skills branch (or start from where you left off)

git checkout skills

• Open the plans/prompt.md file

• Rewrite the 'Execution' section to describe a red green refactor loop.

"Red" means that the agent should write a failing test first. "Green" means that the agent should write the code to make the test pass. "Refactor" means that the agent should refactor the code to make it better.

• Run the ./plans/once-claude.sh script

• Observe the output and see what happens

• If you get curious about my solution, check out the tdd-prompt branch

git checkout tdd-prompt

6. Running Ralph On Your Backlog

Let's try building a backlog and hook up Ralph to that backlog.

• Open the ralph-workshop-001 repo
• Check out the backlog branch. This is a branch where the app has been built up significantly, but has some issues.

git checkout backlog

• Open the plans/flaky-geo-tests.md file and read it.

• Create an issue in the issues folder that describes the problem (imagine this is a GitHub issue - we're not using GitHub here just for ease of setup)

• Run the ./plans/once-claude.sh script

• Observe the output and see what happens

7. Create A Refactor Plan

Let's try creating a refactor plan and putting in our backlog to see what happens.

• Open the ralph-workshop-001 repo
• Check out the backlog branch (or start from where you left off)

git checkout backlog

• Open the plans/repo-cleanup-request.md file and read it.

• Run Claude or OpenCode and initiate the /request-refactor-plan skill.

• Walk through the refactor plan dialogs

• Once complete, run the ./plans/once-claude.sh script

• Observe the output and see what happens

8. Create A PRD

Let's use my write a PRD skill to see how Ralph does when it encounters a PRD in the backlog.

• Open the ralph-workshop-001 repo
• Check out the backlog branch (or start from where you left off)

git checkout backlog

• Open the plans/webhooks-feature-request.md file and read it.

• Run Claude or OpenCode and initiate the /write-a-prd skill.

• Walk through the PRD dialogs

• Once complete, run the ./plans/once-claude.sh script

• Observe the output and see what happens