Explore the guides +
Learn/Lesson 1 of 6
A PRACTICAL GUIDE / 01

Your first useful agent session.

Install a skill, open a project, and make one small change you can actually review.

You don’t need a complicated setup to get value from an agent. I’d start with a project you know, a small task, and a clear way to tell whether it worked. Here’s the whole loop.

01

Start in a project you recognize.

Use a skill-compatible coding agent such as Claude Code or Codex. You’ll also need Node.js and npm available to run the npx installation command. The skill supplies a workflow; it doesn’t install the coding agent or provide model access.

Open your project directory. Check the working tree before you begin so you can distinguish your own edits from the agent’s work. If you use Git, a separate branch makes the result easier to review.

TERMINAL
$ git status

Run terminal commands in your shell. Prompts later in this guide go into the coding agent.

02

Install one skill first.

Start with implement-feature. Run this from the project directory. Follow the installer’s prompts to select your coding agent, then start a new agent session so it can discover the installed skill.

TERMINAL
$ npx skills add akdevsaha-dev/factory-skills --skill implement-feature

For Claude Code, append -a claude-code. For Codex, append -a codex. Add -g if you deliberately want a global install rather than a project install.

03

Pick something small enough to inspect.

A missing empty state is a better first task than rebuilding the entire dashboard. Name the behavior you want and point to a similar part of the app if you know one. Replace the example’s screen names with screens your project actually has.

SEND TO YOUR AGENT
› /implement-feature add an empty state to the invoices list when there are no invoices. Match the empty state in the projects list. Keep the current loading and error behavior.

If your agent doesn’t expose this as a slash command, say “Use the implement-feature skill to…” followed by the same request.

04

Read the result before asking for the next thing.

Look at the changed files. Did the agent reuse an existing component? Does the empty state appear only when the list is empty? Does the report identify the checks it ran?

Try the affected screen yourself, or ask the agent to show the runtime evidence. If a check couldn’t run, decide whether you need to resolve that before accepting the change. Don’t treat “looks good” as the same thing as a passed check.

TERMINAL
$ git diff

Your first successful session is a change you understand and can verify, not a large amount of generated code.

Examples are requests to adapt, not recorded outcomes.Project source