# First Step
[](LICENSE) [](https://x.com/thenikacode)
**Give your agent a vague idea. It asks one question at a time, then hands you a small first step you can check.**
First Step is a skill for Claude Code and Codex. Each turn, it picks the one question that would change your first step the most. In the run above it asked one question, then gave the first step ([full transcript](examples/weekly-commit-update.md)).
## Install
In Claude Code 2.1.275 or later:
```
/plugin install first-step --marketplace thenikacode/first-step
```
Or from your terminal (tested with Claude Code 2.1.310):
```
claude plugin install first-step --marketplace thenikacode/first-step
```
<details>
<summary>Older Claude Code versions</summary>
```
/plugin marketplace add thenikacode/first-step
/plugin install first-step@first-step
```
</details>
With the skills CLI, for Claude Code, Codex and other agents that read skills:
```
npx skills add thenikacode/first-step -g
```
In Codex by hand, copy `skills/first-step` into `~/.agents/skills/first-step`:
```
git clone https://github.com/thenikacode/first-step
mkdir -p ~/.agents/skills && cp -R first-step/skills/first-step ~/.agents/skills/
```
## Use
Type the command, then your idea in your own words:
```
/first-step:first-step <your idea>
```
The command depends on how you installed it:
- Claude Code plugin: `/first-step:first-step`
- skills CLI, in Claude Code: `/first-step`
- Codex: `$first-step`
## What it does
The skill is one Markdown file of instructions, [SKILL.md](skills/first-step/SKILL.md). It tells your agent to:
- start from what you've told it and, when it can, from the project you have open;
- ask one question per turn, with two or three suggested answers when that helps, and take an answer in your own words;
- tell you when two of your requirements conflict, and ask which one comes first;
- stop asking once it can name a small first step and a way to check it, with no fixed number of questions;
- keep the result short, at most about 200 words for a small task;
- skip the questions when your request is already clear;
- write an assignment you can paste into another assistant, if you want to carry on there.
## What it doesn't do
- It has no memory between chats. Each run starts from the conversation you're in.
- It doesn't promise that your plan or business will work.
- Its instructions say that planning with it isn't permission to build, publish, buy or message anyone. The skill can't enforce that. You set what your agent can do in its own permission settings.
- It isn't meant for a clear task, a small edit, a translation or a factual question.
- Answers change from one model to another, and from one run to the next.
## Example
[A week of commits into a post](examples/weekly-commit-update.md) is the full run from the image above, unedited: Claude Code 2.1.310, 26 Oct 2026, on a made-up idea.
## FAQ
### Isn't this just a prompt?
Yes, a long one in a single file, so you can read all of it before you install it. Claude Code and Codex load the same file, and you don't retype it each time. I haven't compared it with a good hand-written prompt.
### It asked a bad question
[Open an issue](https://github.com/thenikacode/first-step/issues/new) and paste the conversation, for example when it asked two things at once or asked for something you'd already told it. Take out anything private first.
## License
MIT. Made by [Nika Code](https://github.com/thenikacode) with Claude Code and Codex. To contribute, see [CONTRIBUTING.md](CONTRIBUTING.md).
Follow the build on X: [@thenikacode](https://x.com/thenikacode).
Not affiliated with Anthropic or OpenAI.