Chapter Lab

When should I use a chat, a Project, a skill, or an automation?

I spent a long time guessing. The thing that helped was doing the job once in a chat, then saving the motion. Answer a few questions for this job. Take a skill that can decide the next one.

Who: Anyone who keeps opening a blank chat for work that already happened last week.
What: A short quiz plus one skill. The skill asks the same questions the next time, in Claude, Cursor, or Codex.
Why: A Project, a skill, and an automation are different jobs. Picking the wrong one wastes a week.
When: Before you build a skill or an automation for work that still changes every time you run it.

Field notes

The quiz is here if you want a recommendation now. The notes are the longer version: what I do first, and what each option is for.

What I learned the slow way

When I started using AI heavily, I kept guessing. Chat, Project, skill, automation. The answer depends on the job, and it depends on you.

Your use case first

Some people like every related chat in one Project. I sometimes keep a long thread out of a Project on purpose. The Project tries to read the whole thing, and the next ask gets slow. That is a taste call. The tools change. Try both if you are unsure.

Do the job once

The thing that unlocked skills for me: do the work in a chat first. For Author29 I needed to redesign a company logo and leave with the assets in a format I would actually use. Once that pass worked, I told the agent: take everything we just did, including the mistakes, and turn it into a skill. Now I can run that shape again.

Then write it down

Do not start by authoring a skill from a blank page. Start by getting one good result. The skill is the notes from that result.

What each one is for

Chat

One-off questions, no files to reuse. Fine for you. A weak place for weekly work more than one person must reuse.

Project

Same briefs or how-to docs, sitting where the next chat can see them. Use this when the files matter more than the exact wording.

Skill

The steps. Voice, order, and a review rule that should not drift. You still press go. Travels into Claude, Cursor, or Codex.

Automation

A trigger starts it: a new email, ticket, file, or a schedule. Write the steps first. Do not let high-risk work ship without a person.

A shared Project in fifteen minutes

One Project, one job

Do not dump every team into one space. Meeting notes and contract review are different places.

Files, then eight lines

Add 2–3 examples of a good answer. Write role, do, do not, and how it should look.

Name an owner

Someone updates the examples when the work changes. Without that name, the Project goes stale.

Tell people where to start

One sentence in Slack: start here. Do not open a blank chat for this work.

Chat, Project, skill, or automation

A chat you start. A Project with the files. A skill with the steps. An automation that starts when something happens.

Chat

One-off. Nothing to reuse. You will start it.

Project

Same files every time. People should start in one place.

Skill

Same steps every time. You still press go.

Automation

It should run when something happens, not when you remember.

Question 1 of 6

Who needs the output?

Take it with you

Keep the decision

This is the thing you keep. One skill that asks the same questions the next time, in Claude, Cursor, or Codex.

How to install it

01Download the Claude skill.
02Open Claude. Start a Project for this job, or open one you already use.
03Paste the file into Project instructions, or add it as a skill / reference file in that Project.
04Start a new chat in that Project.
05Describe the job, then ask it to use the ai-home-decision skill.

What is in the file

# When should I use a chat, a Project, a skill, or an automation?

Help one person put one recurring job in the right place. Works in Claude, Cursor, Codex, or any tool that can follow a skill file.

## The four options

- Chat: one-off, nothing to reuse, they will start it.
- Project: same files every time. People should start in one place.
- Skill: same steps every time. They still press go.
- Automation: it should run when something happens, not when they remember.

If the work is legal, money, or safety, and no reviewer is named, recommend they keep it out of an unattended run.

## Ask these, in order

1. Who needs the output? Just me / my team / several departments
2. Same source files every time? No / sometimes / yes
3. How often? Once / weekly / daily
4. If it is wrong, how bad? Annoying / customer-facing / legal, money, or safety
5. Must the wording and steps stay consistent? No / yes
6. Should this run when something happens, without opening a chat? No / yes

Do not skip ahead. Do not invent answers.

## How to decide

- Trigger yes + high risk → skill a person still runs. Do not automate shipping.
- Trigger yes + it comes back + risk is not high → automation, with the steps written first.
- High risk, solo, one-off → probably not this tool, or a watched chat.
- Same files, more than one person → Project.
- Same steps, more than one person, it comes back → skill.
- Solo, one-off, no files → chat.
- Solo, but it comes back → personal Project or personal skill.

## Output

- The home, in one line
- Why, in three short bullets
- How to set it up, in four steps or fewer
- A starter they can paste into that home

## Voice

Talk like a colleague. Do not call it housing. Do not say the advice has no fluff.