Your first run

Pick the bug you have stopped mentioning in stand-up: the one that shows up twice a month, vanishes when anyone looks, and has a workaround everyone quietly applies. That is a good first problem.

Try one skill on a real problem in your own project. This page uses Diagnose because it reads your project without editing it, starts only when you name it, and leaves a file you can open. Install the skills first: see Install First Four Hours.

Start Diagnose

You start a skill by name with /name : request, the form these pages use. If your agent names skills another way, use its form. With the Claude Code plugin, a skill's name starts with first-four-hours:, so Diagnose is /first-four-hours:diagnose.

For example (illustrative):

/diagnose : The nightly export writes an empty file about twice a month, still exits 0, and works when we rerun it by hand. Find out why.

Replace the request with a real problem from your project. Diagnose traces the symptom through the project and runs only probes that change nothing. It asks before using shared or billable resources or production data, and it never edits the project.

What you get

A sealed diagnosis.md in .first-four-hours/diagnoses/<slug>/. In a Git repository, Git ignores that folder through a rule the skill writes, so the file does not appear as a change to commit. It holds the symptoms, each finding with how well its cause is established, the limitations, and advisory directions. Advisory directions are leads, not a chosen fix.

To see it, run this from the project folder:

ls .first-four-hours/diagnoses

The run is done when the file is sealed. When the evidence cannot support a finding, Diagnose says what evidence is missing instead of guessing.

What to do next

Nothing starts on its own; you decide:

Plan a fix for one finding with Blueprint, then carry out the sealed plan with Construct. For example (illustrative):
```
/blueprint : Plan a fix for finding <finding-id> in .first-four-hours/diagnoses/<slug>/diagnosis.md.
```
Explore options first with Ideate.
Stop here.

If you want to learn a subject rather than change code, start with Tutor instead. For example (illustrative):

/tutor : I study history, and I am tired of people citing Gödel to prove things the theorem never said. Teach me the first incompleteness theorem from scratch, make me build the self-referring sentence myself, then show me popular claims about it and have me say which ones are wrong.

How skills start

Each skill has a mode that controls whether the agent may start it, recorded in the invocation policy.

Mode	Skills	How it starts
`both`	Every other skill	The agent may start it when your request matches its purpose, and you can start it by name.
`user-invoked`	`charter`, `diagnose`, `provision`, `ticket`	Only you start it, by name. The agent never starts it on its own.
`model-invoked`	`router`	The agent starts it when a task matches. Ask for the outcome in plain words instead of naming it.

Name a skill when you want to be sure it runs. On Claude Code and GitHub Copilot, the model-invoked skill is kept out of the skills you start by name; Codex and Cursor have no such setting. Each skill's page in the skill index says how to start it.

If the skill did not start

The agent does not recognize the name: recheck Install First Four Hours.
It is a model-invoked skill: ask for the outcome in plain words.
The agent handled the request without the skill you expected: start the skill by name.