# The kickoff prompt

> Chapter 1 of [Shipping with AI](/guide). The first prompt does most of the steering.

My first prompt is long enough that I saved it as a [Raycast snippet](/uses). Three keystrokes expands into the full Claude Code kickoff: research deeply, plan in phases, then fan the work out to subagents.

The rest changes with the job. I give the agent an outcome, one user action, a definition of done, and any constraints it can’t discover in the repo. Then it reads the nearby code and docs before it proposes anything.

The plan is my first review. I correct the wrong assumptions, cut work, and move the phases around before any code exists. Once it starts building, my prompts get much shorter: keep this, remove that, open the page, run the check.

Types, lint, git, logs, and repo instructions hold the context between those prompts. If I have to explain the whole job again, something went wrong in the plan. The parts I type twice live on [Agent Skills](/agent-skills).

**Read the long version:** [Agent Skills](/agent-skills) · [Skills are the new packages](/writing/skills-are-the-new-packages) · [Ship fast or die](/writing/ship-fast-or-die)

## Continue

[Taste is the job](/guide/taste-is-the-job) →