Stop Using AI Coding Agents Without Understanding These 5 Things

You ask your coding agent to add login, and it picks the wrong framework, pulls in a dependency you didn't ask for, and puts database queries straight into the controller. The model isn't bad at code. Nobody told it how your project works, and a better prompt only lasts for one task. This video goes through the five things you can give a coding agent besides the prompt: CLAUDE.md, skills, hooks, memory files and the prompt itself. It follows one agent and one app all the way through, adding authentication, and shows what each component fixes, using Claude Code as the worked example and noting where other agents differ. The thread through all of it is two questions: when does the agent see this, and who makes sure it actually happens? CLAUDE.md and skills guide the agent but don't guarantee anything. A hook runs because an event happened, not because the AI remembered. The video ends with all five working on one task, and a short test for deciding which one a new rule belongs in. CHAPTERS 0:00 Why a better prompt isn't enough 0:21 The five components 0:49 When the agent doesn't know your project 1:15 CLAUDE.md, the onboarding document 1:50 Context, not a guarantee 2:14 Skills, reusable playbooks 2:59 A code-review skill and how skills load 3:38 Why instructions aren't enough 4:10 Hooks and lifecycle events 4:42 Blocking dangerous commands and hook limits 5:20 Memory between sessions 6:18 Memory vs CLAUDE.md, and stale memory 7:03 Prompts, vague vs specific 7:57 All five on one task 8:58 Which component to use