description: Use when Claude is open and you are still doing all the remembering yourself.
surfaces: Chat, Projects, Cowork, Claude Code
---
Prompting is not the skill. Setup is.
Claude isn't underperforming. It's under-briefed, and you're re-briefing it by hand every single session. Twenty steps, one afternoon, then your prompts get to be one line.
## stop
Three prompts that guarantee a mediocre answer.
"Write me a newsletter about AI."
No context.You gave it nothing, so you got the statistical average of the internet. Everyone else got it too.
"Act as a world-class instructional coach with 20 years of experience who studied under..."
Role-play is not context.It still has no idea who you are, what you ship, or what you'd reject. A costume is not a brief.
"Make it punchier."→ "No, more professional."→ "Try a different tone."
You can't describe what you want.So it keeps guessing. Twelve rewrites later you are back at draft one.
## chat + projects
Where most of the leverage is, and where most people stop.
01Personal preferences are for preferences only. Tone, format, hard nos. Three lines, not your life story.
02Turn on Connectors and code execution and file creation. Skills need that second one to run at all.
03One Project per recurring job. Its instructions and knowledge files load in every chat inside it.
04Write Project instructions like a job spec: what it's for, your voice, the output format, and what to reject.
05Skills are on every plan, including Free. Customize → Skills → Create skill, then upload a zipped folder holding a SKILL.md.
06The description field is the whole game. Name the exact phrases that should trigger it. Vague description, never fires.
07Put the rules that matter at the top of the skill body. Anything buried at the bottom gets skimmed.
## cowork
Same engine as Claude Code, no terminal. Pro, Max, Team, Enterprise.
08Cowork is a mode, not a separate app. Find the message box, switch from Chat to Cowork.
09In your first session type /setup-cowork. It picks a role plugin and connects your tools so you don't start from an empty room.
10Use Work in a folder and scope it tight. One project folder. Never your whole home directory.
11Cowork Projects carry their own files, instructions, and memory. That's the container, not a chat thread.
12Schedule the repeating work. Scheduled tasks run remotely with your machine closed.
13Revoke folder and connector access when the task is done. No standing permissions.
## claude code, week one
Never opened a terminal? These seven steps are the whole beginner path.
14Install it, then cd into a project folder and type claude. It logs you in through the browser once.
15It only sees the folder you launched it in. Standing in the right directory is half the job.
16Run /init. It reads the folder and writes a CLAUDE.md. Read what it wrote before you keep it.
17CLAUDE.md loads at the start of every session. Conventions, quirks, non-obvious commands. Keep it short: it competes with your files for context.
18Start any message with # to append a line to CLAUDE.md without stopping what you're doing.
19Ask for a plan before it edits anything. Read the plan. Then let it run. This is the single biggest novice upgrade.
20Commit to git before you turn it loose. That is your undo button, and you will need it in week one.
## red flags
Signs you're still working like it's 2025. No particular order.
!Your prompt opens with "Act as a world-class..."
!You re-upload the same three reference files every session.
!You paste an about-me document in by hand. That is a skill.
!Twelve chats open, zero Projects.
!You built a skill and it never fires. The description is too vague.
!Code execution is off, so the skill you built can't actually run.
!You're doing ten-step file work in Chat instead of Cowork.
!You pointed Cowork at your entire home directory.
!You ran Claude Code in a folder with no git history.
!Your CLAUDE.md is 400 lines and quietly crowding out your actual code.
!You re-explain your tone every single time.
!Connectors are off, so it can't see the thing you're asking about.
One file, four surfaces
A skill is a folder with a SKILL.md in it. The same file works in Chat, in Cowork, and in Claude Code at .claude/skills/name/SKILL.md. Write it once and stop maintaining four versions of your own voice. Skills you upload are private to your account. Check model names before repeating them, too: Anthropic ships Opus 5, Sonnet 5, Haiku 4.5, and Fable 5. There is no Opus 4.8.
Prompting isn't dead. It got shorter. The work moved into the context it lands in.