The Claude setup checklist that gives you hours back

The hours do not come from finding a better app. They come from the fact that most people use Claude like a search box: new chat, explain everything from scratch, get something generic, fix it by hand. Do that ten times a week and you have spent your savings on setup costs.

Three moves fix it. Give it a memory so it stops starting cold. Hand over the tasks you repeat rather than the ones you agonise over. Let it build you small tools instead of doing small jobs manually forever. Work through the checklist in order and each move makes the next one cheaper.

Find your hours before you save them

Skip this and you will automate the wrong thing beautifully. Spend fifteen minutes listing what you actually did last week, then let Claude sort it into what should be handed over, what should become a tool, and what should stay with you.

I want to work out which of my recurring tasks are worth handing to you and which are not, before I set anything up.

Here is everything I did last week, in the messy order I remember it: [list your tasks, roughly how long each took, and how often it comes round: daily, weekly, monthly].

Sort every task into four buckets and explain your reasoning in one line each. Bucket one, hand to Claude now: repetitive, rule-based, and the output is text or analysis I check rather than trust blindly. Bucket two, worth a small tool: repetitive and mechanical enough that a script should do it, not a conversation. Bucket three, keep for now: it needs my taste, my relationships, or a judgment call only I can make. Bucket four, stop doing: it is not producing anything anyone uses.

Then rank bucket one by hours saved per hour of setup, tell me which single task to start with and why, and estimate honestly how many hours a week this realistically gives me back rather than the most optimistic number.

Give it a memory file

This is the highest-return fifteen minutes in the whole checklist, because it removes the re-explaining tax from every single future chat. The file describes you the way you would brief a new assistant on day one: what you do, how you sound, what you are working on, what you never say.

Do not write it from a blank page. Paste this and answer the questions.

I want a memory file that every future chat with you starts from, so I never have to re-explain myself. Interview me to build it, one question at a time, and do not move on until my answer is specific enough for you to act on.

Ask me about: what I do and who it is for, what I am working on right now, my voice when I write (and get me to paste two real lines I wrote so you can hear it), the three tasks I will bring you most often and what "done" looks like for each, where my files and numbers actually live, and the rules I never break. If I answer vaguely, push back once and ask for the specific version with an example.

When we are done, write it as one markdown document under clear headings, in my words rather than yours, short enough to stay readable (aim for under 200 lines). Finish with a list of the things you still do not know about my work, so I know what to fill in later.

Then put it somewhere Claude reads it automatically. Two paths, same file.

1. 🌱 Browser path: a Project. On claude.ai, open Projects, create one named after your work, then Set project instructions and paste the whole file in. Every chat started inside that project inherits it. Projects appear on free accounts too, capped at five (Anthropic's help centre), with the full walkthrough here.
2. 👩‍💻 Computer path: a CLAUDE.md. Save the same file as CLAUDE.md at the root of your work folder. Claude Code loads it at the start of every session in that folder, which the docs describe as "instructions you write to give Claude persistent context" (memory docs).
3. 📎 Add the proof files. Upload the things you would hand a new hire: your best-performing piece of work, last month's report, the template everyone uses. Claude reads these across every chat in the project.
4. ✅ Test it in one line. Start a fresh chat inside the project and type a single-line request with no context, like "draft the update for this week". If it comes back in your format and voice, the file is working. If it comes back generic, the file is too vague, so go add the missing detail.

Hand over the repeat tasks

A task you do weekly should not be a conversation you rebuild weekly. Turn it into a skill: a small file holding your process, which Claude loads on its own whenever your request matches. Mine draft captions, clean transcripts, and format scripts, and none of those are things I re-explain any more.

Turn the task I repeat most into a reusable skill so I stop rebuilding it every week. The task: [describe it, for example "cleaning a raw transcript into a readable draft in my format"].

Write me the SKILL.md. Include a short lowercase-hyphenated name (do not put "claude" or "anthropic" in the name, or the upload gets rejected), a one-line description saying what it does AND when you should use it, because that line is what makes you trigger it without me naming it, then the steps in my order, then the rules that must always hold, then exactly what the finished output looks like.

First ask me the three or four things you need: what I will hand you as input, the format of the finished thing, and the one mistake you must never make on this task. When it is written, package it as a zipped folder ready to upload and give me one example sentence that should trigger it.

1. ➕ Install it in the app. Open Customize, then Skills, then Add, then Upload a skill, and choose the zip. Skills are available on Free, Pro, Max, Team and Enterprise plans (Anthropic's help centre), so nothing to upgrade for.
2. 👩‍💻 Or install it in Claude Code. Drop the folder into .claude/skills/ inside a project, or ~/.claude/skills/ to have it available everywhere.
3. 🔁 Run it on a real task, then fix one thing. Do not tune it in the abstract. Run it, find the single most annoying difference from what you wanted, and tell Claude to change that one thing in the skill file.
4. 📈 Add the next one only when it hurts. One skill per week is a sane pace. Five well-tuned skills beat twenty half-written ones, for the same reason one tool set up properly beats twenty half-used ones.

Let it build you tiny tools

This is the move most people never make, and it is where the compounding shows up. When a job is mechanical rather than conversational, you do not want a chat about it, you want a small thing that does it. I asked for a content tracker and got the whole tracker, not instructions for building one.

Two versions, depending on your plan. Both start from the same description.

Build me a small tool for one job I keep doing by hand, and build the actual thing rather than telling me how to build it.

The job: [describe it exactly, for example "keeping track of every piece of content I publish, when it went out, which platform, and how it performed, so I can see what is working"].

Before you build, ask me what I want to put in, what I want to get out, and the one view I will look at most, because that view is what the whole thing should be organised around. Then build it as a single self-contained page I can open and use straight away, with sensible sample rows so I can see it working, and a way to get my data back out as a CSV. Keep it simple enough that I could ask you to change it later without you rewriting everything.

When it is done, tell me in two sentences what you built and the one thing I should try first.

The jobs worth turning into tools all look the same: mechanical, repeated, and boring enough that you keep putting them off.

The checklist, in order

Print this, tick it off, done in an afternoon. Nothing here needs you to be technical.

1. 🔎 Audit one real week. Run the audit prompt. Come out with one named task to start with and an honest hours estimate.
2. 🧠 Write the memory file. Run the interview prompt, save the result, and put it in a Project or a CLAUDE.md.
3. ✅ Test it with a one-line request. Fresh chat, no context, see if the output arrives in your voice and format.
4. 🛠️ Build one skill for your top repeat task. Run the skill prompt, upload it under Customize then Skills then Add then Upload a skill, and use it on something real this week.
5. 👩‍💻 Ask for one tiny tool. Pick the most mechanical job on your list and have Claude build the thing itself.
6. 🌱 Keep one habit forever. Every time Claude gets something wrong, add one line to your memory file instead of correcting it in the chat. That is the line that makes next week faster than this week.

The part most people miss

A twenty-tool stack does not compound, because every tool holds one twentieth of your context and none of them learn from the others. One tool with a memory file does the opposite: every preference you record, every skill you tune, and every tool you have it build all sit on the same foundation.

Which is why the move is not learning more AI tools. It is writing your work down once, putting it somewhere one good tool can read it, and then handing that tool more of your week.

Proof

Creating and managing Claude Projects
What Projects are, and the free-plan cap
Claude Code memory docs: what CLAUDE.md is
Creating custom skills, and the plans they are on
What artifacts are and how to use them
Claude Code setup, install commands and account requirements