Your first week with an AI agent – what to realistically expect

On Monday your AI agent runs its first analysis of your offers and finds the comments column before you’ve had a chance to point it out. On Tuesday it saves a file under a different name than yesterday, on Wednesday it asks for the third time how much energy you have today, and on Thursday you hit your usage limit before you’ve properly started. That’s what the first week looks like for many of the people we see in our courses, and none of it means something was set up wrong. Tuesday’s and Wednesday’s problems can each be fixed with one sentence that Claude Code adds to CLAUDE.md when you ask it to, and Thursday’s gets rarer once everything is set up.

Week one: fixing the rules, not yourself

Your first week with Claude Code is mostly about correcting the files your AI agent reads at the start of every conversation, most often CLAUDE.md. The people we work with in our courses usually blame their prompts, when what’s missing is a rule Claude Code had no way of knowing. Without it, it guesses – sensibly, just not always your way.

Tasks with a clear start and finish, on the other hand, work from day one: going through a large spreadsheet, a first analysis of your offers, a contract in two languages, a daily plan – and they usually take far less time than doing them by hand.

You explain the same thing to Claude Code several times, it agrees, and the next day it does it the old way. More than once in our courses, it turned out it had saved that rule in its auto memory rather than in CLAUDE.md. According to the documentation, that’s exactly what “remember this” does: it goes into auto memory, and at the start of a conversation Claude Code loads only the index of those notes, opening a note itself only when it decides it needs it. If a rule should apply in every conversation, tell Claude Code outright to add it to CLAUDE.md rather than to “remember” it.

An AI agent that does more than you asked

In its first week an AI agent is happy to do more than you asked for. You ask for two extra tasks for today and get a plan for the whole month. You ask for meeting notes, and Claude Code saves them under a new name and moves the folder somewhere else. In our courses we’ve also heard about research on a business contact that the AI agent wrote into a calendar event description – a description the invited client could see. Another AI agent was keen to connect a sales system that nobody had asked it to touch.

Part of this comes from the permission mode. In recent versions of Claude Code, a terminal session starts in auto mode by default: Claude Code doesn’t ask for approval, and a separate model judges risky actions, such as permanently deleting files, in the background. Auto mode stops risky actions, but it knows nothing about how you like your files named or how many tasks fit into your day. Claude Code learns that from a rule in CLAUDE.md, for example this one:

Add a rule to CLAUDE.md: before you rename a file, move or delete anything in this folder, connect a new tool, or add anything that isn't in my request, ask me for approval. Show me the wording of this rule first and don't save anything until I approve it.

Claude Code will show you the sentence it proposes, and once you say “yes” it adds it to CLAUDE.md. In later conversations it will have that rule in front of it from the first message. A rule in CLAUDE.md is still guidance that Claude Code tries to follow, not a lock – the documentation calls these files context, not enforced configuration. If you’d rather have Claude Code wait for your approval before most file changes, that’s what manual mode is for.

An AI agent that asks about everything

An AI agent can overdo it in the other direction too. In our courses we see this most often with daily planning: in every conversation the AI asks about your plan and how much energy you have today, even though you answered that in the morning. What’s missing is a rule that says when the question makes sense. People in our courses set their own rules: ask about invoices once a week, and about the daily plan only before ten in the morning.

List the questions you regularly ask me at the start of a conversation or during it. For each one, suggest when it's really worth asking and when you can skip it, and tell me which file these rules should go in. Don't save anything until I approve.

Claude Code will list the questions and propose a rule for each – usually for CLAUDE.md, or, when the questions come from a daily-plan skill, for that skill. Stopping before a file change or a terminal command is a separate matter: the permission mode decides that, not CLAUDE.md.

Using up your limit at the start

You hit your plan’s usage limit faster in the first week than later on, because everything is still being built: settings, CLAUDE.md, the first corrections to it, the first attempts with connectors. Each of these takes a long conversation, and with every message Claude Code sends the whole conversation so far. After a longer break, Claude Code processes the entire conversation from scratch on the first message, so resuming a few old, long conversations can eat a good chunk of your limit before the first new task even comes up. The cheapest way to start a new task is a new conversation, with the /clear command.

We recommend the Pro plan to start with, and the very first week may show you it isn’t enough. If you hit the limit regularly, it’s worth moving to Max. In our courses we see most people pay for Max for the first few months and go back to Pro once everything is set up. If you don’t have time to build out your AI agent yet, you’re unlikely to hit Pro’s usage limit.

Two ways to run out of time

Time can become a problem in two ways when you work with an AI agent, and both are normal. The first is obvious: ten-hour workdays, meetings starting first thing in the morning, and setting up Claude Code keeps getting pushed aside for something more urgent. In our courses we regularly meet people who come to the second session without having opened Claude Code once – and that’s a fine place to start too. For a week like that we recommend one habit: a daily plan with your AI agent, meaning one request in the morning to lay out your day. That’s exactly the task we give people in our courses for their first week.

The second one is more of a surprise. You see how much you can get done, every task suggests two more, and the clock says three in the morning. A plain list helps here: ask Claude Code to write tomorrow’s ideas into a file – they’ll wait there until morning, and you can go to bed.

Corrections to CLAUDE.md as the normal rhythm of the week

The Claude Code documentation suggests adding something to CLAUDE.md when Claude Code makes the same mistake a second time, or when you type a correction into the chat that you already gave in an earlier conversation. In the first week these moments come most often, because the file is new. That’s a good sign: every correction you add carries over into every conversation after it.

The simplest way is to turn it into a short end-of-day routine:

Go through this conversation and list the corrections and notes worth saving permanently in CLAUDE.md. For each one, suggest how to word it. Don't save anything until I approve.

Every rule in CLAUDE.md should also make clear who it’s addressed to. In one of our courses, the line “do not modify without asking”, added to the file as a reminder for a person, was read by the AI agent as a ban on itself, and it stopped updating that file.

CLAUDE.md shouldn’t keep growing forever, either. The documentation recommends keeping each file under 200 lines, because a longer one takes up more room in the conversation and Claude Code applies its rules less consistently.

Your first skill – how to tell it’s time

A skill is a recipe for a repeatable task, saved in a file, that Claude Code loads only when you call it or when it decides it needs it – unlike CLAUDE.md, which it reads in every conversation. The documentation gives two signs: you keep pasting the same instructions, checklist or multi-step procedure into the chat, or your CLAUDE.md has a step-by-step task description where a short rule should be. In our courses the sign usually sounds different: “my AI agent keeps forgetting this”.

A skill is best built from what already works. We’ve seen an AI agent offer to turn a course participant’s prompts into a skill on its own, and its version worked worse than those prompts did. That’s why we recommend going through the task with Claude Code first and writing the skill at the end – from the steps Claude Code actually took.

In short

When your AI agent does too much or asks too much, ask Claude Code to add a rule to CLAUDE.md, and when you want it to wait for your approval before changing a file, switch the permission mode to manual. When you find yourself asking for the same thing yet again, think about a skill. Higher usage at the start is the cost of setting up, and it drops once everything is in place. And when you’re short on time, start with a daily plan in the morning; when you’re short on sleep, with a list of ideas for tomorrow.

Most of this work is done by Claude Code itself, by the way: you say what should change, and it adds the sentences to its own files. The rule about asking before renaming a file is just one example – you can ask it to add other rules the same way. And on Friday you can give your AI agent a task that went well on Monday, and see whether, with a week’s worth of added rules, it does it more your way.