AI 101 for your transition

A model that has read a great deal and knows little about you

Claude is a language model. It was built by having a program read an enormous amount of text and learn, in a statistical way, what tends to follow what. The result can write, summarize, explain, and reason through a problem in plain language. On its own, the model is not a search engine, not a database of regulations, and not a person who knows you.

Two facts matter more than any other.

First, treat it as knowing nothing about you. In a fresh chat with memory and chat search switched off, it starts empty, and what you told it yesterday is gone. Claude also has a memory feature, on by default for Free, Pro, and Max plans, and on paid plans it can search your past chats. With those on, it may carry facts from earlier chats into a new one, and a carried fact can be out of date. So put the facts that matter, like your current date, in every brief, and check the ones it uses. You can review or switch these off under Settings, then Memory.

Second, it can sound sure when it is wrong. A confident tone tells you nothing about whether the answer is right.

Three names to keep apart

• The model is the engine that writes the replies.
• The Claude app (claude.ai on the web, desktop, or phone) is where you chat with it. Depending on your settings it can also search the web, search your past chats, and use memory. Those tools add context you did not paste.
• Claude Code is a tool built on the same models that works in a folder on your laptop. It can read and edit files there and run commands, which is how it sets up your board. By default it asks before it edits a file or runs a command; read what it plans to do before you approve.

That is what the board is for. Muster Out keeps your dates, your path, and the regulation text on every item in one place. When you ask the board a question, your facts are already in the room.

Where to use a draft, and where to check

Used well, the model is a fast drafter, a patient explainer, and a second set of eyes on a plan. Used carelessly, it can hand you a wrong date in the same calm voice it uses for a right one. Anthropic's own guide on reducing hallucinations says even the most advanced models, Claude included, can sometimes produce text that is factually incorrect.

Good at

• Drafting. A leave request, a note to your commanding officer, a question for IPAC (your admin office). It gives you a first version in seconds and takes edits well. You read every line before it goes anywhere.
• Explaining. Paste a paragraph of an order and ask for it in plain words. You get a plain words version to read against the original. Whether the paragraph applies to you is a question for the source and the people who own it.
• Reviewing a plan. Lay out your timeline and ask what looks out of order or missing. Treat its list as leads to check. None of them is a finding yet.

Bad at

• Dates. It can produce a deadline that sounds right and is not.
• Paragraph numbers and form numbers. It knows what a citation looks like and can fill one in from memory.
• Whether a rule applies to you. Eligibility, waivers, and exceptions turn on facts in your record that it cannot see.
• Knowing when it is wrong. It can sound just as sure when it is wrong. Its stated confidence is not a check.

The four line brief

A vague brief is a common reason for a weak answer. The fix is the same one you would use with the new officer: tell them who you are, what you want, what you already have, and what a good result looks like. Four lines take about thirty seconds. A specific brief often makes the answer more relevant to you. It does not remove the model's limits, so a well briefed answer can still be wrong.

Two habits that pay off

• Ask it to ask you questions. "Before you answer, ask me what you need to know, then wait for my answers." This fills in context it lacks. It does not make the answer right, so the one rule still applies.
• Ask for the cite. "Which paragraph says that?" A cite tells you where to look. Open the source and read it, because a cite can be wrong too.

Your account, your data

Everything you type to Claude goes to your own Claude account, under the terms you agreed to when you signed up. Your Muster Out board lives in that same account. Anthropic hosts it under your login. The people who built the board have no account on it. The people who made this lesson cannot see your answers unless you choose to send them at the end, and that is a separate choice.

Anthropic publishes how it handles your conversations and the setting that controls whether they are used for training. Read it there, in their words, and set it the way you want. It can change.

Your own medical history, your dates, your leave balance: those are yours to put in or leave out. The board works with whatever you give it, and more detail gets you a more specific plan. It is your call, made with open eyes.

What deleting covers

• The board. Delete the page from claude.ai and delete the setup folder on your laptop. That removes the board.
• Your chats and Claude Code transcripts. These are separate. You delete them in Claude. Anthropic says a deleted chat leaves your history right away and its back end storage within 30 days. It keeps data longer in some cases: up to 5 years in de-identified form if you allowed training, longer if a chat is flagged for a safety review, and whatever the law requires.
• Everything else. Device backups, any text message you sent, and a lesson result you sent stay where they are until you delete them there.

Do these before your setup

Open Claude at claude.ai or in the Claude app on your phone or laptop, and start a new chat. Each drill takes about three minutes. First, walk through one check by hand.

I am a Marine separating in [year]. Here is a MARADMIN. Give me five plain lines on what it says, then one line on whether it might affect someone leaving active duty. If a date or a paragraph matters, quote it exactly from the text and give the paragraph number. Do not add anything the text does not say.

[paste the MARADMIN here]

I am a [grade] separating on [date]. Draft an email to my commanding officer requesting [number] days of leave from [start] to [end]. Respectful, plain, under 150 words, no bullet points. Then, in a second version, make it half as long without losing the dates.

I am a [grade], [resigning, retiring, or separating at EAS] on [date]. Here is my plan so far: [three to six lines]. Before you answer, ask me up to five questions you need answered, then wait for my answers. After that, tell me what I might be forgetting, in order of how much it could cost me. For each point, name who or what I should confirm it with: IPAC, my TAP counselor, a VSO, or a specific order.

What you will set up

Muster Out is a single page in your Claude account. Across the top sit your date, the countdown, and one line per area of work with a progress bar, and six tabs sit below them.

• Brief. Opens on a Do next panel: one recommended action, its deadline, who owns it, and what blocks it. Then the next ten items, the critical path, and the flags.
• Timeline. Every item on a calendar from today to your date, by area, with the days you will be in the office counted per month.
• Lines of effort. Every item, with the regulation text behind it one tap away. This is where you check things off and add your own.
• Leave. Your balance from your Leave and Earnings Statement (LES), what accrues, and where the plan leaves you on your date.
• Claim. The VA timeline, your condition list, and the evidence to gather, with no rating guesses anywhere.
• Contacts and sources. The offices, the links, and the weekly regulation watch.

Most items are rules counted from your date, and the board says which dates are the regulation's and which are planning targets. Change your date in the profile and the items tied to it recalculate. Dates fixed by people, like booked leave, a SkillBridge start, or an appointment, need your own review when your date moves. When you press Ask, the page sends your board to Claude on your account for that one answer, after you agree once.

Check your understanding

Eight questions. Six right is solid. Nothing leaves this page unless you choose to send it at the end.

One sentence