How to work with Claude

The first session, and the thing you notice here sooner than elsewhere

As with any other chat-based tool, start the first session with a real job and write in that first message what the job is, who it is for, what constraint applies and the format you want back. Up to here nothing differs from the ChatGPT lesson and the same rules work.

The difference starts where your work is long. Here, instead of being a message inside the chat, the output opens in a side panel as a document, and you edit that document, send it back, keep one part and change another. For a thousand-word text that is the difference between editing and getting it again from scratch. Anthropic pricing page lists this on the free plan too, along with creating files, executing code and searching the web.

A habit worth having in that first session: instead of saying "make it better", say which part and why. The sentence "the second paragraph is a slogan, replace it with a real example" is an edit that can be carried out; "make it better" gets you a full rewrite that usually ruins the good part as well.

And a warning that becomes useful sooner than you expect: a long conversation is not free. Every turn re-reads the whole conversation, so a chat with thirty messages gets slow and the model starts ignoring what you said in message three. When you see that, change conversations and take a summary of the decisions with you.

Conversation, project, preferences: which one is for what

The four concepts you read in the previous lesson do exactly the same work here and only their names differ: the conversation is the container for a job, context is what gets re-read from the top every turn, an attachment is a file that gets read without becoming permanent knowledge, and user preferences are the standing instructions pinned to every conversation. Anthropic pricing page lists user preferences on the free plan too.

The fifth concept added here is the project: a workspace that keeps several conversations and documents under one goal and carries its own shared instructions. On that same page Anthropic states that projects are not on the free plan and that Pro gives you unlimited projects. So if you are on the free plan, fill the place of a project with a simple habit: one text file on your own computer holding the fixed decisions of the job, pasted once at the start of every new conversation.

The practical difference between a project and a conversation in one sentence: a conversation is the memory of a job, a project is the memory of a client. If you do ten jobs for one client across three months, there are things you have to repeat every time (tone, prohibitions, product names) and those are what a project is for.

One thing none of these three is, and people mistake it: none of them is your filing system. Chat history is a convenience, not a place of record. Move the output you needed out the moment it is useful.

Five jobs whose output is a document, with ready prompts

These five deliberately differ from the five in the ChatGPT lesson: those were short and frequent jobs, these are jobs whose input or output is long. Swap what is inside the braces for your own.

One. Rewrite a services page under a real constraint. A services page is usually three years old and full of claims you can no longer stand behind.

The text below is my current services page. Rewrite it under these
constraints:
- Keep no claim that has no number or example behind it in my text.
- Delete every unsupported claim and list separately what you deleted.
- Cap: {400} words.
- Add one section titled "who this service is not right for".
Output: the rewritten text first, then the list of deletions.
Current text:
{the page text}

Two. Summarise a long document with references. What separates this from "summarise it" is the word reference: a summary with no references cannot be checked.

Read the attached document and build a summary in this shape:
one point per line, and at the end of each line, in brackets, say which
section or page of the document it came from.
If something is not in the document but you think it matters, put it in
a separate list titled "not in the document".
State no number that is not in the document.
My question about this document: {for example, what am I committing to}

Three. Turn meeting notes into a project specification. Nobody enjoys this job and its absence gets expensive later.

These are my scattered notes from a meeting with a client. Build a
project specification with these headings:
goal, scope, out of scope, deliverables, assumptions, open questions.
Under "assumptions" put anything that was needed to write the document
and was not in the notes, and say explicitly that it is an assumption.
Do not write a date or an amount that I did not write.
Notes:
{meeting notes}

Four. Review your own text for unsupported claims. Before publishing, these ten minutes have caught a wrong claim many times over.

Your role is reviewer. Read the text below and extract only three things:
1) Sentences making a numeric or time claim whose source is not in the text.
2) Sentences that create an obligation (guarantee, promise, always, never).
3) Sentences that can be read two ways.
Rewrite nothing. Give the sentence number and the reason only.
Text:
{your text}

Five. Build a decision table from several options. For choosing a host, a tool or a supplier. The point is in the last line of the prompt.

These are the {three} options I have looked at, with the information I
gathered myself. Build a comparison table with the columns:
option, cost, main limitation, who it suits.
Use only the information given below. For any cell whose information I
did not give, write "unknown" and do not fill it.
After the table write one paragraph: with this information, what is still
missing in order to decide.
Options:
{option details}

Write "unknown" and do not fill it is the most important line in all five. An empty cell in a comparison table means you do not know something, and that is itself information. A cell filled by guesswork destroys that information and looks like fact while doing it.

How far the free plan goes, and where Iran hits a wall

Everything below comes from Anthropic own pricing page and was read from it again on 6 September 2026. Contrary to the usual assumption, the free plan here is not a trial stub: chat on web, mobile and desktop, web search, memory across conversations, creating and executing files, extended thinking, skills, connectors, incognito chats and user preferences are all ticked in the free column.

The Pro column works out to 17 USD a month on an annual subscription, billed as 200 USD up front, and 20 USD if you pay month to month. The Max plan starts from 100 USD a month and Anthropic describes it as a choice between 5x or 20x the usage of Pro.

And now the thing that is not in the table and is the most important line in this section: Anthropic publishes no number for the usage ceiling of the free plan. The pricing page says only that usage limits apply, without saying how many messages, over what window, or with what reset. We write no number either, because there is no number to write. If you find an exact figure for that ceiling somewhere, it is either somebody guess or one account experience on one particular day.

On Iran the situation is clear and unambiguous, and it was read off Anthropic own page: that page carries two separate lists, one for commercial API access and one for Claude.ai, and Iran is on neither. So there is no official signup and no Iranian card is accepted. We wrote this up with its source on the Claude page in our AI reference, and the payment layer separately in the guide to buying AI tools. We do not sell accounts and we do not suggest ways around anything; all we write is what the official position is and where it is stated.

Where your data goes, and which switch you have to flip yourself

Anthropic writes this up in its own privacy centre and that article is dated 16 March 2026. The summary for the consumer plans, meaning Free, Pro and Max: your conversations are used to improve the models when you have allowed it yourself, or when a conversation is flagged for safety review, or when you have explicitly joined a programme. The setting sits in the privacy section of your account and it is called model improvement.

Three details that are written down in few places and all three matter to you.

First, when that option is on, what gets used is the entire related conversation, together with its content, custom styles and conversation preferences. The unit is not the message; it is the conversation.

Second, raw connector content is excluded. If you have connected Google Drive or an MCP server, the raw content coming from there is not in that category, but the same article states plainly that if the same content is copied directly into the conversation, then it is. This is the detail people get wrong: connecting is not the same as pasting.

Third, incognito chats. The article says incognito chats are not used to improve Claude even if you have model improvement enabled. If you have one sensitive job and do not want to go into settings, that is the shortest route.

And one thing nobody expects: the thumbs up and thumbs down button. According to that same article, when you give feedback the whole conversation is stored in Anthropic systems for up to five years, although it is de-linked from your user ID before use and not combined with your other conversations. So that button is not free; do not press it on a conversation with something sensitive in it.

The rest is the usual hygiene and has nothing to do with this particular tool: a unique password, two-step login, and a periodic look at connections and signed-in devices. Plus the rule no setting replaces: do not put anything into a chat that you would not hand to a stranger.