Claude gives generic answers when it knows nothing about you, your entity, or your standards, so it answers the way a capable stranger would. Setup closes most of that gap. The core pass takes about 45 minutes, once; a project per client and a CLAUDE.md folder take an afternoon more.

To set up Claude properly, work through 6 things in order: choose the plan your data allows, switch off model training before any client material goes in, write your standing rules into Instructions for Claude, build one project per recurring body of work, check what memory holds, and brief each task the way you would brief a capable new colleague. If you work in the desktop app's Code tab, add a CLAUDE.md file to each job's folder. The settings paths below are Anthropic's as of October 2026, checked against its help centre on 31 October 2026.

Choose the plan your data allows

For a finance professional the first row to read is model training, not usage. Anthropic's list prices in US dollars, before tax:

PlanPriceWhat it addsModel training (Anthropic's label)
Free$0Sonnet and Haiku, memory, connectors, file creation, up to 5 projects"Opt-out"
Pro$20 a month, or $17 a month billed annually ($200 up front)Opus, Claude Code and the desktop Code tab, Research, the Excel, PowerPoint, and Word add-ins"Opt-out"
Max$100 a month, monthly only5x Pro's usage"Opt-out"
Ultra (formerly Max 20x)$200 a month, monthly only10 times Pro's weekly usage, and no 5-hour limit on verified accounts"Opt-out"
TeamStandard seat $25 a month ($20 billed annually), Premium seat $125 ($100 billed annually), 2 to 150 seatsShared projects, admin controls, SSO"None by default"
EnterpriseUS$20 per seat a month, billed annually, plus usage at API rates; 20-seat minimumSCIM, audit logs, Compliance API, custom retention"None by default"

Usage resets on a rolling 5-hour window, paid plans add a weekly limit, and every app draws on one pool. For an individual, Pro is the sensible start, monthly at first. For a practice handling client data under engagement letters, Team fits better: by default Anthropic does not train on Team or Enterprise inputs or outputs, except feedback you submit. One point for UK readers: Anthropic stores commercial data in the US and documents no EU or UK storage option. If you are still choosing between vendors, ChatGPT vs Claude vs Gemini sets the three side by side for finance work.

Day one: switch off training, then check two switches

On Free, Pro, Max, and Ultra, open Settings > Privacy and switch off "Help Improve our AI models" before any client or confidential material goes in. Switching it off stops future use; anything already used in a training run stays there. One exception applies whatever the setting: a thumbs up or down sends the whole conversation to Anthropic, kept for up to 5 years and usable for training, so do not rate a chat that holds client figures.

For a chat that should stay out of history and memory, such as a personnel matter, start a new chat outside any project and click the ghost icon, upper right. Incognito chats are not used for training, though Anthropic keeps them for 30 days.

Then open Settings > Capabilities and check that the code execution and file creation switch is on (the default on Free, Pro, Max, Ultra, and Team). Without it you get a description of a spreadsheet instead of the spreadsheet: Excel, Word, and PDF files, artifacts, skills, and .xlsx uploads all depend on it. Anthropic warns that hidden instructions in a file could trick Claude into sending data out through the sandbox's network access, so for regulated work ask your administrator to keep network access off or limited.

The day-one pass, in one table:

SettingWhere, as of October 2026What to do
Model trainingSettings > Privacy, "Help Improve our AI models"Off before client material goes in
Instructions for ClaudeSettings > Your Claude (Settings > Account in the older layout)Write your standing rules
MemorySettings > Your Claude, "Memory from chats" (Settings > Memory, "Generate memory from chats", in the older layout); topics under Manage next to Your memoryConfirm it is on; read the topics
Code execution and file creationSettings > CapabilitiesCheck it is on
UsageSettings > UsageCheck before a long job

Write Instructions for Claude (the old profile preferences)

What used to be profile preferences is now Instructions for Claude: a block of standing guidance applied to every conversation, in a project or out of one. Click your initials in the lower left corner, choose Settings, then Your Claude, and find the Instructions section. Tone and format rules belong here too, since Anthropic's personalisation page no longer lists Styles.

The test for what goes in: will it still be true in 6 months? Your role, your readers, and the shape you want answers in pass; one client's particulars belong in a project.

This starter is adapted from The Claude Quick Start. Fill the brackets, cut what misses, and paste the rest:

I am a [financial controller] at [a 200-person group with 4 UK
subsidiaries]. The work I bring you most weeks is [month-end
commentary, board pack drafts, and policy reviews].
Write in [British] English. Keep answers short; expand when I ask.
Lead with the answer or the recommendation, then the reasoning. Use a
table when you compare options, numbered steps for a process.
Most of my output is read by [the CFO and the audit committee], so keep
it plain and specific. Rough notes I paste are raw material to work
with, not prose to critique.
Flag anything you are unsure of. Do not invent figures, client names,
or outcomes; ask me for the real ones instead.

"Do not invent figures" will not stop every confident wrong answer, but it turns many into a question back to you. Then, in a new chat outside any project, paste the Quick Start's two-minute test:

From my standing instructions, describe the business you think you
are working for in two sentences, then name the three jobs I should
hand you this week and say what makes each one a fit.

A reply that names your role and work you recognise means the block is carrying; a bland one means it did not save or is too vague.

Build one project per recurring body of work

A project holds instructions, files, and its own memory for one body of work. Go to claude.ai/projects and click "+ New Project". Write the standing rules under "Set project instructions" and add files with the "+" button in the project knowledge panel on the right.

Three rules decide whether a project helps:

  • Claude cannot see the project's name or description. A rule typed into the description does nothing, and nor does Claude know which entity the project serves unless the instructions say so. Name the client or entity in the first line of the instructions.
  • Instructions are what you would tell a new hire on their first morning; knowledge is what you would hand them to read. The readers, the house rules, and what to do when a figure is missing are instructions. The policies manual, last quarter's signed-off pack, and the chart of accounts are knowledge.
  • One project per body of work whose instructions would differ: a client, an entity, a reporting cycle. If you catch yourself typing "except for the [X] work", the [X] work wants its own project. Free allows 5 projects.

Chats in a project do not share context with each other unless the material is in project knowledge. On paid plans, as knowledge nears the context limit, Claude switches to retrieval, which Anthropic says expands capacity by up to 10 times; name the document you mean when one matters.

The confidentiality point most guides miss: memory made inside a project can be referenced in your chats outside projects. For a confidential client, turn off "Use account memory" on the project's page, which keeps that project's memory separate in both directions.

Before real work goes in, ask Claude in the project for 4 bullets: who the client is, how the output should read, what it must never do, and which files it is drawing on. A vague answer usually means rules stranded in the description.

Check memory, and read what it keeps

Memory is the only layer you do not write: Claude saves topics as you chat. It is on by default on Free, Pro, Max, and Ultra: open Settings > Your Claude and confirm "Memory from chats" is on. On Team and Enterprise an owner enables it first, and it then stays off for each member until they switch it on.

Choose Manage next to Your memory to see every topic, and edit or delete any that are wrong. Pause keeps what is stored but stops Claude using or adding to it; Reset permanently deletes everything, project memory included. To run one chat without memory, switch off Memory in the "+" menu before the first message.

Because memory fills itself, it drifts: a topic saved in March about a restructuring still shapes answers in October. Read the topics list once a month.

CLAUDE.md, if you work in the Code tab

CLAUDE.md belongs to Claude Code: the terminal, the IDE extensions, and the Code tab of the Claude desktop app, on every paid plan. No Anthropic page says an ordinary chat on claude.ai reads it, so chat rules stay in Instructions for Claude and projects.

It is a plain-text file of standing instructions at the top of a folder, read at the start of every session there. In the desktop app, click Code, choose Local and the job's folder, and type /init: Claude drafts a first file for you to correct. Keep it under about 200 lines, as Anthropic advises. A personal file at ~/.claude/CLAUDE.md holds what is true of all your work, the project file what is true of this job, and Claude joins the two, reading the most specific last. Add a line when you type the same correction you typed last session, which is Anthropic's own test.

Two cautions. Anthropic calls these files "context, not enforced configuration": Claude tries to follow them, and where two instructions contradict each other it may pick one arbitrarily. To block an action outright, it points to a hook. And check the permission mode selector under the prompt box: choose Manual for client folders, so Claude asks before each edit or command.

The Code tab matters for confidential files. Since 6 October 2026, new tasks in ordinary conversations on Pro, Max, and Ultra run in Anthropic's cloud and fetch a copy of any local file they use; a Local session works on files where they sit, though what Claude reads is still sent to Anthropic for processing. The Claude Quick Start lays out a one-folder-per-job structure and a worked CLAUDE.md for one engagement.

Brief Claude the way you would brief a new colleague

With standing context in place, the brief carries only what is true today: this entity, this file, this deadline. Anthropic's help article says: "describe what you need, the way you'd brief a colleague".

The Quick Start breaks a brief into 4 parts:

  1. Context. Your role, who the output is for, and which attachment is which: the part people skip, and the one that moves the answer furthest.
  2. Role. One line that sets judgement and register, in the guide's words "act as a second adviser".
  3. Task. The finished state, not a procedure, with a rule for gaps, for example "where the file does not explain a movement, say so; do not infer a cause".
  4. Format. Length, structure, and whether it comes back in the reply or as a file, for example "a one-page Word document, recommendation first".

"Summarise this variance analysis" returns a tidy digest; a brief that names the board as reader, sets a materiality threshold, and forbids invented causes returns a draft you can review against the workbook.

Choose the model from the menu at the start of each chat: Haiku 5.5 for quick, high-volume jobs, Sonnet 5.5 for most drafting, and Opus 5.5 for judgement-led or accuracy-critical work. Leave effort on "Default".

When a draft misses, steer it in the same chat: a new chat gives you a fresh context window, not a fresh usage allowance. "Too hedged; give a recommendation" moves a draft; "make it better" only reshuffles it.

What setup does not do

Every layer above is text Claude reads and tries to follow; none of it checks the output. A well-set-up Claude still states a wrong figure in the same steady prose as a right one, so the review step stays with you: tie every figure to a cell before a draft leaves your desk.

Some work carries a formal duty as well. Anthropic's Usage Policy, effective 12 November 2026, requires a qualified person to meaningfully review a high-risk recommendation before it is delivered or acted on, with personalised investment recommendations and advice on an individual's own taxes among the named examples, and the person who receives it must be clearly told AI was used. Internal drafting and analysis sit outside it.

Where to go next

The Claude Quick Start is the 50-page version of this post, free with your email address: the 45-minute setup click by click, projects, CLAUDE.md, how the layers stack, and the briefing method. If you are an operator new to Claude, The Claude Foundations Playbook ($29, 152 pages) carries the setup into growth workflows and your website. To run Claude itself as a system, with the Claude app and Claude Code, memory, connectors, and hooks, The Claude Power-User Playbook ($47, 224 pages) goes the full distance.

Sources

All checked on 31 October 2026 unless stated.