Setting up Claude like a Pro: Guide to customise chatbot to be your personal assistant

Over the past few months, Claude has seen a notable hike in its consumer base, and the company has been eager to tune it well for professional usage. Hence, it becomes more crucial for you to set it up just right, especially for your work desk. 

But how do you set up Claude? Isn’t it just a generative AI-powered chatbot keen on responding to your queries? 

What most of us often miss out on is the ability to tailor Claude to our needs and liking. Claude, like the rest of the competition, can be tailored to specific professional needs, and all it takes is spending some time and effort to tell the chatbot how you want your AI assistant to serve you.

Hence, in this article, we explore how you can extract the most out of your Claude account and work more efficiently in a corporate workplace scenario.

But first…

Why Claude doesn’t understand your exact needs?

Out of the box, Claude knows nothing about you. It doesn’t know whether you’re a tax lawyer reviewing a cross-border acquisition agreement or a product manager writing a sprint retrospective. It doesn’t even know whether you are a 35-year-old person or a kid. Without that context, it prefers to produce responses calibrated for an average person. The answers you see are mostly accurate and balanced, but maybe always slightly off-target for your actual needs.

While many AI experts recommend refining your prompts, the real fix requires substantial effort. Instead, it requires you to tell the AI bot who you are and what sort of help do you seek. And this should be a one-time thing, not a process you do for every session.

Claude offers two native mechanisms to implement this – a system-level “user preferences” setting that applies globally, and project-specific custom instructions that apply only within a defined workspace. 

1. The global identity layer

Think of your global preferences as a permanent brief that Claude reads before every conversation. This basically tells Claude everything it needs to know about you ahead of every session. 

To access it, open Claude.ai, navigate to Settings, and find ‘User preferences.’ Note that what you write here is appended invisibly to every session.

For example, if you write a global preference note that looks like this, it is less likely to work: “I work in finance and want concise answers.” While it’s not wrong, it’s not giving Claude enough information about you to work with.

Following is an example of a well-written one that gives Claude more information to work with:

“I am a senior financial analyst at a mid-size manufacturing company. My role involves FP&A, board-level reporting, and investment scenario modelling. I use Excel and Power BI daily.

Communication style: Write for a technically literate audience. Skip preamble. Prefer numbered steps or structured tables over walls of text. When I ask about a concept, assume I know the fundamentals and go straight to the nuance.

Output defaults: Always flag assumptions you’ve made. If a task involves numbers, show your working. If I ask for a document, default to a format I can paste straight into Word.”

What to include: 

– Your role and seniority. 

– Your tools and technical stack. 

– The audience you write for. 

– Your preferred output format. 

– Any persistent constraints (e.g. “we follow IFRS accounting standards” or “all external communications must use gender-neutral language”). 

– No long introductions, unnecessary caveats, excessive hedging.

Usually, the limit of four paragraphs is the sweet spot. Less than that and Claude lacks enough information, more than that risks overwhelming the context with boilerplate rather than genuinely useful information.

2. Claude Projects

Claude’s Projects is where you can truly customise the chatbot. A project is a persistent workspace – a folder that retains memory of uploaded documents, conversation history, and custom instructions across sessions. Unlike a single conversation, it doesn’t forget when you close the tab.

The correct mental model is to treat each project as a dedicated specialist you’ve hired for one domain. You can have a Legal Reviewer, a Financial Modeller, a Communications Editor, and many more. Each of these ‘experts’  lives in its own project with its own persona, its own reference documents, and its own output conventions.

Here is how to create it:

Step 1 : Create the project and name it by function, not topic 

“Contract Review Specialist” or “Board Memo Drafter”. The name shapes how you and your colleagues will use it.

Step 2: Write a project-level system prompt 

Under Project Settings, go to Custom Instructions and write a persona prompt. Be explicit about domain knowledge, output format, and behaviour. Example for a legal workspace:

“You are a contract review specialist for an FMCG company operating across GCC markets. Your job is to identify risk clauses, flag non-standard terms, and produce a structured risk summary for the legal team.

For every contract I share: (1) Identify the document type and governing law, (2) Produce a table of key terms, (3) Flag any clause deviating from standard market practice, (4) Give a risk rating: Low / Medium / High with one-line rationale.

Do not paraphrase contractual language — quote directly from the document. Do not offer legal advice — identify, summarise, and flag only.”

Step 3: Upload your reference corpus 

Every project accepts uploaded files that Claude can reference throughout all conversations in that workspace. Upload your style guide, templates, brand guidelines, standard contract library, and company glossary. This is what separates a generic AI from your company’s specialist. It can basically read the documents that a new hire would spend months absorbing.

Step 4: Run an onboarding test conversation 

Before sharing the project or using it full-time, you should test it with a representative real task. Notice where it drifts from your expectations, since you would need to refine the system prompt accordingly. 

3. The prompt architecture

Once your context layers are in place, the quality of your results totally depends on how you structure each request. Structure every non-trivial request using four parts:

Role / Context / Task / Format

Example:

Role: You are acting as my CFO, preparing for a board presentation.

Context: Our Q2 revenue missed the target by 8%. The board is conservative and risk-averse. I have 90 seconds on this slide.

Task: Write the narrative for the revenue variance slide. Lead with what we can control going forward, not the miss itself.

Note that the format should stick to just three bullet points, each under 15 words. Try using plain language and avoid jargon.

Once you start following this structure in your prompts, the quality of Claude’s output improves dramatically, since you essentially ask Claude to stop guessing.

For anything that requires multiple passes — for eg, a market analysis, a policy document, a detailed financial model summary – you should break the task into sequential requests. Ask Claude to create an outline first. Approve the structure. Then ask it to develop each section. This keeps you in editorial control and produces a better final product than a single sprawling prompt ever will.

4. Automating workflows

The email summary manager: Paste your inbox summary or a batch of forwarded emails into a “Communications Manager” project with instructions to categorise by urgency, identify action items, and draft a response for each that requires one. 

The meeting memo manager: Paste raw meeting notes, however messy, and ask your “Documentation Specialist” project to produce “a one-paragraph executive summary, a decision log, and an action item table with owners and deadlines.” The format is consistent every time because the project instructions define it once.

The first-draft manager: Whether it’s a board paper, a performance review, or a project update, upload your previous version as a reference document in the project and ask Claude to write to that standard. It calibrates to your organisation’s actual voice, not a generic AI tone.

Tip: Never paste sensitive or classified data, such as client names, deal specifics, personal employee information, or legally privileged material, unless your organisation has confirmed its Claude deployment is covered by an enterprise data privacy agreement. Prioritise keeping crucial information anonymous.

Conclusion

It is crucial to note that setting up Claude properly gives you an edge in making the most out of the AI bot. Additionally, you should look out for the official documentation, workshops, and prompt engineering guides from Anthropic at no cost, which could make your workflow easier.