How to set up Claude properly: instructions, models and prompt tips

Claude is not just a chat window. Set it up properly, and you work ten times more efficiently than someone who just starts writing. This post shows what it takes: the right settings, the right model, and ten prompt hacks that save time every day.

The five interfaces

claude.ai Web and Mobile is the standard. For anything spontaneous, for projects, for uploads. The mobile app runs on the same account – chats are available on all devices.

Claude Desktop with Cowork is the locally installed app – a fundamentally different mode. Cowork stores processes, decisions and inputs permanently. It can also handle various requests or split complex requests into sub-tasks and agents. The best analogy: Chat is the brief conversation by the coffee machine. Cowork is the weekly workshop with several employees and specialists, source documents and the memo from the last sessions.

Claude Code is aimed at developers. Terminal-based, no IDE needed.

The Chrome Extension brings Claude to every website – in its own browser context with all cookies and logins.

Word and Excel add-in – Claude directly in Microsoft Office. Fun fact: the add-ins beat Microsoft Copilot by a wide margin.

Which model for what

Haiku 4.5 – fast and inexpensive, for simple repetitive tasks. Sonnet 4.6 – standard for 90 per cent of all tasks, default in Pro and Team plans. Opus 4.6 and 4.7 – for demanding tasks. Opus 4.7 uses a third more tokens than 4.6.

Licence – choose it right, once

Individual Pro: 20 dollars per month. Team Standard: 25 dollars per seat, minimum 5 seats. Premium Seat: 125 dollars per month for heavy users. Important: A private account cannot be converted – anyone who switches later starts from scratch.

Instructions – the most important setting

Instructions are permanently stored rules that are active in every chat. They are under claude.ai/settings/general. What belongs there: writing style, language, spelling, interaction style, guardrails. A useful rule: for current questions, Claude should first search the web – LLMs have a knowledge cut-off that can be months old. Instructions are not a finished document. They grow.

Ten prompt hacks from everyday work

1. Ask me 10 questions. For vague tasks, ask Claude to ask ten follow-up questions. The output becomes more precise.

2. Make 2 suggestions. See the range, then cherry-pick.

3. Pre-output check. Forces an extra review loop – not a luxury for quotes and calculations.

4. Edit mid-chat instead of continuing. Back to the last good prompt, new branch, confusing context falls away.

5. Cross-use with other LLMs. The same question in Claude and Gemini – often a better plan emerges.

6. Do not prompt angrily. Aggressive prompting leads to anxious outputs.

7. Build prompts like briefs. Persona, task, context, format.

8. No preamble. Claude gets straight to the point, without introductory phrases.

9. Be brutal, no yapping, Devil's Advocate. Claude challenges its own work rather than confirming it.

10. Think step by step. Errors become visible before they land in the output.

The key settings

Turn on response notifications. Turn on search in chats. Turn on long-term memory. Load tools when needed – noticeable from 10 connectors upwards.

Three terms for the next post

Connectors and MCP – interfaces that connect Claude directly to a CMS, newsletter tool, calendar or CRM. No more copy and paste.

Skills and plugins – skills are stored instructions for specific tasks. Plugins bundle several skills into a package.

Markdown and MD files – Claude processes MD files better than PDFs. Claude automatically reads CLAUDE.md files on start-up.

Details in the follow-up posts.

Next post: Connect Claude with Outlook, Gmail and 200+ tools – MCP explained.