How to Set Up Claude for Better AI Chats
In the last guide, I walked through how to set up ChatGPT for better AI chats.
Claude deserves its own version.
Not because Claude is better at everything. Not because ChatGPT is worse. These tools are starting to overlap in obvious ways, but they still have different personalities, workflows, and strengths. Claude tends to feel especially useful when I want thoughtful writing, long context, project organization, document work, coding help, research, and a calmer back-and-forth.
The setup matters.
If you only use Claude as a blank text box, it is still useful. But the real power shows up when you understand the pieces around the chat: model choice, effort level, extended thinking, Projects, Skills, Artifacts, Connectors, Research, and Cowork.
That sounds like a lot. It is not as bad as it sounds.
Think of Claude as having different rooms:
- Chat is the quick conversation room.
- Projects are the focused work rooms.
- Artifacts are the workbench where drafts, code, diagrams, and small apps can live beside the chat.
- Skills are reusable instructions and workflows Claude can load when the job calls for them.
- Connectors let Claude work with outside apps and data.
- Research is the deeper source-gathering mode.
- Cowork is the agentic desktop workspace for longer, multi-step work.
The goal is not to use every room every day.
The goal is to know which one to walk into.
Start With The Job
Before picking a model or turning on a feature, decide what kind of job you are asking Claude to do.
Most Claude tasks fall into a few buckets:
- Quick help: rewrite this, summarize that, explain something, brainstorm ideas, clean up notes.
- Careful reasoning: debug code, compare options, analyze tradeoffs, plan a project, think through a hard decision.
- File work: read PDFs, analyze spreadsheets, create documents, generate slides, turn messy notes into usable output.
- Research: gather current information, compare sources, produce a cited report, update outdated material.
- Reusable workflow: use the same brand voice, report format, checklist, or process over and over.
- Connected work: search Drive, read emails, check calendars, create tasks, work with GitHub, Slack, Linear, Notion, or other services.
- Long-running work: organize files, create a report, process data, build a deck, or run a recurring task.
That classification tells you what to use.
If it is a simple rewrite, regular chat is probably enough. If it needs recent information, use web search. If it needs a source-backed report, use Research. If it needs a repeatable process, use a Skill. If it needs files and longer execution on your computer, look at Cowork.
Do not start with the fanciest tool.
Start with the smallest tool that can do the job well.
Choosing A Claude Model Without Getting Lost
Claude's model names and availability change over time, and different plans may show different options. So I would not recommend building your whole workflow around memorizing the latest label.
The stable mental model is this:
- Haiku is the fast, lighter option.
- Sonnet is the everyday balance of speed and intelligence.
- Opus is for harder reasoning, coding, analysis, and long-horizon work.
- Newer or restricted models may appear depending on plan, rollout, safety settings, and organization controls.
As of the official Claude Platform model overview I checked on July 1, 2026, Anthropic lists Claude Sonnet 5, Claude Opus 4.8, and Claude Haiku 4.5 in its latest model comparison. The same docs also describe Fable 5 and Mythos 5, but Claude's Help Center says access to Fable 5 was suspended on June 12, 2026. That is exactly why I would treat the model menu as current-account-specific.
Use what is available in your account today.
For most people, the practical setup is:
- Use Sonnet for normal work: writing, planning, summaries, everyday coding, and general productivity.
- Use Opus for hard work: complex reasoning, long documents, tough code, strategy, careful review, or agentic workflows.
- Use Haiku when speed matters more than depth.
- Use higher effort or extended thinking when the task is difficult, not just because the button exists.
Claude's model menu sits next to the send button. Anthropic's Help Center says that menu controls three things: which model you are using, how much effort Claude puts into each response, and whether extended thinking is enabled.
That is an important difference from just "pick a model and go."
Use Effort Like A Gear Shift
Claude has effort levels on supported models. According to Anthropic's Help Center, higher effort means more thorough responses, but it also takes longer and uses more tokens. So yes, you can make Claude think harder. You can also burn through limits faster if you use the heaviest mode for everything.
Use effort like a gear shift:
- Low: fast routine tasks, simple rewrites, summaries, and basic answers.
- Medium: everyday planning, outlines, structured writing, and light troubleshooting.
- High: the best general balance for important work.
- Extra high: long-running coding and agentic tasks on supported newer Opus models.
- Max: correctness-critical work where you are willing to wait.
For a basic email draft, Low or Medium is fine.
For a blog outline, Medium or High is probably right.
For a technical architecture decision, a dense legal-ish contract summary, a difficult bug, or a multi-source research synthesis, use High or higher.
The mistake is treating Max like a badge of seriousness. Sometimes it is just the slow lane with more tolls.
Use Extended Thinking When The Problem Needs A Plan
Extended thinking lets Claude spend more time breaking down a problem before answering. Anthropic describes it as useful for complex reasoning tasks that do not necessarily require recent web information, like math, debugging code, optimization, detailed document analysis, and multi-step planning.
Use extended thinking when:
- The answer needs a careful chain of decisions.
- You want Claude to compare approaches before choosing one.
- You are debugging something with multiple possible causes.
- You are asking for a plan with risks and tradeoffs.
- You need careful analysis of a document, spreadsheet, or technical problem.
Do not use extended thinking for:
- "Rewrite this paragraph."
- "Give me 10 title ideas."
- "Summarize this short note."
- "What is a good subject line?"
The point is not to make every answer feel heavier. The point is to give Claude more time when the task genuinely benefits from it.
Set Up Account Instructions
Claude has account-wide instructions called "Instructions for Claude." Anthropic's personalization docs describe these as general guidance Claude should consider across conversations.
This is where you put stable preferences:
- Who you are.
- What kind of work you usually do.
- How technical you want responses to be.
- Your preferred tone.
- Formatting preferences.
- Whether Claude should ask questions or make reasonable assumptions.
A good setup might look like this:
I am Colin Michaels, an applications developer and personal blogger.I prefer practical, grounded answers with clear next steps.For blog posts, use a personal, lightly funny voice without sounding corporate.For technical work, explain tradeoffs and give concrete implementation guidance.Use official documentation for product claims when possible.Ask clarifying questions only when the missing information would materially change the result.
Keep account instructions broad.
Do not stuff every project detail into them. That belongs in Projects.
Use Memory Carefully
Claude's memory can use chat history to build continuity across conversations. Anthropic says memory is available across Claude web, Desktop, and Mobile, and that project memory is separate from standalone chat memory.
That can be very useful.
Memory is good for:
- Your role and common projects.
- Your writing preferences.
- Your coding style.
- Your usual workflows.
- Your communication preferences.
Memory is not a dumping ground for private information. Claude also has incognito chats, which Anthropic says are not saved to chat history, are not used for memory, and are not used for training. They are still retained for a period of time for safety or organization policy reasons, so "incognito" does not mean "no record exists anywhere."
My practical rule:
- Use memory for stable preferences.
- Use Projects for specific ongoing work.
- Use incognito for sensitive one-off conversations.
- Avoid pasting passwords, private keys, confidential client data, health records, financial identifiers, or anything you would regret storing.
This is boring advice until the day it is not boring.
Use Projects For Work That Has Continuity
Projects are one of Claude's best features for real work.
Anthropic describes Projects as self-contained workspaces with their own chat histories and knowledge bases. You can upload documents, add text or code, and set project instructions. Free users can create a limited number of projects, while paid plans unlock expanded project knowledge through RAG when the knowledge base grows.
Use a Project when the work has context you will need again.
Good project examples:
- AI article series.
- Blog drafts and source notes.
- A coding project.
- A business plan.
- A research topic.
- A health-tracking journal.
- A course or study plan.
- A YouTube or podcast workflow.
- A family planning folder.
For this Claude article series, I would make a project called AI Tool Guides and add instructions like this:
Purpose:Help draft practical articles that teach non-expert readers how to use AI tools more intentionally.Voice:Personal, grounded, useful, lightly funny, and not hype-driven.Rules:- Prefer official documentation for product claims.- Clearly separate sourced facts from opinion.- Include screenshot suggestions when a UI step would help.- Keep examples realistic.- Do not oversell AI as magic.Output:Use blog-ready Markdown with title, excerpt, headings, source links, image direction, and a screenshot checklist.
Then add useful files: the beginner AI post, the ChatGPT guide, source notes, screenshots, style examples, and any future article ideas.
That way, every chat in the project starts with the right background.
Use Artifacts As The Workbench
Artifacts are one of the features that make Claude feel different.
Anthropic's Help Center describes Artifacts as standalone content Claude can show in a dedicated window beside the conversation. Claude may create an Artifact when the content is substantial, self-contained, reusable, or something you are likely to edit later.
Common Artifact examples include:
- Markdown documents.
- Code snippets.
- Single-page HTML websites.
- SVG images.
- Diagrams and flowcharts.
- Interactive React components.
In plain English: if the output is something you want to keep working on, ask Claude to put it in an Artifact.
Examples:
Draft this as a Markdown artifact so we can revise it section by section.
Create a simple interactive checklist artifact for this workflow.
Turn this process into a flowchart artifact.
Artifacts are especially good for article drafts, diagrams, one-page tools, checklists, small prototypes, and working documents. They keep the big piece of work from getting buried in the chat scroll.
Claude can also create and edit downloadable files such as Excel spreadsheets, Word documents, PowerPoint decks, and PDFs when code execution and file creation are enabled. That is a separate but related capability, and it is useful when you need an actual file instead of a preview beside the chat.
Use Skills For Repeatable Work
This is where Claude gets especially interesting.
Anthropic describes Skills as folders of instructions, scripts, and resources that Claude loads dynamically for specialized tasks. They use progressive disclosure, which means Claude can look at the available skills, decide which ones are relevant, and load only what it needs instead of stuffing everything into the context window.
That is a very good idea.
Use a Skill when you have a repeatable workflow:
- Apply a writing style.
- Follow a brand guide.
- Create a recurring report.
- Format meeting notes.
- Generate a proposal.
- Review code using team conventions.
- Analyze data using a standard method.
- Build a presentation from a known template.
Projects and Skills overlap, but they are not the same.
Use a Project for context.
Use a Skill for procedure.
For example:
- Project: ColinMichaels.com Blog
- Skill: Blog Post Package Creator
The project stores the posts, notes, screenshots, and working context. The skill teaches Claude the exact repeatable steps: title, excerpt, categories, tags, SEO metadata, draft, image prompt, sources, and CMS JSON when needed.
That is better than re-pasting the same process every time.
Anthropic's skill creation docs say the best skills are focused, have clear instructions, include examples when helpful, and define when the skill should be used. A simple skill can be just a skill.md file. A more advanced one can include reference files, resources, and scripts.
Start simple.
Do not create one giant "do everything for my life" skill. That sounds efficient until it becomes impossible to reason about. Focused skills compose better.
Use Connectors Slowly
Connectors let Claude access apps and services, retrieve data, and sometimes take actions. Anthropic's connector docs give examples like Linear, Slack, Google Drive, GitHub, and Google Workspace.
This is where Claude becomes more useful and more serious.
If Claude is just drafting text, the risk is limited. If Claude can read your Drive, search email, create calendar events, write to Linear, or interact with Slack, you need to be much more deliberate.
My connector rules:
- Connect only what you actually need.
- Review what the connector can read and write.
- Use approvals for anything that changes data.
- Disconnect services you are no longer using.
- Avoid connecting sensitive accounts just to experiment.
- Be explicit in the prompt about what Claude may and may not do.
Anthropic says connectors inherit your permissions from the connected service. If you cannot access a file or channel in the source system, Claude cannot reach it through that connector either. That is good. It also means Claude can reach things you personally can reach, so you still need judgment.
Claude also has tool access modes for connectors:
- Auto: Claude decides what to load, and this is the best starting point for most users.
- Always available: useful when you have fewer than 10 connectors that you use constantly.
- On demand: better for large connector libraries or when conversations are hitting length limits.
That matters because every connected tool takes up room in the conversation. More tools is not automatically better.
Sometimes the best setup is one carefully chosen connector, not a dozen shiny ones waiting to confuse the chat.
Web Search, Research, And Extended Thinking Are Different
Claude has a few features that sound similar until you use them.
Use web search when you need current information and a few sources are enough. Anthropic says web search is best for straightforward factual queries that can be answered with one or two tool calls. It can also fetch specific URLs when web search is enabled.
Use Research when you need a more complete source-backed answer. Anthropic describes Research as agentic: Claude conducts multiple searches that build on each other, explores angles automatically, and delivers thorough answers with citations. Research can also work across connected sources like Gmail, Google Calendar, and Google Docs when those are connected.
Use extended thinking when the problem needs reasoning, not necessarily new information. It is for hard thinking, planning, debugging, analysis, and multi-step decisions.
The quick guide:
- Normal chat: explain, draft, brainstorm, revise.
- Web search: check current facts.
- Extended thinking: reason carefully.
- Research: gather and synthesize multiple sources.
- Connectors: use your apps and data.
- Cowork: run longer, multi-step work on your desktop.
Do not use Research for every question. It can use more of your limits because it gathers more sources. Save it for work where sources matter.
When To Use Claude Cowork
Claude Cowork is the more agentic side of Claude.
Anthropic describes Cowork as bringing Claude Code's agentic capabilities into Claude Desktop for knowledge work beyond coding. It is available on paid plans through Claude Desktop. Instead of answering one prompt at a time, Cowork can work on complex, multi-step tasks, use local files you grant access to, coordinate subtasks, create polished outputs, and run scheduled tasks.
Good Cowork tasks:
- Organize a folder of documents.
- Turn messy notes into a formatted report.
- Analyze CSV files and create charts.
- Create a spreadsheet with formulas.
- Build a slide deck from source notes.
- Research a topic and write a cited brief.
- Process transcripts into themes and action items.
- Run a recurring weekly summary from connected tools.
Bad Cowork tasks:
- "Rewrite this paragraph."
- "Give me five blog titles."
- "Explain what an API is."
- "Summarize this short note."
Regular chat is still the right tool for simple things.
Cowork matters when the task is big enough that you would normally open files, run commands, compare sources, organize outputs, and come back later. It also comes with real permissions. Anthropic says Cowork can read and write files in folders you connect, and that Claude can make real changes. It asks before permanently deleting files, but you should still review its plan before letting it run.
Use the cautious mode first.
The speed boost is not worth blindly granting access to important folders.
Scheduled Tasks In Cowork
Claude Cowork also supports scheduled tasks on paid plans. Anthropic says scheduled tasks can run recurring work or run on demand, using the same Cowork capabilities such as connectors, skills, and plugins.
Good scheduled tasks:
- Daily briefing from email, calendar, and Slack.
- Weekly competitor research.
- Monthly file cleanup in a specific folder.
- Recurring blog research summary.
- Weekly status report from project management tools.
Write scheduled prompts carefully. The task might run when you are not sitting there to clarify.
Good:
Every Monday morning, check the official Anthropic Claude release notes and Help Center pages for changes related to models, Projects, Skills, Artifacts, Connectors, Research, Cowork, and privacy. Create a short Markdown summary with links and note if nothing relevant changed.
Too vague:
Keep me updated on Claude.
"Claude" is too broad. Give the task a fence.
One major limitation: Anthropic says Cowork scheduled tasks only run while your computer is awake and Claude Desktop is open. That is important. This is not a cloud-only background worker you can forget about.
Computer Use Is Powerful, So Treat It Like Power
Claude can also use your computer in Cowork and Claude Code in supported environments. Anthropic describes computer use as a research preview where Claude can navigate your screen directly by clicking, typing, and opening apps when a connector or browser path is not available.
That is useful.
It is also the kind of feature where you should slow down.
Anthropic says Claude takes screenshots to understand your screen, which means it can see visible personal data, sensitive documents, or private information in apps you have allowed. It recommends not granting access to sensitive apps like banking, healthcare, or government services.
My rule: use connectors first, browser second, computer use last.
Connectors are usually faster and more reliable. Computer use is for the weird gap where the tool you need does not have a cleaner integration.
Start with small tasks. Watch what Claude is doing. Keep sensitive windows closed. Do not use it for high-stakes actions.
A Better Claude Prompt Template
Claude responds well when you give it structure. You do not need a magical prompt, but you do need a clear one.
Use this template:
Context:[What is this about? Who is it for? What background matters?]Goal:[What should Claude produce or decide?]Inputs:[Paste notes, attach files, list sources, or describe constraints.]Claude setup:[Use web search, Research, extended thinking, a Project, a Skill, an Artifact, or connectors if needed.]Output format:[Markdown article, checklist, table, spreadsheet, slide outline, report, etc.]Standards:[Tone, length, source requirements, accuracy expectations, privacy limits.]Process:[Ask clarifying questions only if needed. Otherwise make reasonable assumptions and proceed.]
Example:
Context:I am writing a medium-depth blog post for readers who understand basic AI chat but want to use Claude more effectively.Goal:Create a practical guide about setting up Claude with model choice, effort levels, extended thinking, Projects, Skills, Artifacts, Connectors, Research, and Cowork.Inputs:Use official Anthropic or Claude documentation where possible. Mention that features vary by plan and rollout.Claude setup:Use web search or Research for current product claims. Put the article draft in a Markdown artifact so we can revise it section by section.Output format:Blog-ready Markdown with headings, short paragraphs, examples, source links, screenshot suggestions, and image direction.Standards:Keep it practical, grounded, lightly funny, and not hype-driven. Clearly explain when not to use heavier features.
That gives Claude a job, a format, and a standard.
That is much better than:
Write an article about Claude.
Claude is good. It is not telepathic.
My Recommended Claude Setup
If you are past the beginner stage, here is the setup I would use.
First, add account-wide instructions. Keep them short and stable. Put your role, preferred style, and general answer expectations there.
Second, use memory if you want continuity, but review what Claude remembers. Use incognito chats for sensitive one-off conversations.
Third, make Projects for ongoing work. A blog series should have a project. A coding project should have a project. A research topic should have a project. Add files and project instructions.
Fourth, use Sonnet for everyday work and Opus for the hard stuff, based on what your account offers. Adjust effort only when the task actually benefits from it.
Fifth, turn on extended thinking for hard reasoning. Use web search for current facts. Use Research for serious source-backed synthesis.
Sixth, use Artifacts whenever the output is something you want to edit, reuse, or keep beside the conversation.
Seventh, create Skills for repeatable workflows. Keep them focused. A blog packaging skill, a proposal skill, a document-review skill, or a brand-style skill can save a lot of repetitive prompting.
Eighth, connect apps slowly. Keep approvals on. Do not give Claude access just because the connector exists.
Ninth, reserve Cowork for bigger tasks that benefit from local files, longer execution, scheduled runs, or actual deliverables.
That setup gives you most of Claude's value without turning the whole thing into a control panel you need a pilot's license to operate.
Screenshot Examples To Include
Because Claude changes often and features vary by plan, account type, and rollout, capture screenshots from your own account right before publishing. These are the screenshots I would include:
- Claude model menu next to the send button
- Show model selection, effort level, and extended thinking settings.
- Suggested caption: "Claude setup starts with model, effort, and thinking, not one magic prompt."
- Official reference: Change the model, effort, and thinking settings
- Instructions for Claude
- Show the account-wide instructions field.
- Suggested caption: "Use account instructions for stable preferences, not project details."
- Official reference: Understanding Claude's personalization features
- Claude memory and chat search settings
- Show memory, chat search, and incognito controls if available.
- Suggested caption: "Memory is useful, but it should be intentional."
- Official references: Claude chat search and memory, Use incognito chats
- A Claude Project with instructions and files
- Show project instructions, project files, and a focused project name.
- Suggested caption: "Projects give ongoing work its own context."
- Official reference: What are projects?
- An Artifact beside a chat
- Show a Markdown draft, diagram, or small HTML tool in the Artifact pane.
- Suggested caption: "Artifacts are the workbench for things you want to keep editing."
- Official reference: What are artifacts and how do I use them?
- Skills under Customize
- Show built-in skills and a custom skill if one is enabled.
- Suggested caption: "Use Skills for repeatable workflows, not one-off chats."
- Official references: What are skills?, Use skills in Claude
- Connectors and tool access
- Show connected apps, permission choices, and tool access mode.
- Suggested caption: "Connectors are powerful because they touch real data. Slow is smooth here."
- Official references: Use connectors to extend Claude's capabilities, Manage Claude's tool access
- Research enabled in the chat input
- Show the Research indicator or menu item.
- Suggested caption: "Research is for deeper source-backed answers, not every quick question."
- Official reference: Use research on Claude
- Claude Cowork task view
- Show the Cowork tab, task plan, or scheduled task screen.
- Suggested caption: "Cowork is where Claude moves from chat into longer desktop work."
- Official references: Get started with Claude Cowork, Schedule recurring tasks in Claude Cowork
Official Sources To Bookmark
- Claude Platform Docs: Models overview
- Change the model, effort, and thinking settings
- When should I use web search, extended thinking, and research?
- Enable and use web search
- Use research on Claude
- Understanding Claude's personalization features
- Use Claude's chat search and memory
- Use incognito chats
- What are projects?
- What are artifacts and how do I use them?
- Create and edit files with Claude
- What are skills?
- Use skills in Claude
- How to create custom skills
- Use connectors to extend Claude's capabilities
- Manage Claude's tool access
- Use Google Workspace connectors
- Get started with Claude Cowork
- Schedule recurring tasks in Claude Cowork
- Let Claude use your computer in Cowork
- Sensitive data in Claude chats
Final Thought
Claude gets much better when you stop treating it like a blank box.
Give it account instructions for the way you work. Use Projects when the work has continuity. Use Artifacts when the output needs a workbench. Use Skills when the workflow repeats. Use web search, Research, and extended thinking when the task actually needs them. Use Connectors and Cowork when the job touches real files, apps, or longer execution.
That is the real setup.
Not a single perfect prompt.
A thoughtful workspace.
That is where these tools start feeling less like toys and more like leverage.