Claude.ai in depth

Mastering Claude.ai: Projects and artifacts

Beyond plain chat, Claude.ai has two tools beginners often miss.

Projects: a space that holds a persistent context (custom instructions + knowledge documents) reused across all the project's conversations. Instead of re-pasting your brief every time, you put it once in the project. It is the consumer equivalent of a CLAUDE.md.

Artifacts: when Claude produces a self-contained deliverable (a document, a piece of code, an HTML page, a table), it opens it in a separate panel, editable and versioned, next to the conversation. You can iterate on it without drowning the chat.

Good habits:

• Put your recurring instructions (tone, language, format, constraints) in the Project instructions, not in each message.
• Load your reference documents into the Project's knowledge base.
• Use artifacts for anything that is a deliverable, not a discussion.

Move a paid seat into a Team plan without losing anything

A company usually starts with one person paying for an individual plan, then decides to roll Claude out to the whole team. At that point a natural fear appears: will creating a Team plan wipe out the account that already exists, with its conversations, its projects, its memory? It will not, but the actual mechanics have a right order and a wrong order, and getting it backwards costs both money and content.

Creating an organization (a Team plan) from an existing individual paid account does not destroy that personal account. The two coexist afterwards, and an account selector lets the same person switch between their personal space and the organization's space. The team owner can keep a personal subscription running alongside the new organization without conflict.

The part that trips people up is what happens to a teammate who already has their own paid plan. Migrating that person's individual account INTO the organization is a one-way move: it transfers their conversations, projects, files and memory into the organization's space, and it automatically cancels their individual subscription as part of the same operation. That automatic cancellation only fires in that exact order.

• Right order: the teammate joins the organization first, and the migration itself cancels their old individual plan and carries their content over.
• Wrong order: the teammate cancels their individual plan first, then joins the organization separately. This loses the automatic content transfer entirely: conversations, projects and memory built on the old account do not follow them in.

One more thing worth checking before a rollout: a premium or upgraded individual tier is not automatically equivalent to a standard team seat. If someone was on a high-usage individual plan, dropping them onto a standard team seat can be a real capability regression, not just an administrative formality. Open the actual billing screen and confirm exactly what each seat includes, and while you are there, confirm whether the organization owner's own seat is itself counted and billed as one of the paid seats. Assuming a stated headcount matches the billed headcount is an easy way to be off by one on the invoice.

What breaks first when you arrive from ChatGPT

Anyone who trains a group of new users coming from a different AI assistant sees the same handful of complaints show up, in roughly the same order, every single time. Knowing that order in advance turns a training session from a series of surprises into a session where you answer the question before it is even asked.

The first thing a new user tries, almost every time, is asking for a photo or an illustration. Claude has no native image generation. It is genuinely good at diagrams, charts and vector visuals built from structured instructions, and it can be connected to a third-party image generator through a connector (a plugin-like bridge to an outside service), but it does not produce photos or illustrations itself out of the box. This is reliably the first disappointment in a demo.

The second friction point is memory, and it is subtler because it looks like a bug rather than a design choice. Memory here is two separate systems with no bridge between them: an ongoing rolling synthesis of a person's regular, general conversations on one side, and memory that is scoped to a specific project workspace on the other. A user who expects one continuous memory across everything will conclude "it forgot" the moment they are simply looking in the wrong of the two places.

The third friction point is files, and it produces about half of all early "this doesn't work" reports combined with the memory confusion above. A spreadsheet with live formulas needs the code execution capability (Claude actually running code to compute the result) to be processed correctly; without it, formulas are not evaluated. A Word-format document, meanwhile, is read as plain text only: no bold, no headings, no layout awareness, just the words.

Finally, a straight vocabulary translation saves a lot of confused questions from someone used to a competing assistant's terms:

• Their custom assistant / custom GPT maps to a project workspace: shared inside an organization, not published to a public marketplace.
• Their canvas or rendering surface maps to an artifact: strong for building and displaying a finished output, weaker for fine incremental prose editing.
• Their plugin or action system maps to a combination of a skill (a packaged capability) plus a connector (the actual bridge to an outside service or account).

Public links, training opt-out and residency are three questions

A rushed team rollout tends to collapse three separate privacy questions into one comfortable assumption: "it's a paid business plan, so everything is private." Each of the three has a different answer, and treating them as one settled fact is exactly how a real exposure happens.

The first question is sharing. A conversation or an artifact shared through a public link is genuinely public, not private-but-hard-to-find: content shared this way has actually been found indexed by search engines. The safe posture is private by default, with active shares reviewed periodically, not shared-and-forgotten.

The second question is training opt-out. On consumer-tier plans, allowing conversations to be used to help improve the underlying models is an active setting that a user has to go find and turn off themselves; it is not off by default. Anyone assuming their consumer account is automatically excluded from training should go check the setting rather than assume it.

Business and professional tiers change this: the contractual commitment there is specifically that customer content is not used for model training. That is a real and meaningful guarantee, but it answers only the training question. It says nothing on its own about where the data physically lives.

That brings the third, separate question: data residency, meaning whether data is actually hosted within a specific region for a regulatory or contractual reason. This is unresolved by the training-opt-out guarantee, and the standard service does not commit to a specific region by default even on business tiers. Residency needs its own answer, confirmed in writing by the vendor for the specific compliance need at hand. It is never something to state from memory or to infer from the training commitment, because the two are genuinely different promises.

Put together: sharing controls who can see a specific piece of content, training opt-out controls whether content shapes future models, and residency controls where the data sits. A rollout that only checked one of the three has not actually checked privacy, it has checked a third of it.

Projects in depth

A Project in Claude.ai is a persistent workspace that holds three things together: a set of uploaded files (called project knowledge), a block of custom instructions, and the full history of every conversation you start inside it. Unlike a plain chat, which forgets everything the moment you close it, a Project keeps its context alive indefinitely, so every new chat you open inside the Project already "knows" your documents and rules.

Project knowledge is the library Claude reads before answering. You upload files (PDFs, text files, spreadsheets, code) or paste raw text directly. Claude indexes that content and can quote, summarise, or reason over it in any conversation inside the Project. Good candidates for project knowledge include:

• Company style guides or brand voice documents
• Product specs, FAQs, or pricing sheets
• Recurring reference data (glossaries, org charts, templates)
• Previous research you want Claude to build on

Custom instructions are a standing system prompt (a set of rules injected before every conversation) that you write once and never repeat again. Use them to set the tone ("always reply in British English"), define Claude's role ("you are a senior copywriter for a B2B SaaS company"), or enforce constraints ("never recommend competitors"). Instructions in a Project override the global custom instructions you may have set in your account settings, so you can have different personas for different Projects.

Because knowledge and instructions persist, Projects are the right tool whenever you have a recurring workflow: a weekly report, a client account, a research topic, or a product you support. Opening a new chat inside an existing Project costs you nothing in setup time, and Claude picks up exactly where the shared knowledge left off.

Custom instructions and styles

Both Claude.ai (the chat interface) and Claude Cowork (the desktop file agent) let you define custom instructions: a block of text Claude reads silently before every conversation. Think of it as a permanent briefing you write once, so you never have to repeat yourself.

In Claude.ai, open your profile and find Custom Instructions. You can tell Claude your role, preferred language register, output format, and things to avoid. In Claude Cowork, the equivalent lives in the project settings panel under Project Instructions. Both support plain prose and bullet lists.

Good custom instructions cover four areas:

• Role context: who you are and what you do ("I manage marketing for a B2B industrial equipment company").
• Tone and register: formal, casual, technical depth expected.
• Format preferences: bullet lists vs prose, short paragraphs, no em-dashes, no filler phrases.
• Hard constraints: things Claude must never do ("never invent statistics, never use exclamation marks").

Once saved, every new chat inherits these rules automatically. You can override them for a single session by simply telling Claude in that conversation, but the saved instructions stay unchanged for next time.

Artifacts you can share

When you ask Claude.ai to produce a document, a block of code, or a small interactive app, the result appears in a dedicated panel called an artifact. The artifact sits beside the conversation rather than inside it, so you can read, copy, or interact with it without losing track of the chat.

Three kinds of artifact exist. A document artifact holds formatted text such as a report, a contract draft, or a how-to guide. A code artifact shows syntax-highlighted source code you can copy in one click. A mini-app artifact (also called a rendered artifact) runs live HTML, CSS, and JavaScript directly in the browser, letting you see a working calculator, a quiz, or a chart without leaving the page.

Every artifact comes with a Share button. Clicking it produces a public link anyone can open, even without a Claude account. The recipient sees the artifact in a clean viewer and can interact with mini-apps as if they built it themselves. You can also remix a shared artifact: open the link, click Remix, and Claude starts a new chat pre-loaded with that artifact ready to edit.

• Document: markdown or rich text, ideal for reports and templates.
• Code: any language, copyable with one click, not executed in the browser.
• Mini-app: live HTML/JS, runs in the artifact panel, shareable as an interactive link.
• Remix: opens a shared artifact into a fresh Claude chat for further editing.

Claude on mobile and by voice

The Claude.ai mobile app (iOS and Android) brings the full conversational model to your phone. You can type, paste text, share images from your camera roll, and upload documents. The interface adapts to small screens, so long responses are paginated and easy to scroll.

Voice input lets you dictate your prompt using your phone's built-in microphone, then read Claude's reply on screen. This is not a true voice-to-voice assistant: Claude receives your transcribed text and returns written text. It shines for hands-free drafting, quick questions while commuting, or when typing is inconvenient.

Mobile Claude is best for tasks that fit in a single exchange or a short back-and-forth. It is not designed for long coding sessions or file-heavy workflows. Those belong in Claude Code (the CLI and IDE agent) or Claude Cowork (the desktop file agent). On mobile, stick to writing, brainstorming, summarizing, translating, and light research.

• Strengths on mobile: quick questions, voice dictation, image analysis (describe a photo, read a receipt), translation on the go, and drafting short messages.
• Limitations on mobile: no terminal, no file system access, no multi-step coding pipelines, no persistent memory beyond the current conversation (unless a Project is open).
• Projects on mobile: if you created a Project on Claude.ai, you can continue it on mobile and the project context (instructions, uploaded files) carries over.
• Model choice: the mobile app lets you select the model. Opus (claude-opus-4-8) gives the deepest reasoning; Sonnet (claude-sonnet-4-6) balances speed and quality for everyday use.

The growing Claude app family

Between late June and early July 2026, Anthropic shipped a wave of new surfaces for Claude, meaning new places and ways you can reach the same underlying model. The lesson here is not to memorize a product catalog, but to understand a simple idea: the model is the same brain everywhere. What changes from surface to surface is the job it is wired to do and the tools it can reach. Picking the right surface for the right job is now part of using Claude well.

On June 23, 2026, Anthropic launched Claude Tag in beta for Team and Enterprise plans. It lets Claude join Slack as a shared team member: instead of a bot you message privately, Claude becomes a presence you can tag (mention with @) directly into a channel, the same way you would tag a colleague. Once tagged, Claude can be delegated a multi-stage task, meaning a piece of work with several sequential steps, and it works through that task visibly inside the channel where the rest of the team can follow along.

On June 30, 2026, two things shipped on the same day. First, Claude Science, a dedicated AI workbench app built specifically for researchers, entered beta on macOS and Linux for Pro, Max, Team, and Enterprise plans. A "workbench app" here means a standalone application tailored to a workflow (in this case scientific research) rather than a general chat window. Second, Claude Desktop for Linux shipped as an official beta, supporting Ubuntu 22.04 and later plus Debian 12 and later, on both x86_64 (standard Intel/AMD 64-bit processors) and arm64 (the 64-bit ARM architecture used in many newer laptops and servers) processor types.

Claude in Chrome, the browser extension for consumers, is in beta on all paid plans and integrates with Cowork (Anthropic's collaborative workspace) as a research layer, meaning Claude can browse and gather information from the web to feed into Cowork sessions. Separately, the developer-facing side of this integration matured further: the Claude Code integration inside Chrome became generally available (GA, meaning fully released and stable rather than beta) in version v2.1.198 on July 1, 2026.

On the coding side, Claude Code Artifacts entered beta on June 18, 2026 for Team and Enterprise plans. This lets a coding session running in Claude Code publish a live, self-contained interactive HTML page to a private URL hosted on claude.ai, so a working prototype or tool built during a session can be shared with a link instead of a code export.

Finally, on July 7, 2026, the Microsoft 365 connector (the integration that lets Claude interact with Outlook, Calendar, OneDrive, and SharePoint) gained write tools. Before this date it was read-only, meaning Claude could look at your email, calendar, and files but not change anything. Now it can draft and send email, manage calendar events, and create or update files in OneDrive and SharePoint, which turns it from an information source into an active assistant that can take action on your behalf inside Microsoft's ecosystem.

Taken together, these releases show a pattern: Anthropic is not building separate AI products, it is building more doors into the same Claude. A Slack mention, a Chrome tab, a Linux desktop app, a research workbench, a live HTML artifact, and a Microsoft 365 write action are six different doors into one brain. The skill for an intermediate user is learning which door fits which task.

Update, July 11, 2026: on July 9, 2026 Claude.ai gained a well-being pair: Reflect (Settings, then Reflect), a monthly recap of your topics, activity patterns and work insights built from your memory, and Time and focus, optional break reminders and quiet hours. Both shipped in beta on the Free, Pro and Max plans, web and desktop.