Claude Just Had a Crazy 2026: The 18 Features You Need to Stay Ahead of 95% of Users
Anthropic shipped non-stop this year. We tested, analyzed and documented all their new features, with a 60-second try-it for each.
Anthropic started with one surface on Claude. Now there are three popular ones.
The chat window at claude.ai came first. It is the conversation surface most people still meet Claude through, the place you go to ask, draft, and think. Claude Code came next, built for developers who wanted Claude inside the terminal and the IDE so it could actually write, run, and ship code instead of just describing it. Then came Cowork, a desktop assistant that organizes your files, runs scheduled tasks, and works alongside you on everything else on your computer. Three surfaces, three jobs.
That alone would have been a year of shipping. Anthropic kept going. This year they stacked new features inside each surface. Chat got Claude Design, Inline Interactive Charts, Output Styles, Fork, /btw, /rewind, and /recap. Code got Routines, /ultrareview, Auto Mode, Ultrathink, Subagents, Hooks, and Plugins. Cowork got Live Artifacts, Routines, and the plugin marketplace. Three surfaces, each with a quietly growing toolkit, all rolled out in the first four months of the year.
What Anthropic shipped between January and April is not a feature list. It is a strategy. They built the workflow, then wrapped a consumer product around it. Claude Design for non-designers. Routines for people who do not know what a CRON job is. Live Artifacts in Cowork for anyone who gave up on building dashboards. Inline Interactive Charts for every Claude user on every plan.
I have been using every one of these. Most Claude users are still on 3 or 4 of them. This article walks through 18 features that changed how I work, grouped into 6 jobs you can hand to Claude, each with a 60-second try-it and the mistake most people make on setup.
If you used Claude in January and have not touched it since April, this is your catch-up.
If Claude is already your daily driver, this is your audit.
In This Post, we will cover:
Section 1: Run Claude While You Sleep (Routines, Managed Agents, /ultrareview, Auto Mode)
Section 2: Control Your Conversation (Fork, /btw, /rewind, /recap, Transcript View)
Section 3: Make Claude Think Harder (Ultrathink, Subagents)
Section 4: Customize How Claude Works (Hooks, Output Styles, Plugins)
Section 5: New Things Claude Can Now Make (Claude Design, Live Artifacts, Inline Interactive Charts)
Section 6: The Base Layer (Claude Opus 4.7)
The point is not to use all 18
A Note on Interfaces
Every feature carries an interface category(s) so you know where it lives.
Chat: claude.ai browser app.
Code: Claude Code in web or desktop app.
Code (CLI only): Claude Code terminal/CLI. Slash commands that do not work in the web or desktop app.
Cowork: Claude Desktop’s Cowork environment.
API: Claude API with SDK or direct HTTP.
If a feature works in more than one, all are listed.
Section 1: Run Claude While You Sleep
Four features ship the “cloud automation” vision. Routines is the headline. The other three matter more than people realise.
New to Claude Code? Start here, the primer for non-developers 👇
1. Routines: Claude that runs on its own schedule (Done)
🏅 Level: Beginner
🖥️ Interface: Code Desktop, Cowork
What it does in action. Set up a task once, have Claude run it for you automatically. Give it a prompt, connect the tools it needs, pick when to run it, and the rest is hands-off. If you choose a remote routine, it runs on Anthropic’s cloud and your laptop does not need to be open. If you choose a local routine, it runs on your machine, so the laptop has to be on and awake when the routine fires. The output is waiting when you check back.
When to use it.
Any task that needs to run on a schedule without you starting it each time.
Work that needs to happen while you sleep or while your laptop is closed.
Jobs another tool should trigger automatically, like a webhook from your CRM or a Slack alert.
When NOT to use it.
You picked the “local” routine option by mistake. Local routines only fire while your computer is awake. If the whole point is “while I sleep,” you need cloud.
The task needs fresh human judgement every run. You will get a consistent wrong answer.
Try it in 60 seconds. Open Claude Code (web or desktop) → Claude Code → Click the Routines panel- > Click New routine → Choose Local/ Remote → Name the routine, paste your prompt, choose a trigger → select any triggers it needs (Slack, Notion, Linear, etc.) → Create.
Local routines live on your machine and only fire while the computer is awake, which suits tasks that need access to files or tools that sit on your laptop. Cloud routines run on Anthropic's servers regardless of whether your device is on, which suits anything you want to happen while you sleep or while you are away from your desk. Many workflows use a mix of both, depending on where the data lives.
What the community figured out.
From Daria Cupareanu: Connect all your tools upfront before you build your first routine. You will not need to pick them one by one each time. Claude figures out which connectors to use based on what your prompt mentions.
From Ilia Karelin: Routines, /loop, and Desktop Tasks are three different products. Routines runs in the cloud, /loop runs locally in one session, Desktop Tasks lives on your machine. Pick based on where the work has to happen.
We publish builds like this weekly — workflows, prompts, and the mistakes we make so you do not have to. Subscribe to get the next one.
2. Claude Managed Agents: Always-on agents without the infra
🏅 Level: Advanced
🖥️ Interface: API
What it does in action. You describe the agent you want in plain English, and Claude builds it for you. Tell it what the agent should do, what tools it needs, and what files it should work on. Anthropic spins up a private cloud workspace where the agent can read files, run commands, and browse the web. The agent runs on its own, picks up where it left off across sessions, and you can step in any time to redirect it.
When to use it.
Long-running work that takes minutes or hours and chains many tool calls.
Async jobs you want to fire off and check back on, with state preserved across runs.
Anything where you would otherwise spend a sprint building your own sandbox, retry logic, and event stream.
When NOT to use it.
Single-turn prompts. The Messages API is simpler and cheaper.
Workloads where you want to swap Claude for another provider mid-stack. Managed Agents is Anthropic-only.
Try it in 60 seconds. You do not need a dev team for this. Even a one person operation stitching their own tools into one workflow can spin up an agent that runs in the cloud. Open the Claude Console → Click managed agents on the side panel → Agents → New Agent → Either build a guided agent with Claude or leverage some of their pre-built agents. Full quickstart: https://platform.claude.com/docs/en/managed-agents/overview
Pro tip: pair it with the Advisor tool. Inside a Managed Agent, a cheap executor (Sonnet or Haiku) can ping a smarter advisor (Opus) mid-task for a quick plan, then keep going. You get Opus-grade thinking on the architecture decisions while paying executor rates for the bulk of the work. Docs: https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool
What the community figured out.
From Daria Cupareanu: The biggest unlock is not the infra, it is a prompting shift. Daria’s writeup pulls a Rakuten lesson worth stealing even if you never ship a Managed Agent: stop scripting steps, hand the agent the outcome you want and let it figure out the route.
3. /ultrareview: Cloud bug hunt before you merge (Done)
🏅 Level: Intermediate
🖥️ Interface: Code Desktop
What it does in action. Before you ship code, a team of Claudes reviews it together in the background. They look for real bugs, confirm each one is actually a bug and not just a style opinion, and send you back a list. Runs in the cloud, so you keep working while they review.
When to use it.
Before you merge something that touches payment flows, auth, or data writes.
After you wrote something in a language you do not use every day.
When the change is large enough that a human reviewer would miss something.
When NOT to use it.
Trivial PRs. You burn $5 to $20 per run after your 3 free runs. That cost adds up fast if you fire it on every commit.
Code that is still mid-draft. Run it on a clean diff, not a thought in progress.
Your working tree is clean. /ultrareview needs a branch with changes ahead of main — if there is nothing to diff, it will not run. You need uncommitted changes on a branch or a PR number to point it at.
Try it in 60 seconds. In Claude Code (desktop or CLI), push your branch and open a draft PR. Run /ultrareview <PR-number>. No argument runs it on your current branch diff including uncommitted changes. Review takes 5 to 10 minutes in the background.
What I found. I ran it against a set of recent commits: config changes, resource file updates, and a content file. It caught a metadata inconsistency between two files that contradicted each other, flagged a stale type definition that no longer matched the codebase, and suggested a small improvement to a deployment config. Nothing earth-shattering, but every finding was real. No style nitpicks, no false positives. For a basic review of a handful of commits, that signal-to-noise ratio is genuinely better than what most linters give you.
What the community figured out.
Pro and Max subscribers get 3 free ultrareview runs. They are a one-time allotment, do not refresh, and expire on May 5, 2026. After that each run is billed as extra usage at roughly $5 to $20 per review. Use them on real code, not toys. https://code.claude.com/docs/en/ultrareview
Check out Karo’s article for great tips on when to use /ultrareview.
The value is verified findings, not volume. Testers consistently flag that every reported bug was independently reproduced, which means less noise than local /review.
4. Auto Mode: Claude works unsupervised, with a safety net (Done)
🏅 Level: Intermediate
🖥️ Interface: Code, Code (CLI)
What it does in action. Claude does multi-step work without stopping every 10 seconds to ask permission. A safety checker watches each action in the background and blocks anything risky. You get longer stretches of uninterrupted work without babysitting the screen.
When to use it.
Long refactors across many files.
Build-and-test loops where you want Claude to iterate.
Deep research tasks where stopping to approve every step slows you down.
When NOT to use it.
Any task touching production credentials, payments, or irreversible actions.
First time with a new codebase. Watch Claude first, trust it second.
Try it in 60 seconds.
Claude Code desktop: open Settings and flip the “Allow auto permissions mode” toggle on. Then pick Auto from the permission mode dropdown in your session.
Claude Code CLI: tap Shift+Tab to cycle through permission modes until it lands on Auto. Start your task.
/permissionsis a slash command that opens Claude Code’s permission rules view: the list of tools, Bash commands, and file paths the agent is allowed, asked about, or blocked from touching in your session. It only works in the Claude Code CLI terminal, not the desktop app.*
Section 2: Control Your Conversation
Five features fix the biggest daily friction. Your context window, your thread, your undo button, your memory.
5. Fork: Branch your chat without losing the original (Done)
🏅 Level: Beginner
🖥️ Interface: Code Desktop, Code CLI
What it does in action. Clone any Claude conversation at any point, then continue in two directions at once. Your original chat stays untouched. The copy starts with the full history. Use it when you want to try a different angle without losing what you have.
When to use it.
You want to test a different direction without abandoning what works.
You are about to run a risky prompt and want a safe copy.
You want two threads exploring the same context at once.
When NOT to use it.
You think it will save tokens. It will not. The fork inherits the entire conversation, so a 50% full thread forks into two 50% full threads.
You want a clean slate. Use Projects or start a new chat.
Try it in 60 seconds. In Claude Code desktop, hover any past message and click Fork from here. A new thread opens with full history intact. In Claude Code CLI, run /branch inside a session.
6. /btw: Ask a question without polluting the main thread (Done)
🏅 Level: Beginner
🖥️ Interface: Code Desktop
What it does in action. Ask Claude a quick side question without cluttering your main conversation. The question and answer stay separate, so your main thread keeps its focus. You can even ask while Claude is busy with something else.
When to use it.
You need a quick fact or file lookup without derailing the main task.
You want to test a phrasing without committing to it.
You are deep in a long session and do not want every small question bloating the conversation.
When NOT to use it.
The question is actually the next step. Then ask it normally so Claude remembers.
Try it in 60 seconds. In Claude Code, type /btw followed by your question. On desktop, a side chat opens as an overlay on top of the main chat. In the CLI, the answer appears inline in the response while your main task keeps running. Either way, the question does not persist in the main thread.
7. /rewind and Checkpoints: Esc twice, it never happened (Done)
🏅 Level: Beginner
🖥️ Interface: Code Desktop, Code CLI
What it does in action. An undo button for Claude. Before every change Claude makes, it saves a snapshot. If something goes wrong, tap Esc twice and Claude puts everything back the way it was. Works for both your files and the conversation itself.
When to use it.
Claude made a bad edit across 4 local files on your machine (code, docs, scripts, configs) and you want it undone in one shot.
A prompt went sideways and you want to retry from the message before.
You want to fork the conversation. Rewind to a message, send a different follow-up, and explore Path B without polluting Path A.
When NOT to use it.
You expect it to undo edits Claude made to Notion, Slack, GitHub, or any connector. Rewind only restores files on your machine. External edits are already live on those servers.
Try it in 60 seconds. Tap Esc twice after any change you want to undo, type /rewind, or hover any prior message and click Rewind to here. Pick whether to restore code, conversation, or both.
Rewind only rolls back what lives inside Claude. It does not check or revert state in external tools. If you ask Claude to edit a Notion page, then four messages later rewind to before that request, the chat goes back but the Notion page stays exactly as Claude left it. Same rule for Slack messages, GitHub commits, and anything else sent through a connector.
Enjoying the breakdown? Free subscribers get round-ups like this every month. Paid gets the step-by-step builds, prompts, and templates we use ourselves.
8. /recap: Catch up on a session you left hanging (Done)
🏅 Level: Beginner
🖥️ Interface: Code CLI
What it does in action. Come back to a session a day later, type /recap, and Claude summarises where you left off. No need to scroll through the whole conversation trying to remember what you were doing.
When to use it.
You left a long CLI session mid-task and came back fresh.
You switched machines and Claude’s session synced but your memory did not.
When NOT to use it.
No need to waste tokens on a smaller conversation you can scroll through quickly.
You are in Claude Code web/desktop. /recap is CLI only.
Try it in 60 seconds. In Claude Code CLI, type /recap. Read. Continue.
9. Transcript View: Choose how much of Claude’s process you see (Done)
🏅 Level: Beginner
🖥️ Interface: Code Desktop
What it does in action. Every response Claude gives can be viewed at different levels of detail. Normal shows the conversation cleanly with tool calls folded into short summaries. Thinking shows the reasoning Claude worked through before it responded. Verbose opens everything: every file read, every tool called, every step taken. Summary strips it back to final answers and changes only. Switch between them any time without losing the conversation.
When to use it.
Normal: the standard chat layout for everyday use. Your prompts and Claude’s final answers, with tool calls collapsed into one-line summaries you can expand if needed.
Verbose: the full execution trace. Every tool call, file read, command, and reasoning step rendered in order, so you can see the entire path Claude took. Use it when something went wrong and you want to see exactly what happened.
Summary: an outcome-only view with a dynamic title at the top that updates as the chat evolves. Only the current prompt and Claude’s output are shown, everything else is captured in the tile. Use it when you are juggling multiple sessions and only care about results.
Thinking: the reasoning view. Claude’s chain of thought renders inline above each answer, separate from the final response and from the tool plumbing. Use it when Claude made a decision you want to follow or question.
When NOT to use it.
Verbose as your default. It is a lot to read on routine tasks.
Try it in 60 seconds. In Claude Code Desktop, click the Transcript view dropdown (ref: first image). Switch between Normal, Thinking, Verbose, and Summary. You can also press Ctrl+O to cycle through modes.
Section 3: Make Claude Think Harder
Two features control depth of reasoning. One is a keyword. One is a team of smaller Claudes.
10. Ultrathink: One word, maximum reasoning (Done)
🏅 Level: Beginner
🖥️ Interface: Code CLI
What it does in action. Type the word “ultrathink” in your prompt and Claude takes much longer to think before answering. The answers are deeper and more considered. Takes about 3x longer than a normal response and costs more, so save it for decisions that actually matter.
When to use it.
Architectural decisions that will shape a codebase for months.
Debugging a bug that has survived 3 other attempts.
A business decision you want Claude to think through end-to-end.
When NOT to use it.
Routine asks. Most prompts do not need 31,999 tokens of thinking.
In claude.ai web chat or the API. The keyword only works in Claude Code CLI.
Try it in 60 seconds. In Claude Code CLI, write your prompt and include the word “ultrathink” anywhere in it. Run it.
What the community figured out.
From Nir Diamant: Save ultrathink for the problems that have already cost you time. Architectural decisions, or a bug that has stumped you for hours.
11. Subagents and /agents: Delegate to purpose-built mini-Claudes (Done)
🏅 Level: Advanced
🖥️ Interface: Code CLI
What it does in action. A subagent is a specialised mini-Claude with its own context window, its own system prompt, its own tool permissions, and its own model (Haiku, Sonnet, Opus). When Claude sees a task that matches a subagent’s description, it delegates automatically and the subagent returns only the summary. The benefit is twofold: your main thread stays focused on the work instead of being flooded with logs, search results, or file dumps, and you can route the heavy lifting to a cheaper model like Haiku while Opus stays on the main conversation.
Where they live. Project subagents travel with your repo so the whole team gets the same set. Personal subagents stay on your machine and work across every project.
How they differ from skills. Skills add knowledge to your main thread. Subagents do the work in a separate context window and hand back only the summary. Skills make the main thread smarter. Subagents keep it clean.
When to use it.
A side task would flood your main thread with search results, logs, or files you will not reference again. The subagent does the work in its own window and hands back the summary.
You keep spawning the same kind of worker with the same instructions: code reviewer, test generator, doc writer, researcher.
You want to constrain tool access for a specific job, like a reviewer that can only read and not write.
You want to route routine work to a cheaper, faster model like Haiku and keep Opus on the main thread.
When NOT to use it.
One-off tasks you do occasionally. A slash command is lighter.
You cannot write a clear description of what the subagent should do. Claude uses that description to decide when to delegate, so a vague description means flaky delegation.
Try it in 60 seconds. In Claude Code CLI, run /agents. The panel opens with two tabs: Running and Library. Full guide: https://code.claude.com/docs/en/sub-agents Open Library and hit Create new agent. Choose where it lives: Project if the team should share it, or Personal if it is just for you. Pick Generate with Claude (recommended) and describe what the subagent does in plain English. Claude writes the markdown file for you. Use Manual configuration only if you want to edit the YAML by hand. The description you write is what Claude reads to decide when to delegate, so be specific. Source: https://code.claude.com/docs/en/sub-agents
See how chaining 4 Claude Code skills builds one automated pipeline 👇
Section 4: Customize How Claude Works
Four features turn Claude from a tool into a platform. Three are for individuals. One is for teams.
Writing your own skills or subagents? The 5 fixes that actually work, and the failure modes to avoid 👇
12. Hooks: Run your own scripts when Claude does something (Done)
🏅 Level: Advanced
🖥️ Interface: Code Desktop, Code CLI
What it does in action. Think of hooks as automated tripwires. Strict rules you set up once that fire automatically at specific moments in Claude’s workflow. Instead of hoping Claude remembers to do something, hooks guarantee it happens. Claude uses judgment to make decisions. Sometimes you do not want judgment, you want absolute control. Hooks give you that.
Four things hooks can do for you.
The tap on the shoulder. Send a desktop notification when Claude is waiting for permission, so you can step away while it works.
The bodyguard. Block edits to sensitive files like password files or
.gitfolders before they happen.The janitor. Run a code formatter like Prettier the moment Claude finishes writing a file, so every output matches your team’s style.
The auto-approve. Auto-confirm safe actions Claude keeps asking permission for, so you stop clicking approve.
How they work. As Claude works it passes through events: SessionStart when you open Claude Code, PreToolUse right before Claude uses a tool, PostToolUse right after. You set up hooks once in a small config file. You tell Claude: when you hit this event, pause, run my script, then check the result. Your script either lets Claude proceed or blocks it.
Hooks are powerful, but they run real commands on your machine with your permissions, so a careless one can do real damage. That is exactly why Chris from ToxSec and I are writing a deep dive on hooks: when to use them, how to install your first one, the risks to watch for, and the best practices for setting them up safely. Going live 5th May, subscribe to get it delivered to your inbox.
Try it in 60 seconds. Hooks live in your Claude settings file. You pick the moment you want Claude to pause (like right after it edits a file), point it at the script you want to run, save, and you are done. For the step by step on where the file lives, what to write, and a few ready to use examples, check the official guide at https://code.claude.com/docs/en/hooks-guide or check your inbox on 5th May for the deep dive Chris and I are publishing.
13. Output Styles: Tell Claude how you want it to talk to you (Done)
🏅 Level: Beginner
🖥️ Interface: Chat
What it does in action. Five preset styles change how Claude responds without changing what it knows. Concise keeps answers short. Explanatory adds the reasoning behind each decision. Learning curates a learning plan tailored to your profile. Formal tones it up for professional contexts. You can also create a custom style: upload a writing sample or describe your preferred style in words, and Claude builds a tailored version from it. Content from your sample is not stored after the style is created.
When to use it.
You keep getting responses that are longer or more formal than you need.
You are learning something new and want Claude to explain the why, not just the answer.
You want Claude to match your team’s writing voice across every response.
When NOT to use it.
You want different answers, not a different tone. Styles change how Claude communicates, not what it knows.
Try it in 60 seconds. In Claude.ai chat, click the + button and select Use style. Pick a preset (Normal, Concise, Explanatory, Formal, Learning) or click Create & edit styles to upload a writing sample or describe your preferred tone. Claude builds a custom style from it.
14. Plugins: One install, a full workflow ready to go (Done)
🏅 Level: Beginner
🖥️ Interface: Code Desktop, Code CLI, Cowork
What it does in action. A plugin is a bundle. One install gives you a set of skills, slash commands, sub-agents, hooks, and connectors that work together for a specific job. A skill is one tool: a folder of instructions Claude reads when the task matches. A plugin is the toolbox: it can hold several skills plus the slash commands and automations that wire them together. So a code-review plugin might ship the review instructions (skill), a /review command, a sub-agent to run reviews in parallel, and a hook that fires the review on every commit. You install it once, and all of that is set up. The Plugin Marketplace is where you find them: Anthropic ships official ones, and community marketplaces add dozens more, so you are not hunting GitHub.
When to use it.
The job needs a full setup, not just one instruction. Skills, connectors, defined sub-agents, all working together.
You want a workflow other people have tested instead of building it from scratch.
You want to standardise a team’s Claude Code setup. Everyone installs the same plugin and gets the same behaviour.
When NOT to use it.
A skill alone would do the job. If all you need is instructions, write a skill. Plugins are heavier.
You will install 20 plugins in one afternoon because they look cool. Every plugin is context you pay for.
Try it in 60 seconds. From our own build: Anthropic shipped 11 plugins. I needed a 12th — a Product Research plugin using the same evaluation framework behind every Cash & Cache tool review. The lesson: plugins are only as good as the inputs you prepare. Four structured files in, three publish-ready outputs back. The mid-task corrections barely mattered. If you are building your own, spend the time on input structure, not prompt engineering.
You do not always have to borrow someone else’s. Build your own, here is how 👇
What the community figured out.
From Jenny Ouyang: Testing 10 plugins and keeping 4 is roughly the right ratio. Most are fine in isolation and mediocre in combination. Pick based on workflow fit, not demo appeal.
Section 5: New Things Claude Can Now Make
Three launches that moved Claude from “answers questions” to “makes things you can use.”
15. Claude Design: Decks, prototypes, one-pagers from prompts (Done)
🏅 Level: Beginner
🖥️ Interface: claude.ai/design
What it does in action. Describe a slide, a one-pager, or a prototype in plain English. Claude builds it for you. You tweak it with more plain English (”make the header bigger,” “change the colour scheme to navy”), and when it looks right, export to PowerPoint, PDF, Canva, or straight into code.
When to use it.
Investor one-pagers, internal decks, pitch materials where speed beats polish. Handoffs from design intent to code. Claude Design bundles everything for Claude Code with minimal friction. Non-designers who need something that looks professional fast.
When NOT to use it.
Brand-critical design work. Figma is still better for sustained design systems. Pixel-perfect precision. Claude Design is good at drafts, Figma is good at finals. Hero images and cover art. It produces structured HTML layouts, not photorealistic imagery. This is the mistake a non-designer is most likely to make in the first session: reaching for the new tool for everything.
Try it in 60 seconds. Go to Claude.ai. New chat. Select Claude Design from the tool menu. Describe what you want to make. Iterate on the canvas. Export.
What I actually found. I tested Claude Design across 5 real Cash & Cache tasks. The article one-pager and the sponsor pitch deck came back clean, branded, and PPTX-ready in one prompt. The biggest surprise: piping our Tool Evaluator output into Design produced a spider graph and a four-tab comparison that says in one glance what takes three paragraphs of prose. Evaluator becomes the engine, Design becomes the showroom.
The full breakdown of the 5 tasks, the design system setup, and the exports is here 👇
What the community figured out.
From Ruben Hassid: The strongest workflow chains commands like /discover, /strategize, and /handoff to move from research to UX strategy to developer handoff inside one session:
16. Live Artifacts: Cowork dashboards that refresh themselves (Done)
🏅 Level: Beginner
🖥️ Interface: Cowork
What it does in action. Standard Claude artifacts are snapshots: Claude builds something once and it freezes there. Live Artifacts are different. Build a dashboard or tracker once in Cowork, connect it to your data sources (Notion, Slack, Stripe, Gmail, HubSpot), and every time you open it, it pulls fresh data and rebuilds itself. No re-prompting. No manual refresh. The output becomes a living tool you keep, not a one-time deliverable you paste somewhere and forget. Every version is saved automatically, so you can roll back anytime.
When to use it.
You need one view to see activity across multiple tools. Claude pulls from Notion, Linear, Slack, and others and builds that all-in-one view for you.
Trackers that need to stay current: campaign performance, customer health, project status, competitor moves.
Any report that used to require a data person and a sprint. If the data lives in a connector, Claude can build and maintain it.
When NOT to use it.
One-off static output. Use a regular artifact.
Data that lives outside any Cowork connector. Live Artifacts refreshes from connected sources, not arbitrary URLs.
Try it in 60 seconds. Open Claude Desktop in Cowork → Live Artifacts → New Artifact → Chat with Claude → Claude will automatically run the following prompt:
I want to make a live artifact. Explain what live artifacts are in Cowork, then look at my connectors (MCP servers), and ask me a few questions to figure out what kind of live artifact would be most useful for me.
It saves to the Live artifacts tab. Reopen from that tab anytime and it refreshes with current data. Click the refresh icon on the top right anytime to get an updated dashboard
Want to see what a full Cowork session looks like end-to-end? Here is how one session produced an entire newsletter 👇17. Inline Interactive Charts: Visuals Claude draws inside the chat (Done)
🏅 Level: Beginner
🖥️ Interface: Chat, Cowork
What it does in action. Paste any text that has numbers in it and ask Claude to chart it. It reads the data, picks the right chart type, and builds an interactive visual inside the chat. Click on it, ask Claude to change the format or add more data, and it updates. No spreadsheet or formatting required.
I broke this down in a quick note, the exact prompt and a worked example 👇
When to use it.
Turning unstructured text (meeting notes, research summaries, strategy docs) into a chart without cleaning the data first.
Comparing competitors by pasting their pricing page URLs and letting Claude extract each tier into a single chart.
You want to understand a topic through numbers. Ask Claude to explain LLMs, market trends, or any concept in data, and it builds the data story from scratch.
When NOT to use it.
Direct deck imports. You can export to a standalone HTML file, but plugging that into PowerPoint is a separate step.
Heavy, live-data dashboards. Those belong in Live Artifacts.
Try it in 60 seconds. Go to claude.ai (free tier with Sonnet 4.6 is enough). Paste any text that contains numbers, trends, or comparisons. Then add this prompt:
Create an interactive chart from the data in the text above.
Identify the key metrics and relationships.
Don't code. Interactive chart only.Claude extracts the data and builds the chart. To share it, follow up with Code this as a downloadable HTML file. and you get a standalone file you can email or upload anywhere.
Section 6: The Base Layer
One model upgrade. Everything else rides on it.
18. Claude Opus 4.7: Best Opus to date, same price (Done)
🏅 Level: Beginner
🖥️ Interface: All surfaces (Chat, Code Desktop, Code CLI, Cowork, API)
What it does in action. The newest and smartest Claude model. Writes noticeably better code than the previous version. Reads images, charts, and graphs much more accurately. Handles long, complex tasks without losing its place. Already the default for Pro, Max, and Team accounts in Claude Code.
When to use it.
Any hard coding task. This is the best model Anthropic has for long, complex engineering work.
Anything that involves reading images. Chart interpretation, graph reading, parsing overlapping data series, and pulling information from screenshots are all noticeably better.
Creative work where craft matters. The model is noticeably better on professional writing, design, and brand output than 4.6.
When NOT to use it.
Pulling specific facts out of long documents. If you regularly hand Claude a contract, a long report, or a fat PDF and ask it to find a specific clause or number, 4.7 is noticeably worse at this than 4.6. Anthropic confirmed the drop in their own release notes. Test it on your documents before you switch. https://blog.wentuo.ai/en/claude-opus-4-7-long-context-regression-en.html
Flowing prose. Writers report 4.7 reaches for bullets and headers where 4.6 held a narrative flow. One reviewer summed it up as “it used to feel like talking to a thoughtful colleague. Now it feels like getting a memo.” Check your output if long-form writing is your core use case. This breakdown from Hamza Farooq puts Opus 4.6 against 4.7 to show you what’s changed 👇
Simple everyday tasks. Sonnet is more than capable, and cheaper. Use Opus for planning, strategy, and decisions that change how a workflow runs. Use Sonnet for BAU work: draft editing, summarizing meeting notes, formatting a doc, answering a quick factual question, drafting a routine email.
Try it in 60 seconds. Claude Code and Cowork already default to Opus 4.7 on the Max plan, and on Claude.ai it sits at the top of the model picker. Have you tried Opus 4.7? What has been your experience? Drop your experience in the comments, I read every one.
What the community figured out.
From Zvi Mowshowitz: The 48-hour reaction was unusually split for an Anthropic point release. Coding users are enthusiastic but prose writers have concerns. Run your own tests before committing.
Watch token economics. 4.7 ships with a new tokenizer, so the same text counts as roughly 1.0 to 1.35x more tokens than on 4.6 (code and technical docs hit hardest). Anthropic also recommends a higher default
xhigheffort level for coding and agentic work.
The Point Is Not to Use All 18
Anthropic's shipping pace this year has been wild. Keeping up with it takes real time, and they may have shipped something else by the time you read this. Treat this as a snapshot, not a finish line.
You do not need every feature in this list. You will not use every feature in this list. Nobody should. Tool overload is real and it costs more than the tools themselves.
The advantage is not using more, it is choosing better. Read back through the 18, mark the ones that match a job you actually do every week, and test them on your workflows.
Which of these are you keen to get started on, and which would you like us to publish a deep dive on next? Drop it in the comments or message me directly.
💬Did I miss anything?
What is the feature Anthropic shipped recently that changed your workflow the most? Drop it in the comments or message me directly. If enough of you weigh in, I will turn this into a living document that captures what the community figures out, not just what one of us does.
💡 Enjoying this article or other ones about frameworks to work with AI? Share this piece with someone who can benefit.
Hard to keep up with Anthropic these days 🫠 putting all of this together and making sense of the flood (literally a flood!) of new features is suuuper helpful. Awesome roundup, guys, and thanks for the shoutout!
Amazing article with amazing work form others too!