Most people use Claude like a slightly smarter search engine. Type a question, get an answer, close the tab. That’s leaving an enormous amount on the table. Claude has grown into a layered platform with agentic tooling, persistent workspaces, and an interactive runtime — and the gap between a casual user and someone who’s actually configured it properly is massive. Here’s the tour I wish I’d had.
</figure>
First: Pick the Right Model
This sounds obvious until you realize most people just use whatever the default is and call it a day.
Claude comes in three tiers — Haiku, Sonnet, and Opus — and matching the model to the task is itself a productivity hack. Haiku is fast and cheap, great for high-volume classification, summarisation, or anything where latency matters. Sonnet is the daily driver: balanced, capable, handles most writing and coding tasks without drama. Opus is for the hard stuff — complex reasoning, long document synthesis, anything where you’d rather wait a few extra seconds and get it right.
Running everything through Opus because it sounds most impressive is like using a sledgehammer to hang a picture frame. Pick Haiku when speed matters, Sonnet when you just want to get work done, and Opus when the task genuinely warrants it.
Artifacts: The Hidden Interactive Runtime
This is the feature that changes everything and somehow most people walk right past it.
Artifacts is a side panel that renders live, interactive output — HTML pages, SVG graphics, React components, dashboards, mini-apps, even games — directly inside the chat. You’re not just reading generated code; you’re running it. Ask Claude to build a calculator, a data visualisation, or a resume builder, and it appears on the right, live and clickable.
The stack it supports out of the box is genuinely impressive: React with Tailwind CSS, Shadcn UI, Lucide icons, and Recharts. A single prompt can produce a fully styled, interactive component.
A prompt that actually works:
Build a React component using Tailwind CSS that shows a kanban board
with three columns (To Do, In Progress, Done) and drag-and-drop cards.
Each card should have a title, tag colour, and due date.
Make it functional with local state — no external libraries beyond what's available.
You get a working kanban board. In the chat. Without spinning up a dev environment.
A few things worth knowing:
Adding “using AI” to your Artifact prompt tells Claude to embed reasoning logic into the output — making it smarter, not just prettier.
Artifact version history works like a lightweight Git: you can inspect and revert to any previous iteration if Claude goes off in the wrong direction.
Custom Visuals (a newer addition) lets you request bespoke flowcharts, comparison tables, and diagrams — Claude generates them as proper SVG rather than ASCII art.
Tip: If you want to share an Artifact externally, Claude can publish it to a shareable URL. Great for quick prototypes or internal demos where spinning up a Vercel deployment would be overkill.
Projects: Give Claude a Permanent Memory
</figure>
The biggest friction point with AI assistants is the amnesia. Every session starts from scratch — you re-explain the project, re-paste the context, re-specify the tone. Projects fix this.
A Project is a persistent workspace that retains:
Your uploaded files (PDFs, spreadsheets, code, images) as a searchable Knowledge Base
Custom instructions — a permanent system prompt that shapes how Claude responds in that project
Full conversation history across every session
I have a Project for every non-trivial ongoing context: one for a research area I’m tracking, one for a codebase I maintain, one for writing work with a specific style guide. Claude walks into each session already briefed.
Here’s what a useful custom instruction looks like:
You are assisting with an open-source Rust library.
The codebase targets stable Rust, no nightly features.
Always suggest idiomatic Rust — prefer iterators over explicit loops,
use proper error propagation with `?`, and avoid `.unwrap()` in library code.
When in doubt, ask a clarifying question before generating code.
Set that once. Never type it again.
Tip: If you’re working on multiple related files, upload them all to the Project Knowledge Base. Claude can cross-reference between documents — useful for things like “does this API design contradict what we documented in the spec last month?”
Claude’s native Memory feature (plan-dependent) goes a step further and persists preferences across projects. Combined with a file-based memory system using a local MCP server, you can get cross-session recall that works simultaneously in Claude.ai, Claude Code, and the API.
Prompt Engineering That Actually Works
There are a few patterns worth internalising. Not because they’re magic, but because they cut rewrites.
The Role → Context → Task → Format → Constraints pattern:
Role: You are a senior backend engineer reviewing code for production readiness.
Context: This is a Node.js API endpoint that handles payment webhooks from Stripe.
Task: Review the code below and identify issues with error handling, idempotency, and security.
Format: Bullet points grouped by severity — Critical, Warning, Info.
Constraints: Be direct. If something is fine, say so. Don't pad the output.
Structuring prompts this way turns Claude from a chatbot into a structured thinking tool. The output is predictable and usually needs one revision rather than three.
Reverse prompting is the underrated one. Instead of jumping straight to output, ask Claude to interview you first:
I want to write a technical post-mortem for a production outage.
Before you draft anything, ask me the questions you need answered
to write a thorough and useful post-mortem.
This surfaces requirements you didn’t know you had. I use it constantly for anything where scoping matters more than speed.
Constraint cascades — iteratively layering constraints rather than front-loading them — work surprisingly well for refining outputs. Start broad, get a draft, then add: “Now tighten it to 400 words,” “Now make the tone more direct,” “Now add a concrete example in the second paragraph.” Each constraint sharpens without losing the thread.
Tip: Claude’s Custom Styles feature (in the UI: Normal / Explanatory / Formal / Custom) applies your preferred tone across every response without you specifying it per prompt. Set it once for your account.
Verification loops are what Anthropic engineers apparently use internally — asking Claude to check its own answer before returning it:
... Before responding, re-read your answer and verify that:
1. The code compiles without errors
2. You haven't introduced any off-by-one errors
3. The approach handles the empty list case
If any check fails, fix it first.
Anecdotally, this reduces obvious errors by a meaningful amount on tasks where correctness is load-bearing.
MCP and Connectors: Where It Gets Agentic
Claude Connectors are the consumer-facing integration layer: link Claude to your email, calendar, Google Drive, and external databases so it answers questions grounded in your actual context, not generic training data.
MCP (Model Context Protocol) is the developer-facing version of the same idea — a standardised protocol that lets Claude connect to thousands of external tools via a common interface. Live databases, filesystems, Supabase, GitHub, web scrapers — all accessible from chat or Claude Code without custom glue code.
The Advanced Tool Use API (currently in beta) pushes this further with two features that matter for complex workflows:
Tool Search Tool — Claude dynamically finds the right tool from a large registry without stuffing the entire toolset into the context window
Programmatic Tool Calling — Claude writes code to orchestrate multiple tools in a single pass, collapsing multi-step pipelines into one request
Combine Projects + MCP + the 200K token context window on Opus, and you can drop an entire codebase or legal document set into a persistent workspace and have Claude reason across it indefinitely. That’s a qualitatively different capability from what most people associate with “chat AI.”
Tip: Start with one MCP server, not ten. The Context7 MCP is a good first one — it fetches current library documentation on demand, so Claude stops hallucinating outdated API signatures.
The Payoff
The tooling has matured enough that the bottleneck is no longer what Claude can do — it’s whether you’ve configured the environment to let it. That setup investment pays back quickly. An afternoon spent on Projects, a prompt pattern you trust, and one MCP integration that hits a real data source you use daily will change how you work with it. Start there.
Enjoy Reading This Article?
Here are some more articles you might like to read next:
