[ entry ]claude

Meet Claude (and the era it opens)

What Claude actually is, the three primitives every surface relies on (MCPs, Skills, Plugins), the four rooms it lives in (Chat, Code, Cowork, Design), and why this is the era where the execution layer collapses into the model. A working introduction with one 50x example from my quarter.

Published: 10 May 2026
Reading time: 11 min read
Topic: claude
Language: EN

On a Wednesday in April I sat down to mock a clinic dashboard. I described what I wanted, Claude Design built the first version, I refined it through conversation, exported the prototype to Code, and shipped it to staging before lunch. Six months ago that day was a sprint.

This post is the working introduction I wish someone had handed me when I started. What Claude actually is, the three primitives every surface relies on, the four rooms it lives in, and the thesis I keep repeating to founders, students, and ops leads in Guatemala City who ask me what changed: the execution layer is collapsing into the model. Anyone, in any field, who does not learn to operate these tools will be moving manually next to peers moving by machine. That gap compounds.

What Claude is, briefly

Claude is Anthropic's model family. Three tiers right now: Opus 4.7 for the heaviest reasoning and agentic coding (the most capable generally available model, released April 2026), Sonnet 4.6 as the default production tier (most of your work), Haiku 4.5 for cheap and fast (high concurrency, low latency). You pick the tier by task difficulty. That is the engine.

The interesting part is the chassis. In the last twelve months Anthropic shipped MCP, Skills, Plugins, Claude Code, Claude Cowork, and Claude Design. Each one is a different way of letting the same model work in your real systems instead of in a chat window. The thesis of this post is that right now the chassis matters more than the engine.

The essence

Claude has stopped being "a smarter chatbot." It is closer to a worker you can hire, brief, and reshape. It reads your real files, writes back to your real apps, ships real outputs (decks, code, designs, reports), runs on a schedule without you in the chair, remembers context across sessions when you let it, and shows you its work before it acts. The chat surface is the smallest piece of all of that.

If you only chat with Claude, you have used about ten percent of what is on the shelf.

The three primitives every surface uses

Learn these names. You will hear them in every plugin description, every product page, every internal demo.

MCPs are the cables. MCP (Model Context Protocol) is the open standard Anthropic published so any app can let Claude read and write to it: Slack, Gmail, Notion, GitHub, Linear, Stripe, your CRM, your accounting system, your custom internal tool. Without MCP, Claude only sees what you paste into the chat. With MCP, it talks to the apps directly. Think of MCP as USB-C for AI: one shape that fits every device. The cost is that every connected app gets exactly the permissions you grant, no more and no less, so the security model is on you. The gain is a model that operates inside your stack instead of outside it. As of March 2026 there are more than 5,000 community MCP servers tracked across registries, with Anthropic curating around 2,000 in the official directory.

Skills are the recipes. A Skill is a folder of instructions, templates, sample data, and standing rules for one narrow task. "When I ask you to draft a monthly close, here are the steps, here is what good looks like, here is the template, here is who it's going to." Once written, the recipe runs the same way for every team member who invokes it. Anthropic and the community publish hundreds; you also write your own as you find tasks that recur. The point is to stop re-explaining your business to Claude every conversation.

Plugins are the kitchen-in-a-box. A Plugin packages skills, MCPs, slash commands, and sub-agents into a single install. A marketing plugin gives a marketer their full Claude setup in one click: the skills for drafting, the MCPs for HubSpot and Klaviyo, the sub-agents for campaign planning. A finance plugin does the same for a controller. Anthropic shipped 11 open-source plugins for Cowork at launch (Productivity, Marketing, Sales, Enterprise Search, and more); the community publishes more weekly.

Mental model: MCPs are the plumbing, Skills are the recipes, Plugins are the kitchen-in-a-box. Learn the three names and the rest of the ecosystem makes sense the next time someone shows you a demo.

The four rooms where Claude lives

The model is the same. The room you put it in changes what it does for you.

Claude Chat (claude.ai)

The conversation surface. Where most people start. Where most people stop. Use it for exploratory thinking, drafting, analysis, research. Inside Chat, the Projects feature persists context across sessions: upload your SOPs, your brand guide, your customer data once, every conversation in that Project starts already knowing them. This is the cheapest way to convert "smart intern who forgets you each morning" into "specialist who already knows the company."

Claude Code (CLI / IDE)

Claude with hands. Runs in your terminal, reads your repo, edits files, runs tests, opens pull requests, deploys to staging. The most underestimated surface among non-engineers, who assume "Code" means "for devs only." It does not. Anyone uses Code for scheduled data pulls, document generation, lightweight automations, scripted reports. You describe the work; Claude writes the script and runs it. Anthropic's 2026 Agentic Coding Trends report shows a single Claude Code agent at Delivery Hero merging more than 100 pull requests a day. The cost is a steep first week of learning the discipline. The gain is that the shell of your computer becomes a place where work happens for you instead of with you.

Claude Cowork

Knowledge-work power on the desktop. Cowork (research preview Jan 2026, expanded Feb 2026) is Claude reading and writing in your local files and apps directly: your Excel sheets, your PowerPoint decks, your Slack channels, your Google Workspace, your DocuSign, your Notion. Connectors include FactSet and S&P Global on the enterprise side. Open-source plugins for Productivity, Marketing, Sales, and Enterprise Search are available at install time. Cowork is the answer to "how do I get Code-level leverage on knowledge work that does not live in a repo?" Most non-engineers should start here.

Claude Design

Visual generation in conversation. Design (Anthropic Labs, April 2026, on Opus 4.7) builds prototypes, slides, one-pagers, and full-app designs. You describe what you want, Claude produces the first version, you refine through conversation, inline comments, direct edits, and custom sliders. During onboarding it reads your codebase and design files to learn your design system, so every project after the first uses your colors, typography, and components automatically. Export to PDF, URL, PPTX, or send straight to Canva. It hands off to Code cleanly: design intent in, production code out.

How they integrate in a real day: I draft a strategy doc in Chat with the company Project loaded, prototype the UI in Design, hand the prototype to Code for implementation, run the daily ops in Cowork. Same model, four rooms, four kinds of work. The handoffs are explicit, the audit trail is intact, the judgment lives in me.

Why this is a different era

Every previous wave of productivity tooling (the spreadsheet, the database, the web browser, the smartphone) accelerated specific pieces of knowledge work but left the execution layer mostly human. You still typed the email. You still wrote the SQL. You still designed the slide. You still pushed the deploy. The tools made each step faster, they did not remove the step.

What is shipping in 2026 removes the step. The execution layer (the part where intent turns into output) is collapsing into the model. A founder describes a feature, Code ships it. A marketer describes a campaign, Cowork drafts it, plans it, schedules it. An ops lead describes a recurring report, a Skill runs it forever. A designer describes a screen, Design builds it. The model still needs your judgment about what to do, for whom, with what trade-offs. Commercial judgment is still yours. The execution layer is not.

The field you are in does not protect you from this. A doctor with Claude does triage faster. A lawyer with Claude finds clauses faster. A teacher with Claude personalizes lessons faster. A bookkeeper with Claude closes books faster. A student with Claude learns faster. The leverage cuts across every category that produces or processes information. Anyone who does not learn to operate these tools will spend the next decade producing work manually next to peers producing it through machines, and that gap compounds quarter by quarter.

The entry cost is low. Twenty hours of disciplined practice with Chat plus Projects, plus one MCP wired into your daily app, plus one Skill written for a recurring task, plus one Plugin installed for your role. That is enough to start the leverage flywheel. The compounding does the rest.

What 50x looks like in practice

The cleanest example I have is from my own work this quarter, building Mozt's clinic-operations product.

The "before" baseline. A competent IE consultant in Guatemala designs operations for one clinic: process map, SOPs, training material, basic dashboard, three to six weeks of work, charges roughly $4,000 to $6,000, maybe touches one or two clinics in a quarter. That is the conventional throughput.

The "after" the toolchain shifted, in five working days:

Day 1, morning. Chat: draft the discovery questionnaire and process-map outline, bilingual, against my IE template Project. 30 minutes.
Day 1, afternoon. Cowork: pull the clinic's Excel reports, parse them, build the baseline-throughput chart, identify the binding constraint. 90 minutes including the reading.
Day 2. Code: scaffold the dashboard from scratch (Next.js, the design system, auth), wire it to the clinic's WhatsApp inbound through MCP, deploy to staging. Half a day.
Day 3. Design: produce the patient-facing UI, three iterations, exported to Canva for the printed material. Two hours.
Day 4 to 5. Skills + Plugins: package the entire flow as a clinic-onboarding plugin so every future clinic gets the same baseline in one install. One day.

End-to-end: one IE, five days, ready to repeat. Onboarding the second clinic takes a day. The tenth takes hours, because the plugin carries the standardized work and the judgment lives where it actually adds value.

Fifty times is not a round number for marketing. Six weeks (240 hours) compressed into five days (40 hours) is 6x by clock alone. Multiply by the parallelism (the same plugin onboarding ten clinics in the time the consultant onboards one), the embedded design system that previously needed a designer, the audit trail that previously needed a project manager, the agent that runs the daily checks instead of a junior, and the practical multiplier on what one operator delivers across a quarter sits in the 30x to 50x range. The number stops being a slogan once you put it on a calendar.

The trade-off is worth naming. The IE judgment about which constraint matters, which trade-off to take, which clinic to onboard, which question to ask the owner did not get cheaper. If anything it got rarer, because the models are now good enough that the bottleneck moved upward, from execution to selection. That migration is the whole story of this era.

What's next

This post is the working introduction. The next ones go deep, one tool at a time, with the patterns that actually move the needle in a LATAM operator's stack:

Claude Code for non-engineers. Scheduled data pulls, monthly reports, document generation. The shell as a workshop.
Claude Cowork for finance and ops. The SOP-as-Skill pattern. The audit-trail discipline. Why this is the highest-leverage surface for SMBs.
Claude Design for solo founders. Prototype to production without hiring a designer. The Code handoff that ships.
MCP for SMBs. Which integrations earn back their setup cost the fastest in a Guatemalan operations stack.
Plugins worth installing on day one. The short list that pays for itself in the first week.
The 50x patterns named. Which combinations of the four rooms actually compound, and which combinations look impressive but don't.

The model is good. The chassis is finally good. What's left is your reps.