How to actually use Claude, ChatGPT, and Gemini in 2026: the workflow stack

Most of the prompt engineering advice circulating on X and YouTube was written for 2023. The platforms have changed since then. In 2026, the meaningful skill is no longer crafting a single clever prompt — it's designing a workflow that uses Projects, Skills, MCP connectors, Workspace Agents, and Gems together. This post walks through what each of those actually is, where they overlap, and how a working professional should think about the stack.

The shift you probably missed

If your AI use still looks like opening a chat window, typing a request, and copying the output somewhere else — you're using a 2023 workflow on 2026 tools. The platforms have all quietly added a persistence layer underneath the chat. Most people haven't moved up to it.

Here's the shift in one sentence: the meaningful unit of AI work is no longer the prompt. It's the configured workspace.

Claude has Projects, Skills, MCP connectors, and Plugins. ChatGPT has Workspace Agents and Codex-powered Apps. Gemini has Gems, Deep Research, and a 1M-token context window across Workspace. Each of these does the same fundamental thing — it captures context, process, and tool access in a reusable artifact so you don't retype the same instructions every time.

This matters because almost all the "prompt engineering" content currently being published treats AI like a stateless chatbot. It teaches techniques for getting a better single response. That's still useful at the margin. But the leverage in 2026 is upstream of the prompt: it's in the layer that holds your context, your tools, and your standards so that every prompt starts from a better baseline.

The five components that matter

Here's how the current AI toolchain actually fits together for a working professional. I'll cover what each thing is, what job it does, and where the differences matter.

1. Projects (Claude) — persistent workspaces

A Project in Claude is a container that holds context, reference files, and custom instructions for an ongoing area of work. You set it up once. Every conversation inside that Project automatically inherits the context.

Concrete example: a Project called "Client X — Marketing Strategy" with the client's brand guidelines uploaded, three previous strategy documents as reference, and a system prompt saying "You are a senior marketing strategist familiar with this client's voice and goals. Every recommendation should reference the brand guidelines."

Every time you start a chat inside that Project, Claude already knows the context. You don't paste the brand guidelines into every conversation. You don't re-explain who the client is. The Project does that work.

Same output. One-twentieth the typing.

2. Skills (Claude) — reusable process recipes

A Skill is a folder with a SKILL.md file that teaches Claude how to do one specific task the same way every time. It's a saved playbook. Anthropic launched Skills in October 2025 and substantially expanded them in early 2026.

The mechanical structure: a SKILL.md file with YAML frontmatter (a name and a description that tells Claude when to invoke the skill) plus markdown instructions for the task itself. You can also include supporting files — templates, reference data, brand guidelines — that the skill loads when invoked.

Concrete example: a Skill called "weekly-client-report" that produces your weekly client report in your exact format, with your voice, pulling from the right reference files. You write it once. From then on, you say "draft this week's client report" and Claude follows the saved playbook.

The crucial detail most people miss: Skills don't all load into memory at once. Claude reads the skill names and descriptions, decides which is relevant for the current task, and loads only that one. You can have a library of 100 skills and Claude won't get cluttered. This is called "progressive disclosure" and it's the architectural reason Skills scale better than long custom-instruction blocks.

3. MCP connectors — direct access to your tools

MCP (Model Context Protocol) is an open standard that lets AI assistants connect directly to external tools, databases, and apps. Instead of copying data into a chat, MCP servers let Claude query your actual systems in real time.

As of mid-2026, Claude has 20+ MCP connectors for major business software — Slack, Notion, Linear, Gmail, Calendar, Drive, plus industry-specific ones like the 20+ legal-industry connectors Anthropic released this month. Gemini Enterprise now supports custom MCP servers as well. ChatGPT Workspace Agents use a similar concept under different naming.

What this changes practically: AI is no longer a bounded chatbot. With MCP connectors enabled, you can ask Claude to "summarize the last week of customer support tickets and surface the three most common complaints" and it will actually query your support system, read the tickets, and produce the analysis. Without MCP, you'd have to export the data, paste it into chat, and lose half the context.

The trap most users fall into: enabling MCP connectors but still working as if the AI doesn't have them. They still copy and paste. They still summarize their own data before showing it. The leverage comes from trusting the connection and asking questions that assume AI can see what you can see.

4. Workspace Agents (ChatGPT) — persistent multi-step automation

OpenAI launched Workspace Agents in April 2026 as the successor to custom GPTs. They're powered by Codex and run continuously in the cloud — meaning they can take actions on schedules or in response to triggers, even when you're not online.

A Workspace Agent is closer to a configured employee than a chatbot. You describe a workflow in natural language ("every Friday at 4pm, pull the week's sales data from Salesforce, generate a summary, and post it to the #sales-leadership Slack channel"). The platform builds it, connects the tools, and runs it on schedule.

Where this diverges from Claude: Workspace Agents run autonomously and can be shared across an organization. A team builds an agent once and the whole team uses it from ChatGPT or directly in Slack. They have persistent memory across runs. They can request approvals at decision points and wait for human input.

The relevant 2026 detail: Workspace Agents are available on Business, Enterprise, Edu, and Teachers plans. Custom GPTs still exist for consumer users but OpenAI has signaled they'll be phased out of business plans in favor of the Workspace Agents model.

5. Gems (Gemini) — grounded custom assistants

Gems are Google's version of custom AI roles, integrated tightly with Google Workspace. You configure a Gem with a specific persona ("sales email reviewer," "meeting notes summarizer") and it applies that role automatically without you re-explaining context each session.

The distinguishing feature: Gems can be grounded in your Google Drive content. If you build a Gem for "client X brand review" and point it at the relevant Drive folder, every interaction draws on the actual files in that folder. Combined with Gemini's 1M-token context window, this lets you ask questions across entire knowledge bases without the manual setup other platforms require.

Gemini also has Deep Research — an autonomous research mode that searches across multiple sources, synthesizes findings, and produces a structured report. As of Google AI Pro ($19.99/month), users get 20 Deep Research sessions per day. This isn't a chatbot feature; it's a different kind of AI use entirely, designed for the kind of multi-source analysis that previously took hours.

How they fit together

The mistake people make once they learn about these features is treating them as competing options — "which platform should I pick?" That's the wrong question. Most working professionals will end up using all three platforms because each has genuine strengths.

The right framing: each component answers a different question.

Most professionals use this stack badly because they only know one or two pieces. They have a few Projects but no Skills. They have a single ChatGPT chat thread they keep reusing. They've heard of Gems but never set one up. The unused capacity is the largest source of wasted time in most knowledge work in 2026.

The skill that still matters

Briefing AI clearly still matters. Context, role, constraints, examples, verification — these fundamentals didn't go away. But they now operate at a different layer.

In 2023, the brief lived in the prompt. You typed the whole brief every time, and the quality of your typing determined the quality of the output.

In 2026, the brief lives in the workspace. You configure your Projects, Skills, and Gems with the briefing already encoded. Then your day-to-day prompts can be one sentence because all the context is already loaded.

This isn't a small change. It means the people getting reliable output from AI in 2026 aren't necessarily better at writing prompts. They're better at building the persistent layer. The prompt is the tip of the iceberg.

A practical diagnostic

Here's an honest test for whether your AI workflow has moved past 2023. Answer these questions truthfully:

• Do you have at least one Project (or equivalent) configured for an ongoing area of work?
• Have you ever created a Skill, Workspace Agent, or Gem and reused it more than five times?
• When you ask AI for help with a recurring task, does it already know your standards, brand voice, or process — or do you re-explain those every time?
• Have you enabled at least one MCP connector that lets AI read data from your actual tools (Slack, Drive, Notion, etc.)?
• Have you scheduled any AI task to run automatically (weekly report, daily summary, monitoring)?

If you answered "no" to most of these, you're using 2023 workflows on 2026 tools. The good news is that fixing this is the highest-leverage thing you can do with AI right now. The technology is already capable; most users just haven't moved up to it.

Where to start

You don't need to set up all five components at once. Pick the one that hurts most.

If you keep retyping the same context every time you start a new chat, your first move is Projects (Claude) or a Workspace Agent (ChatGPT). Pick the platform you use most.

If you keep doing the same kind of task with slightly different inputs — drafting emails in your voice, writing weekly reports, formatting data the same way — your first move is a Skill (Claude) or saved process (Workspace Agent).

If you keep copying and pasting data into chat from another tool, your first move is an MCP connector (Claude) or the equivalent Workspace Agent tool integration. The friction of copy-paste is the friction of an outdated workflow.

If you need to do real research across many sources, your first move is Gemini Deep Research. It's the single best capability in this category as of mid-2026.

What this changes about how you should think about AI education

Most current AI education — courses, YouTube videos, X threads — is written about prompts. That made sense in 2023. It makes less sense in 2026.

The information that's actually scarce in 2026 isn't "how to write a better prompt." It's "how to build the workspace layer so your prompts can be shorter and your outputs can be more reliable." That requires a different kind of teaching: less clever phrasing, more structural setup.

This is why Albis exists. Single prompts are still useful at the margin. But the leverage in 2026 is upstream — in the persistent layer where context, process, and tools live. That's what a structured workflow course teaches that scattered free content rarely does.

Frequently asked questions