How to Actually Use AI in 2026: The Complete Guide

Most people are still using ChatGPT the way they used Google in 2005: type a question, get an answer, close the tab. A lot of people aren't even asking the AI to use web search, and just relying on its "training date." Le gasp!

That worked fine when AI was a novelty in 2023, or 2024. In 2026, it's like owning a professional kitchen and only using the microwave. If you're feeling attacked right now, good. Channel that energy to upgrade your AI skills and keep scrolling.

We've been reflecting on this here at The Neuron, especially since so many of our readers are totally new to AI. This guide is our attempt to make it concrete. We're going to walk you through two things: the five levels of AI proficiency (the "what to learn" framework) and which specific tools to use for what (the "what to pick" guide). Ten minutes, actionable by tomorrow.

Let's get into it.

First up, the TL;DR

Most people are still using ChatGPT the way they used Google in 2005: type a question, get an answer, close the tab. A lot of people aren't even asking the AI to use web search, and just relying on its "training date." Le gasp!

That worked fine when AI was a novelty in 2023, or 2024. In 2026, it's like owning a professional kitchen and only using the microwave. If you're feeling attacked right now, good. Channel that energy to upgrade your AI skills in 2026 and keep scrolling…

We've been reflecting on this here at The Neuron, especially since so many of our readers are totally new to AI. So here's the framework we recommend for getting real, compounding value out of AI. Think of it as five levels.

Here's the stack:

• Level 1: Projects. Stop chatting in the main window. Create a project folder (ChatGPT, Claude, and Gemini all have them).
• • Inside, add custom instructions (persistent rules the AI follows every time), upload reference documents (style guide, brand voice, codebase), and set memories (facts it remembers across sessions).
  • This is the foundation for your work. Don't do any sort of work without this set up.
• Level 2: Prompting. Ya that's right, this is level two (not one). After your project is set up, you can then focus on how to prompt.
• • Simplest formula: Persona + Task + Context + Format. "You are a senior content strategist. Create a content plan for a tech blog targeting AI beginners. Present as a bulleted list." Goal, context, constraints. That's it.
• Level 3: Skills. Once you've gone back and forth enough to nail a task, package that conversation into a reusable skill. Then you can ask your AI at any time to use that skill to do that same task without memorizing or saving the prompt somewhere.
• • Ask: "Reverse-engineer this conversation into a skill using your skill creator skill I can call anytime." If it doesn't give you a doc you can "install," it didn't work right; see below for more.
  • This is a one click trick that will save you twenty minutes of prompting for something you already got your AI to do for you once before.
• Level 4: Automations. Once you've got skills you can call any time, now you can schedule them for your recurring tasks. Claude's Cowork, OpenAI's Codex, and Gemini's Opal and Scheduled Actions all support this.
• Level 5: Agents. These are AI that reason, act, and use tools in a loop.
• • Automations run tasks on a schedule; agents run toward a goal. They reason about what needs to happen, pick the right tools (or skills), act, check if it worked, and loop until the job is done.
  • Three ways to use them: for you (an OpenClaw or Claude Code agent that manages your calendar, triages your inbox, and files your expenses without being told each step), for your customers on your behalf (a support agent that reads tickets, pulls up account data, resolves issues, and only escalates what it can't handle), or as the product itself (an AI tutor, financial advisor, or research assistant where the agent IS the thing you sell).
  • The difference from Level 4: at Level 4, you decide what runs and when. At Level 5, the AI decides what and when. You give it "keep my inbox under 20 unread" and it figures out the filtering, replying, and archiving on its own.

Why this matters: The gap between "I use ChatGPT sometimes" and "AI saves me 10 hours a week" is almost entirely about moving up this stack. Most people are stuck at Level 2. The real productivity gains live at Levels 3-5.

Our take: You don't need to be a developer. You need to stop treating AI as a search engine and start treating it as a coworker who needs onboarding. Projects are the onboarding. Skills are the training. Automations are the job to be done every day. And agents are the coworker you interact with to get it all done.

Part 1: The 5-Level AI Proficiency Stack

Level 1: Projects (Your AI's Home Base)

Every major AI platform now supports project folders: ChatGPT Projects, Claude Projects, and Gemini Gems. The concept is the same across all three: a persistent workspace where your chats, files, and instructions live together.

Inside a project, you set up three things:

Custom instructions are the rules your AI follows every conversation inside that project. In ChatGPT, you access these through Project Settings. In Claude, they're called "Custom instructions" in the project sidebar (they might be called something slightly different in ChatGPT, like "Project Instructions"). Think of these as a permanent system prompt.

Drop in things like: "You are a marketing strategist for a SaaS startup. Write in a conversational tone. Use bullet points for lists. Always suggest A/B test ideas." Every chat in that project inherits those rules automatically. OpenAI's docs note that each custom instructions field has a 1,500-character cap, so keep them focused.

OpenAI's own prompt engineering guide calls this out: "A GPT model is like a junior coworker. They'll perform best with explicit instructions to create a specific output." Custom instructions are how you give those explicit instructions once instead of repeating them forever.

Project knowledge is where you upload reference documents the AI can draw from. Your brand guidelines, your product spec, your company's writing style guide, competitive research, client briefs. Whatever the AI needs to do its job without you re-explaining every time. ChatGPT supports up to 20 files per custom GPT (512 MB each). Claude Projects support uploaded docs that sit in context across all chats.

Memories are facts the AI remembers across sessions. In ChatGPT, you choose when creating a project whether memory is project-only (contained to that workspace) or default (shared across your account). Claude has both project-specific memories and account-wide memories you can edit. The practical difference: use project memory for "this client prefers formal tone" and account memory for "I'm a marketing director at a fintech startup."

The key insight: most people skip this step entirely. They open a blank chat, re-explain their context every time, and wonder why the AI gives generic answers. Projects solve this. They're the difference between talking to a stranger and talking to a coworker who's been briefed.

Level 2: Prompting (The Conversation)

Once your project is set up, prompting gets much simpler because half the context is already loaded. You don't need to say "I'm a marketing director at a fintech startup" every time; the project already knows.

For the prompts themselves, the simplest useful formula is: Persona + Task + Context + Format.

• Persona: "You are a senior data analyst."
• Task: "Analyze this sales data and identify the top 3 trends."
• Context: "This is Q1 2026 data for our SaaS product. We're preparing for a board meeting."
• Format: "Present as a 1-page executive summary with bullet points."

A more structured approach some people like is RASCEF:

• Role: You are a senior content strategist.
• Action: Create a content plan for a new tech blog.
• Steps: 1. Research trending topics. 2. Suggest 5 article titles. 3. Outline main points for each. 4. Identify keywords.
• Context: The blog targets beginners in AI.
• Examples: Use a friendly, engaging tone.
• Format: Present as a bulleted list in a markdown code block.

But honestly, you don't need to memorize any framework. Both OpenAI and Anthropic publish excellent prompt engineering guides, and they boil down to the same three ingredients: a clear goal ("what do I want?"), relevant context ("what does the AI need to know?"), and output constraints ("how should the answer look?"). If your prompt has those three elements, it will work.

Here are the highest-leverage tips from those guides that most people miss:

Anthropic's golden rule for Claude: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too." The Claude 4.6 best practices page also recommends being explicit about what you want: if you want above-and-beyond behavior, explicitly request it rather than relying on the model to infer this from vague prompts. Claude 4.6 is also more concise than previous versions; it may skip verbal summaries after tool calls and jump straight to the next action. If you want more visibility, tell it: "After completing a task, provide a quick summary of the work you've done."

OpenAI's key insight for GPT-5.4: The GPT-5.4 prompt guidance says the biggest gains come from three things: choosing the right reasoning effort for the task, using explicit grounding and citation rules, and giving the model a precise definition of what "done" looks like. GPT-5.4 also has a reasoning effort setting (none, low, medium, high) that controls how much it "thinks" before responding. Their recommendation: start with none for simple tasks, medium for research or analysis, and only go higher if your results actually improve. Most people overcrank this and just burn tokens.

Both agree on one thing: tell the AI what to do, not what not to do. Instead of "Don't use jargon," try "Explain in plain language a non-technical reader would understand." Positive instructions consistently outperform negative ones across both platforms.

Anthropic's prompting best practices also recommend wrapping different types of content in XML tags (stuff that goes inside a side carrot like this: "<" and then you do on on the other side like this: ">") to help Claude parse complex prompts. For example, providing three to five examples of what you're looking for inside tags dramatically improve accuracy and consistency.

One important nuance: the field is shifting from "prompt engineering" to what practitioners call context engineering. Anthropic published a blog post in late 2025 arguing that as we build more capable agents, "we need strategies for managing the entire context state," not just the prompt. The Prompt Engineering Guide (the most widely used open resource in the field) now has an entire context engineering section. The core idea: the prompt is just one piece of what the model sees. The documents, the memories, the conversation history, the tool definitions; all of that is context, and managing it well matters more than wordsmithing one message.

This is why Level 1 (projects) comes before Level 2 (prompting). The context you set up in your project is the foundation. The prompt is just the question you ask inside that foundation.

Level 3: Skills (Your AI's Muscle Memory)

This is where most people plateau, and it's where the real compounding gains begin.

Here's the scenario: you've spent 15 minutes going back and forth with Claude on a specific task; say, formatting a weekly report from raw data. You've refined the instructions, corrected the tone, specified the sections. The result is perfect. And next week, you'll have to do it all over again.

Skills fix this. Instead of re-prompting, you ask the AI: "Reverse-engineer this conversation and package it into a reusable skill I can call anytime I need to do this again."

On Claude, Skills are structured SKILL.md files that Claude loads dynamically when relevant. Anthropic published a 32-page guide on building them, and there are now 1,000+ community skills available in the official repository. The Anthropic Skills announcement explains the concept with a kitchen analogy: "MCP provides the professional kitchen (access to tools, ingredients, and equipment). Skills provide the recipes (step-by-step instructions for creating something valuable)."

ChatGPT now also has Skills. If you use ChatGPT and never made a skill before, click this!

1. To get to your Skills library, follow these steps:
2. 1. On Claude, go to Customize > Skills
   2. 1. Or just type & bookmark this url: https://claude.ai/customize/skills
   2. On ChatGPT, click on your Profile in the bottom left corner > Skills
   3. 1. Or just type & bookmark this url: https://chatgpt.com/skills

The mental model: Level 2 is like cooking a meal from scratch every night. Level 3 is like meal-prepping. Same quality, fraction of the time. And here's what makes it compound: every skill you build makes the next project faster. After a few months, you have a library of skills that covers most of your recurring work.

For enterprise teams (or those on the ChatGPT Teams account): you can also share these skills across your organization / teams account!

Level 4: Automations (AI That Works While You Sleep)

Once you have skills that run reliably, the next step is scheduling them. This turns your AI from a tool you use into a worker that runs independently.

Claude's Cowork supports scheduling automations and, with the new computer use feature, Claude can open apps, click through UIs, and complete tasks on your machine. OpenAI's Codex is purpose-built for this: scheduled coding tasks, background processes, recurring reports. Google's Opal (Gemini's automation layer) connects automations across Google Workspace.

Practical examples: generate a weekly competitive analysis every Monday morning. Summarize your Slack channels at 5 PM daily. Audit your codebase for security vulnerabilities on a schedule. Pull and format data from three sources into a single dashboard every Friday.

The key difference between Level 3 and Level 4 is human involvement. At Level 3, you trigger the skill. At Level 4, the skill runs itself. You're moving from "AI as a tool I pick up" to "AI as a process that runs in the background."

Level 5: Agents (AI With Agency)

Automations run tasks on a schedule. Agents run toward a goal. They reason about what needs to happen, pick the right tools (or skills), act, check if it worked, and loop until the job is done. You stop assigning tasks and start assigning outcomes.

The difference from Level 4 is who's making the decisions. At Level 4, you decide what runs and when. At Level 5, the AI decides what to do next. You give it "keep my inbox under 20 unread" and it figures out the filtering, replying, and archiving on its own.

Three ways to use them:

For you: An OpenClaw or Claude Code agent that manages your calendar, triages your inbox, and files your expenses without being told each step. The infrastructure has matured rapidly: Anthropic's Claude Agent SDK, MCP (Model Context Protocol, now governed by the Linux Foundation with 97M+ monthly SDK downloads and adoption by Anthropic, OpenAI, Google, and Microsoft), and persistent memory all support agents that run for hours or days autonomously.

For your customers on your behalf: A support agent that reads tickets, pulls up account data, resolves issues, and only escalates what it can't handle. A sales agent that qualifies leads by researching companies and drafting personalized outreach. You build the agent once; it serves every customer who comes through the door.

As the product itself: An AI tutor, financial advisor, or research assistant where the agent IS the thing you sell. This is where the economics change fundamentally: the marginal cost of serving a customer approaches zero.

OpenAI's GPT-5.4 prompt guidance dedicated an entire section to agentic scaffolding, noting that GPT-5.4 is trained to operate anywhere along the spectrum, from making high-level decisions under ambiguous circumstances to handling focused, well-defined tasks. Anthropic's prompting best practices for Claude 4.6 now include an entire agentic systems section covering tool use, multi-turn inference, and context management for long-running agents.

If you're starting to build agents, the official docs surface three practical patterns worth knowing:

Context awareness: Claude 4.6 can track its remaining context window throughout a conversation, so it knows how much space it has left to work. Anthropic recommends telling it: "Your context window will be automatically compacted as it approaches its limit, allowing you to continue working indefinitely. Do not stop tasks early due to token budget concerns." This single instruction prevents agents from artificially wrapping up work too soon.

Completeness contracts: OpenAI's GPT-5.4 guide recommends telling agents to treat a task as incomplete until all requested items are covered or explicitly marked as blocked. A common failure mode is the agent finishing after partial coverage. The fix: "Keep an internal checklist of required deliverables. Confirm coverage before finalizing."

Safety guardrails: Both platforms recommend telling agents to confirm before taking irreversible actions. Anthropic's suggested framing: "You are encouraged to take local, reversible actions like editing files or running tests, but for actions that are hard to reverse or affect shared systems, ask the user before proceeding." This is the difference between a useful agent and one that force-pushes to main at 3 AM.

Part 2: Which AI Tool for What

Now that you know the stack, here's which tool to use at each level.

Step 1: Pick Your Daily Driver

Your main AI for everyday work. The two most reliable options right now:

• GPT-5.4 (ChatGPT's paid tier): The broadest model with the most integrations. Strong at web search, image generation, and general-purpose tasks. OpenAI's GPT-5.4 prompt guidance notes it's designed for "production-grade assistants and agents that need strong multi-step reasoning, evidence-rich synthesis, and reliable performance over long contexts."
• Claude Opus 4.6 (Claude's paid tier): The strongest writer and deepest reasoner. Anthropic's docs describe it as having "a more concise and natural communication style," more direct and less verbose than previous models. Best for nuanced analysis, long-form writing, and tasks that require genuine understanding.

Gemini 3.1 and Grok 4.20 are strong contenders, but less consistent in our testing; though of course you can choose those. If you want to compare models on actual performance data instead of marketing claims, Artificial Analysis tracks benchmarks, pricing, speed, and quality across every major model in real time.

Step 2: Pick Your Platform App

This is the tool that does work on your behalf, on your actual computer. Our two recommendations:

• Claude Desktop app (with Cowork): Best for knowledge workers who want file management, scheduling, computer use, and automations in one place. Cowork lets you describe the outcome you want and Claude creates a plan, reasons across your tools and files, and carries work forward.
• OpenAI's Codex app: Best for developers who want a coding agent that lives in their terminal. Purpose-built for scheduled coding tasks, background processes, and codebase management.

We don't find Gemini or Grok's platforms as compelling as the two above, even though there are times when we may need to or prefer to use Gemini or Grok for specific tasks. Both apps have room to grow and become more useful as central hubs; Grok has a better UI/UX, while Gemini has lots of cool tools and features spread across Google's broad product catalogue, making it difficult to have a true central hub like the two above.

Step 3: Use Each Tool for Its Superpower

Every major AI tool has something it does better than the rest. The biggest mistake people make is trying to use one tool for everything; while you need a daily driver for the majority of your task, there are edge cases where you should just use the best tool for the job.

Here's our current ranking on which of the major AI apps is best for what:

• ChatGPT (basic app) for: One-off web searches and image generation. The fastest path from "I have a question" to "here's an answer with sources." GPT-5.4's web browsing is the most seamless of any model, and Image integration means you can generate images mid-conversation. Pro tip from OpenAI's docs: GPT-5.4 has a reasoning effort knob. For quick searches and simple questions, keep it on "none" or "low." Only crank it to "medium" or "high" for research-heavy synthesis where you actually need deep thinking. Most people leave it on the default and burn tokens for no reason.
• Gemini for: Accessing YouTube links (it can watch and summarize full videos, something no other model can do natively), creating images (with Nano Banana), building automations with Opal (Google's automation layer that connects across Gmail, Calendar, Docs, and Sheets), and deep research for pulling links and sources. Caveat: the link quality from Gemini's deep research is excellent; the report text itself tends to be sloppy in our experience. Use it for the sources, then rewrite.
• Claude for: Deep research for niche links (it can't access many mainstream websites since it doesn't have the same licensing agreements as OpenAI or Gemini, but it's good at searching far and wide for diamond in the rough links), working with files on your computer via Cowork, scheduling recurring automations, and writing. The strongest writer of the bunch (imo, though Corey would disagree), and the most reliable for tasks that require following complex, multi-step instructions. Claude's Skills system is the most mature implementation of reusable AI capabilities. Pro tip from Anthropic's docs: Claude 4.6 has improved vision and can analyze images, screenshots, and even videos broken into frames. Give it a "crop tool" instruction (tell it to zoom into specific regions of an image) and accuracy goes up significantly (that said, Claude's vision is still far worse than Gemini or others). Also, Claude 4.6 naturally delegates to subagents for parallel work; if it's spawning too many, tell it: "Use subagents only when tasks can run in parallel or require isolated context. For simple tasks, work directly."
• Grok for: Reading X (Twitter) links. Grok can access posts, threads, and profiles that other models can't fetch. If your workflow involves monitoring X for trends, competitive intelligence, or industry commentary, Grok is the only model that can pull that data natively.
• Codex (from OpenAI) for: Scheduled automations, organizing coding work, and coding (obviously). Purpose-built for developers who want background coding processes. OpenAI's GPT-5.4 prompt guidance has an entire section on optimizing coding tasks, including frontend engineering with Cursor integration. Pro tip from OpenAI's docs: GPT-5.4 can over-engineer by default, adding unnecessary abstractions or building in flexibility you didn't ask for. Add this to your instructions: "Only make changes that are directly requested or clearly necessary. Don't add features, refactor code, or make 'improvements' beyond what was asked." Also supports compaction for multi-hour sessions that exceed a single context window.

Beyond these core tools, there are many more niche applications for standalone use-cases. These might appeal to you if they are in your business vertical or you just want some out of the box solutions for these specific use-cases. They are...

🏢 Professional Services

• AI for Legal — Harvey ($1.2B raised, $11B valuation as of March 2026, Sequoia/GIC/a16z, used by Paul Weiss, O'Melveny)
• AI for Healthcare — Abridge ($212M raised, partners with Epic, used by 100+ health systems for AI clinical documentation); Hippocratic AI ($137M, AI health agents for non-diagnostic care)
• AI for Insurance — Tractable ($185M raised, AI computer vision that assesses auto and property damage from photos with 95% accuracy, settling claims in minutes instead of days); Lemonade (public, $631M raised, AI-native renters, homeowners, pet, car, and life insurance; chatbot processes 55% of claims instantly and famously paid a stolen laptop claim in 3 seconds)
• AI for Accounting — Puzzle ($66.5M raised, AI-native general ledger that automates 98% of categorization for startups); Numeral ($57M raised, $350M valuation, AI sales tax compliance across 11,000+ jurisdictions)
• AI for HR/Recruiting — Eightfold AI ($424M raised, $2.1B valuation, AI talent intelligence for hiring, retention, and workforce planning; used by Bayer, Capital One, Moderna); Paradox ($200M+ raised, Olivia AI recruiting assistant used by McDonald's, Unilever)

🎨 Creative Tools

• AI for Voice — ElevenLabs ($500M Series D, $11B valuation, $330M ARR, used by Meta/Salesforce/Revolut); Voxtral TTS by Mistral (open-source, runs on a smartwatch, 9 languages, 90ms time-to-first-audio, matches or beats ElevenLabs on naturalness in human evals; launched March 26, 2026)
• AI for Video — Runway ($237M raised, Gen-3 Alpha model); Seedance 2.0 by ByteDance (went viral Feb 2026, 2K resolution with lip-synced audio); Luma (Dream Machine, fast video generation from text and images); LTX Studio by Lightricks (AI video creation platform with storyboard-to-video pipeline and consistent characters across scenes)
• AI for Music — Suno ($125M Series B, generates full songs from text prompts); Udio ($50M, similar capabilities)
• AI for Design — Canva (Magic Studio AI suite, $40B+ valuation); MagicPath (AI design tool with Figma Connect)
• AI for Presentations — Gamma ($30M raised, AI-generated presentations/docs/sites); Beautiful.ai

💻 Developer & Builder Tools

• AI for Coding — Cursor (Anysphere, $900M+ raised, $9.9B valuation, AI-native IDE with inline editing and tab completion); Windsurf (Codeium, $300M raised, AI IDE with multi-file "Cascade" flows); Factory ($70M+ raised, Sequoia/NEA/NVIDIA, autonomous "Droids" agents ranked #1 on Terminal Bench for migrations, refactors, and feature development; used by MongoDB, EY, Bayer)
• AI for App Building — v0 by Vercel (AI that generates full-stack web apps and components from natural language prompts, deploys instantly to Vercel); Replit ($200M+ raised, $1.16B valuation, AI-powered cloud IDE that builds, hosts, and deploys apps from a single prompt)

🧠 Knowledge & Productivity

• AI for Search — Perplexity ($400M Series E, $24B valuation, 1B+ monthly queries, bidding on Chrome)
• AI for Education — NotebookLM (Google, free, turns any document into an interactive AI tutor with auto-generated podcasts, study guides, and Q&A; went viral multiple times); Oboe ($20M, a16z-led Series A, generates comprehensive personalized courses on any topic in seconds)
• AI for Language Learning — Speak ($78M+ raised, AI-powered conversation practice, big in South Korea/Japan); ELSA ($53M+ raised, 50M+ users, AI pronunciation coaching)
• AI for Translation — DeepL ($300M raised, consistently outperforms Google Translate on accuracy benchmarks)
• AI for Organizing Your Work — Notion AI (AI Q&A, summarization, and drafting embedded across your entire workspace); Obsidian (local-first markdown knowledge base with AI plugins for search, linking, and surfacing connections across your notes); Letta (AI agents with persistent memory that remember context across conversations and evolve over time)
• AI for Meetings & Transcription — Otter.ai (AI meeting assistant, real-time transcription and summaries); Granola (AI notepad that enhances your own meeting notes with transcript context, not a passive recorder); Fireflies.ai (transcription + auto-generated action items)
• AI for Finding Things — Yutori Scouts (always-on AI agents that monitor the web for flights, deals, price drops, papers, or anything you describe; just launched iOS app March 2026)

📈 Sales, Marketing & Commerce

• AI for Sales/CRM — Clay ($178M raised, AI-powered data enrichment and outreach); Apollo.io (AI prospecting platform)
• AI for Customer Support — Intercom (Fin AI agent resolves 80%+ of queries); Sierra ($485M raised, $14B valuation, founded by ex-Salesforce CEO Bret Taylor)
• AI for Marketing — Typeface ($165M raised, a16z + Lightspeed, enterprise content generation that learns your brand voice; used by Nestlé, Danone, Bayer); Helena (autonomous AI marketer that runs your SEO, ads, social, and email campaigns while you sleep)
• AI for Content at Scale — Jasper ($300M raised, enterprise AI content platform); Copy.ai (AI workflows for GTM teams)
• AI for Payments — Stripe (Agent Toolkit + Agentic Commerce Suite that lets AI agents create payments, issue virtual cards, and sell through AI surfaces; partners with Anthropic, OpenAI, Perplexity)
• AI for Shopping — Perplexity Shopping (AI-powered product search and comparison built into Perplexity Pro); Amazon Rufus (AI shopping assistant built into the Amazon app that answers product questions, compares items, and makes recommendations from Amazon's catalog)

💰 Finance & Operations

• AI for Finance — Ramp ($2.1B raised, AI-powered corporate cards + expense management); Brex (AI financial OS for startups)
• AI for Data/Analytics — Glean ($260M Series E, $4.6B valuation, enterprise AI search across internal tools); Databricks ($134B valuation, IPO planned Q2 2026)
• AI for Task Automation — Tasklet (from the Firebase co-founder, backed by USV/Lightspeed, describe what you want in plain English and it runs 24/7)

🏗️ Industry & Physical World

• AI for Defense — Anduril ($3.4B raised, $28B valuation, autonomous defense systems); Shield AI ($900M+ raised, $5.3B valuation, autonomous drones and AI pilot)
• AI for Construction — Procore (public, AI scheduling and risk forecasting, used in 70%+ of major builds); Buildots ($166M raised, Intel Capital backed, AI computer vision that tracks construction progress from 360° hardhat cameras against the BIM model; reduced project delays 50%)
• AI for Robotics — Figure AI ($675M raise, Amazon deal validated commercial deployment); Skild AI ($1.4B raised in Jan 2026 for robot foundation models)
• AI for Science & Drug Discovery — Isomorphic Labs ($600M raised, Alphabet/DeepMind spinout, Nobel Prize-winning AlphaFold tech, $3B+ in pharma partnerships with J&J/Novartis/Eli Lilly/Sanofi/Merck); Insilico Medicine ($410M+ raised, first end-to-end AI-designed drug in Phase 2, $2.75B Eli Lilly deal)
• AI for Autonomous Vehicles — Waymo ($16B round, largest AV deal in history, operating commercially in 6 US cities, expanding to Tokyo and London); Waabi ($600M+ raised, autonomous trucking with Uber Freight partnership)
• AI for Cybersecurity — Wiz ($2B+ raised, $12B valuation, cloud security platform; Google nearly acquired for $23B); Abnormal Security ($285M raised, AI-native email security)

🤖 Agents & Platforms

• AI for Running Your Whole Life — OpenClaw (the "operating system for personal AI," fastest-growing open-source project ever; enterprise variants include NVIDIA's NemoClaw for security and NEAR's TitanClaw for decentralized agents); Manus (autonomous agent that executes complex tasks in a cloud browser); Genspark (deploys specialized AI agents for research, travel, shopping, and coding)

Note: we may update this list as time goes on; we've been meaning to put together a definitive tool list like this at some point.

Bonus Level: Running Models Locally

If you want to run AI models completely offline, on your own hardware, with zero cloud fees and full privacy, then you'll want to use one of these tools:

• LM Studio: The easiest GUI for non-technical users. Download the app, search for a model (it tells you which ones your hardware can handle), click download, click chat. Free, works offline, supports Mac, Windows, and Linux. Built on llama.cpp under the hood.
• Unsloth Studio: Launched March 2026. A local web UI for running, training, and comparing models side-by-side. Their dynamic GGUFs (optimized compressed model files) are the gold standard for local inference; they consistently outperform other compression methods on accuracy benchmarks. You can even connect local models to Claude Code or Codex as backends.
• Ollama: One terminal command: ollama run llama3.1. Best for developers who prefer CLI (command line interface) communications. Over 100K GitHub stars, the most popular local model runner.
• Hugging Face: The library where all open models live. Browse, download, and deploy thousands of free models. Also home to the community that uploads new models daily.

The Bottom Line

The gap between casual AI users and power users comes down to one thing: structure. Power users don't use fancier prompts. They build systems: projects that remember context, skills that encode expertise, automations that handle the recurring work, and agents they interact with to get it all done.

The industry is recognizing this shift. What used to be called "prompt engineering" is evolving into context engineering; the discipline of managing everything the model sees, not just the words you type. Andrej Karpathy, Shopify CEO Tobi Lütke, and Gartner have all named context engineering as a defining skill for 2026. You could take this one step further and say we're all going to be agent engineers; as per usual, Latent Space was ahead of the curve with this.

Anyway, you don't need to reach Level 5 this week. Start with Level 1. Create a project for your most important recurring task. Add your context. Write custom instructions. And iterate. See what works. Adjust, edit, and tweak. You'll feel the difference as you continuously update and change your instructions.

Then slowly, when you're ready, package your first skill. As you learn what works best on a task by task basis, turn those successful task completions into skills. And turn those recurring skill requests into automations. Sooner than you think, you'll be able to schedule your first automation and effectively build your first agent.

Each level compounds on the last. Projects are the onboarding. Skills are the training. Automations are the job to be done every day. And agents are the coworker you interact with to get it all done.

Resources mentioned in this guide:

• OpenAI Prompt Engineering Guide
• OpenAI GPT-5.4 Prompt Guidance
• Using GPT-5.4 (Model Guide)
• Introducing GPT-5.4
• ChatGPT Projects
• ChatGPT Custom Instructions
• Anthropic Prompting Best Practices (Claude 4.6)
• Anthropic Prompt Engineering Overview
• Anthropic Interactive Prompt Engineering Tutorial (foundational concepts; exercises use older models but techniques apply to Claude 4.6)
• Claude Skills Guide
• The Complete Guide to Building Skills for Claude (32-page PDF)
• Anthropic Skills Repository (1,000+ community skills)
• OpenAI Custom GPTs Guidelines
• Artificial Analysis (model benchmarks)
• LM Studio
• Unsloth Studio
• Ollama
• Hugging Face
• Prompt Engineering Guide (promptingguide.ai)