The power user

Model pickers, reasoning modes, data analysis, tools you build by talking, and assistants that act.

6 lessons · For: AI is already in your week. This course is about making it infrastructure: model choice, serious research, data work, and the first taste of agents. · Product details last reviewed July 2026

Past the everyday features, the products reward people who learn their depths: the model menus most users never open, research agents that work for twenty minutes at a stretch, analysis tools that write real code against your data, and the new agent modes that click, browse, and book on your behalf. This course covers that layer, product by product.

Fair warning: this is where the study guides' judgment stops being optional. Choosing models, verifying research, and supervising agents are exactly the competencies the exam measures — the tools below are where those muscles get their workout.

1. Lesson 1

   Learn the model picker

   Inside every major product hides a menu most users never touch, and it is the single highest-leverage setting there is. Every provider now ships a ladder: fast, inexpensive models for routine work, and deeper reasoning models — ChatGPT's thinking modes, Claude's extended thinking on its Sonnet and Opus models, Gemini's Pro and Deep Think tiers, Grok's expert modes — that take seconds or minutes longer and actually work the problem instead of answering on reflex.

   The judgment for when each pays is taught product-neutrally in the "Choosing the right model" study guide, and it transfers exactly: reasoning depth and error cost decide the tier. Reformatting notes on a thinking model wastes your time; analyzing an acquisition on a fast model wastes your judgment. The power user's habit is switching deliberately, mid-conversation when needed — and noticing which tier their current task actually deserves.

   Try it this week

   Run one genuinely hard task from your work — a strategy question, a tricky analysis — on both a fast model and a thinking model, then a routine task on both. Note where the gap is dramatic and where it is invisible. That boundary, in your own work, is the lesson.

2. Lesson 2

   Deep research, competitively

   Once deep research matters to your work, stop treating the engines as interchangeable. They have characters: ChatGPT's Deep Research produces the most exhaustive reports and needs a tight brief to stay on target; Gemini's shows you its research plan before running it — edit the plan, that is the power move — and draws on Google's reach; Perplexity's Research mode is the fast, lighter option when you need cited orientation in two minutes rather than a commissioned report.

   The professional pattern for decisions that matter: run the same brief through two engines and read the disagreement. Where the reports diverge is precisely your verification worklist — one of them is wrong, and finding out which teaches you more than either report alone. The brief template from the everyday course still governs: question, decision it feeds, sources to prefer and avoid, format. A deep research agent with a vague brief is an expensive way to get a long document.

   Try it this week

   Take one real decision this month and run an identical brief through two research engines. List every point where they disagree, resolve the top two disagreements against primary sources, and note which engine earned your default.

3. Lesson 3

   Data work without a data team

   Every major assistant will now take your spreadsheet and actually analyze it. ChatGPT's data analysis writes and runs real Python against your file — statistics, charts, cohort breakdowns — and shows you the code, which matters more than it seems: code you can inspect is analysis you can audit. Claude's analysis tool does the same in JavaScript, and Claude also works directly inside Excel; Gemini analyzes from Sheets and Drive, where your data already lives.

   The craft, straight from the "Giving good context" guide: explain your columns and the decision the analysis feeds — "column C is churn reason, free text; I need to know if pricing complaints are growing" outperforms "analyze this" by a mile. And keep the examiner's discipline from "Checking the work": recompute one or two headline figures yourself, and ask what rows were dropped or assumed. The classic silent failure in AI data work is a cleaning step you never knew happened.

   In the toolbox

   • ChatGPT data analysis

     OpenAI · Included; heavier use on Plus

     Upload CSVs or spreadsheets; it writes and executes Python, returns charts and findings, and shows the code. The strongest general-purpose option.

   • Claude analysis tool & Claude in Excel

     Anthropic · Included; Excel integration on paid plans

     Runs code against your data in-chat, and brings Claude directly into the spreadsheet where finance-shaped work actually happens.

   • Gemini in Google Sheets

     Google · Included with Workspace and AI plans

     Analysis where the data lives: formulas, summaries, and charts generated inside Sheets, with Drive files a mention away.

   Try it this week

   Export a month of real data — sales, tickets, hours, anything. Have an assistant answer three questions you have never had time to ask it, then verify one headline number by hand and ask the model what rows it excluded. Keep the transcript; it is your audit trail.

4. Lesson 4

   A second brain that cites itself

   The power user's knowledge setup has two layers. NotebookLM is the deep archive: load an entire domain — every doc from the project, a quarter of board papers, the full technical manual — and get answers grounded only in those sources, cited to the passage. Because it will not stray beyond what you loaded, it is the rare AI surface where "where did that come from?" always has an answer. Its audio and video overviews turn the corpus into briefings you can absorb on a commute.

   Projects (in Claude and ChatGPT) are the working layer: smaller, faster, tied to standing instructions, and conversational. The division of labor that works: NotebookLM for studying and interrogating a fixed body of material; a Project for the living, recurring work that draws on it. Feeding a NotebookLM-refined summary into a Project's reference files is the power move that connects the layers — prepared context, reused forever, exactly as the "Working economically" guide prescribes.

   In the toolbox

   • NotebookLM

     Google · Free; higher limits on AI plans

     The grounded archive: dozens of sources per notebook, passage-level citations, audio overviews. The best answer to "I need to genuinely master this pile of material."

   • Projects

     Anthropic · OpenAI · Free and paid tiers

     The working layer: standing instructions plus reference files plus history, scoped per job. Where your prepared context lives and compounds.

   Try it this week

   Pick the domain you most need to master this quarter. Build a notebook with every source you have, interrogate it for an hour, then distill what you learned into a one-page brief and install that brief in a Project. You have just built the two-layer system.

5. Lesson 5

   Small tools you build by talking

   Somewhere in your work is a spreadsheet doing a job a small app should do — a calculator with your team's actual formula, a tracker, a decision matrix, a practice quiz. Claude's Artifacts and ChatGPT's Canvas will build that as working, interactive software from a description, in minutes, no programming knowledge required. The trick is treating v1 as a draft: use it, notice what is wrong, say so, get v2. Five rounds in, you have a real tool.

   Custom GPTs (ChatGPT) and Gems (Gemini) are the other kind of building: not apps but shareable assistants — your instructions, your reference files, wrapped up so a colleague can use the result without knowing how to prompt. The team onboarding guide that answers questions, the style-checker loaded with your standards. This is the E4 move from the "Asking well" guide made concrete: your best prompt, productized for people who will never write one.

   Try it this week

   Build one working mini-tool in Artifacts or Canvas for a real recurring annoyance, and iterate it at least three rounds. Separately, turn your best prompt into a custom GPT or Gem and hand it to one colleague — their confusion is your revision list.

6. Lesson 6

   Connectors, agents, and where to be careful

   Connect your accounts and the assistant stops being an amnesiac: with Gmail, Calendar, and Drive linked (all three major products do this now), "what did I promise people this week?" becomes answerable. Scheduled tasks — ChatGPT's tasks, Gemini's scheduled actions — run standing requests on a clock: the Monday-morning briefing, the weekly summary of a folder. This layer is safe, reversible, and immediately useful; set it up this week.

   The frontier layer is agents: ChatGPT's agent mode and Atlas browser, Perplexity's Comet, Claude's browser extension — assistants that click, browse, fill forms, and complete multi-step tasks. Treat a new agent like a new intern with your logins: watch it work the first several times, start with read-only and reversible tasks, and never leave it unattended around payments, sending, or deletion. And carry the "Keeping it safe" guide's E4 warning into practice: agents read web pages, and web pages can contain instructions planted for them. An agent that suddenly wants to do something you did not ask for is not being helpful — close the tab and check the source.

   In the toolbox

   • Connectors

     OpenAI · Anthropic · Google · Free and paid tiers

     Link Gmail, Calendar, Drive, and workplace tools so the assistant can search your actual life. The highest value-to-risk ratio on this list.

   • Scheduled tasks

     OpenAI, inside ChatGPT · Google, inside Gemini · Included with paid (and some free) tiers

     Standing requests on a timer — briefings, digests, reminders with research attached. Automation with training wheels, in the good sense.

   • Agent modes & AI browsers

     OpenAI (Atlas) · Perplexity (Comet) · Anthropic (Claude in Chrome) · Mostly paid tiers

     Assistants that act — browse, click, fill, book. Genuinely useful, genuinely new risk surface. Supervised use only until trust is earned per task.

   Try it this week

   Connect calendar and email to one assistant and schedule a Monday-morning briefing for two weeks. Then write your personal agent policy — three tasks you would delegate unattended, three you never would — and the reason for each line. That document is E4 judgment, on paper.

The products above change; the judgment the exam scores does not. The study guides teach that durable layer, and the free check scores you on it in about fifteen minutes.