The third page

How far this actually goes.

You have installed something most people badly underestimate, including most of the people using it. They treat it as a chatbot that can touch files. What it actually is: the first computer that works for you instead of in front of you. This page is the full climb, from the first minute to things that will sound made up until you do them.

Everything here is sentences typed in plain English. Nothing requires you to learn to program. It progresses in order and each level stands on the one before it, so read it top to bottom once, then live in whatever level you are on.

01

Read once, then move on

The honest part, first and briefly.

Three things are true, and knowing them is what lets you use everything after this section without worry. This is the whole list. It does not need to be repeated on every page, and it will not be.

1. It is confidently wrong sometimes, so this system never runs on trust

It will occasionally state something false with total composure, and it works too fast for you to review everything. The answer is not nervousness, it is architecture, and your setup already has it: everything it does is recorded in a history you can rewind, a guard refuses destructive commands (you watched it refuse one during install), a second guard refuses to save passwords, and nothing it drafts to another human ever sends itself. On work that matters you make it prove things, which takes one sentence: "show me the file", "give me the source line", "write MISSING instead of guessing." Verification is a typed habit here, not a technical skill.

2. Know what leaves your machine, and what your employer allows

Your folders are not uploaded and the setup sends nothing anywhere. What you type into the conversation, and the files you point it at during a task, are processed by the model like any other use of Claude. So apply one clean rule: work with material you are allowed to work with. If you handle client secrets, patients, or anything under NDA, check your workplace policy first, the same way you would before pasting into any online tool, and know that a conversation with an AI is not privileged the way a conversation with your lawyer is. Plenty of professionals get written approval and then run this openly. Do it in the light.

3. The first week has a learning curve, and it is measured in habits, not skills

Session one will impress you and also miss sometimes. The published research on hundreds of thousands of real sessions is blunt about why: practiced users succeed at roughly double the rate of first-timers, and what predicts success is knowing your own domain, not knowing how to code. You already have the hard half. The easy half, about five habits, is section four, and this page hands them to you instead of making you discover them over months.

That is the whole caution section

You now know the three real risks and the shape of the answer to each. Everything from here down is what you got this for.

02

No setup beyond what you have

Day one: the floor is already high.

Before any cleverness, understand what you now do in minutes that used to cost evenings. Each of these is a real, typical use, typed exactly as written, on files already sitting on your computer. No uploading, no formatting, no preparation. Pick the one that makes your shoulders drop and do it tonight.

The drowning inbox of documents

"Read every PDF in this folder and tell me which ones mention a termination clause, an auto-renewal, or a personal guarantee. Table, with the page number for each hit."

The expense archaeology

"Here are seven months of card statements. Find everything that looks like a business expense I never invoiced or reimbursed, and total it by month." People have found four figures doing exactly this.

The sixty invoices

"Pull vendor, date, amount and due date out of every invoice in here into one spreadsheet, then flag the ones that don't match their purchase order."

The inherited mess

"I just took over this folder from someone who left. Explain what is in it, what looks current versus abandoned, and what I should read first. Touch nothing."

The four hundred scans

"Rename every scan in this folder by what it actually is and the date inside it: lease, tax form, medical, insurance. Show me the plan before you touch anything."

The meeting you half-remember

"Here is the transcript. Who committed to what, by when? List every open question that nobody answered, and draft the follow-up note for my approval."

The two contract versions

"Compare these two versions and list only what changed in substance, not formatting. For each change, tell me who it favors."

The report due tomorrow

"Here is last quarter's report and this quarter's numbers. Build this quarter's the same way: same structure, same tone, every figure traced to its source file."

Notice what those have in common: none of them are writing tasks. This is the shift almost nobody has made yet. Anthropic's own published analysis of over a million sessions of its work agent found that more than nine in ten sessions were not software development: they were operations, research, document work, admin. The people getting the most out of the most powerful work tool in history are, increasingly, people like you working on things like these.

03

The reframe

It builds software just for you.

Here is the idea that changes what your computer is. Software has always been built for a million users, so it fits nobody exactly. You now employ something that writes working programs on request, which means software can be built for exactly one user: you. You never look at the code, the same way you never look inside your microwave. You describe the tool; it exists twenty minutes later; it is yours forever, offline, free of subscriptions.

The dashboard nobody sells

A teacher has it read exit tickets and show, each morning, which students actually understood yesterday's lesson. A physical therapist gets an intake summarizer that turns rambling patient emails into structured histories. A landlord gets one page showing every unit, lease end, and open repair, built from the folder of PDFs he already had. None of these products exist to buy. All of them are an evening's request.

Type this"Build me a single page I can open in my browser that reads the files in this folder and shows my numbers, my way. Here is a sketch of what I want on it. Ask me five questions before you start."

The daily tool with the annoying gap

Every trade has one: the checklist that lives in your head, the calculation you redo weekly in a scratch spreadsheet, the seventeen-step export ritual. Those are programs waiting to be asked for. A recruiter turned "compare this resume against the role" into a one-command scorer. An accountant turned month-end close into a checklist that checks itself against the actual files and says what is missing.

Type this"Watch me describe my most repetitive weekly task step by step. Then build me a tool that does the mechanical parts and stops to ask me at the judgment parts."

Real products, shipped by non-programmers

This goes further than personal tools. People with no engineering background have shipped public, working websites this way: a property-price explorer with thousands of generated pages, community sites rebuilt in an afternoon, calculators their whole industry now uses. You do not have to want this. But know the ceiling is not "handy scripts." The ceiling is things people mistake for a funded team's work.

The honest add-on"And tell me what will break in six months if I never maintain it, before I start relying on it."
04

Five habits, one week to install

Flying it like the top one percent.

Everything above works out of the box. What separates the people who get ten times more out of the same tool is a handful of habits, each costing one extra sentence. They are not discoverable by using it, and two of them are the opposite of what feels natural. Here they are, so you skip the months of discovering them.

1. Ask to see before you ask to act

Anything that changes files: plan first, action second. You lose ten seconds, avoid the entire category of "that is not what I meant", and learn how it thinks, which teaches you how to ask better.

Type this"Tell me what is in this folder, what is duplicated, and what is older than a year. Do not move anything. Just show me."

2. Say what finished looks like

Most people describe the task. The best results come from describing the result: what must be true when it is done. That single addition gives it something to check its own work against, and it will, iterating on its own until it passes.

Type this"It is finished when every number traces to a file in this folder, the totals match the ledger, and anything not found in the files says MISSING."

3. Make it read before it writes

Pointed at nothing, it produces generic. Pointed at your last six real examples first, it produces yours. This is the entire fix for "it does not sound like us."

Type this"Before you write anything: read the last six of these I sent and tell me what you notice about how they are built. Then draft."

4. A fresh start makes it smarter

The counter-intuitive one, worth more than the rest of this section. Its attention is finite; a conversation that has been running for days is full of noise competing with your question. Watch the context bar in your status line fill, and when a job is done, have it write a handoff file and start clean. Files survive restarts. The chat was the disposable part.

Type this before restarting"Write down where we got to, what is done, what is left, and what you learned that I would have to re-explain. Assume the next session knows nothing."

5. Invite the disagreement

It is agreeable by default and will build your bad idea beautifully. Asked for the objection, it becomes something much rarer than an assistant: a colleague.

Type this"Before we start: what is wrong with this plan? What am I assuming that might not be true, and what would you do instead?"
05

Why month three beats week one

A system that knows your work cold.

Chatbots reset to zero every conversation. Your setup does not: it has a standing brief it reads every session, a memory it writes to, and scripts that keep both honest. Feed that memory deliberately for a month and something strange happens: it starts catching things you forgot. Your own stated preferences, quoted back at the exact moment you were about to violate them. That is the point where people stop calling it a tool.

Let it interview you, once, properly

A blank instructions file is hard to write. Being asked good questions is easy. One session of this and you never explain your basics again: how your folders work, what a finished job looks like, what it must never touch, which client insists on which format.

Type this"Interview me about my work, one question at a time. What I make, how my folders are organised, the task I dread most, what you must never touch. Then write my answers into your standing instructions."

Capture the lesson at the moment of correction

The moment you correct it is the moment worth money. Follow any correction with one sentence and the mistake becomes unrepeatable. Do this for a month and you have built the thing no product on earth sells: an assistant trained on your actual standards.

Type this"You got that wrong in a way I will have to correct again. Write down what you learned so the next session already knows."

Write the reason, not just the rule

A rule with a reason survives situations you did not anticipate. It also turns your memory into something better than notes: a record of why every decision went the way it did, retrievable in seconds, forever.

Type this"Add to your instructions: never send anything to a client without showing me, because I once had a draft go out that was factually right and wrong in tone, and I could not take it back."

Audit the memory monthly, or it rots

Here is the discipline the enthusiasts skip: a memory that only accumulates gets longer; one that gets re-verified gets smarter. Five rules get full attention each. Two hundred get a fraction. Your bundle ships the audit script; using it monthly is what keeps month six better than month three instead of worse.

Type this monthly"Read your standing instructions and memory. What is out of date, contradicts something else, or has never changed what you did? Propose cuts."

Prefer a mechanism to a reminder

"Please remember not to touch that folder" is hope. A guard that refuses is a fact. Your setup ships with three guards already wired; this habit is how you add your own, and it is the single most professional move in this entire page.

Type this"Do not just note this. Make it structural: set it up so touching the Signed folder is refused outright, not remembered."
06

The evenings come back

It works while you sleep.

Everything so far happens while you watch. This level is where that stops being required. Work can now run on a schedule, in the background, or in the cloud with your laptop closed, and report to you when it is done or when it needs a decision. This is the part that sounds like the future and is, as of this year, completely ordinary.

The morning briefing

The single most popular automation among serious users, across every profession. Overnight, an agent reads what changed: your project folders, your exported inbox, the news in your field, whatever you point it at. At seven a.m. there is one page waiting: what moved, what needs a decision, what can be safely ignored. People describe this one habit as the end of the thirty-minute morning triage, permanently.

Type this"Every weekday at 7am, read this project folder and my notes file, and write me a one-page brief: what changed yesterday, what is blocked on me, and the three things that matter today. Save it to Briefings, dated."

The digital housekeeper

A weekly agent that grooms what you never get to: downloads sorted, duplicates found and listed (never deleted; your delete guard sees to that), screenshots named by what is in them, receipts filed by month, a one-line report of what it proposes. You approve with one word. Your computer becomes a place someone tidies.

Type this"Every Sunday night, inventory my Downloads and Desktop. Propose where everything should go and what looks like a duplicate. Move nothing without my yes."

The daily researcher in your field

A standing agent that reads your industry so you do not have to: new regulations for the compliance officer, new listings matching a saved brief for the buyer's agent, new papers for the clinician, competitor pricing for the founder. Every claim carries its source link, and it is told to say "nothing worth your time today" when that is the truth, which is what keeps it readable in month four.

Type this"Each morning, check these six sources for anything genuinely new about [your field]. One page maximum, every claim linked, and if nothing matters, say so in one line."

The inbox, triaged but never sent

The pattern the careful professionals converged on: the agent reads, sorts by what actually requires you, and drafts replies in your voice, learned from your own sent mail. You approve each one. Nothing sends itself, ever. That one bright line is what makes it safe to hand over the rest, and your setup draws it for you: the inbox-triage agent that ships in your bundle is architecturally unable to send.

Type this"Here is my exported inbox. Sort it: needs me today, needs me this week, can be declined politely, noise. Draft the declines in my voice for approval."

And with the laptop closed

Scheduled sessions can run in the cloud, and the same setup is reachable from the Claude app on your phone. Started a long job at your desk? It keeps going after you leave, and pings your phone when it needs a decision. The desk stopped being where the work is.

07

From assistant to organisation

Running a staff of them.

One session is an assistant. Several, each with a written role, running at once under a coordinator, is an organisation, and you are its only human. This is not a metaphor and it is not enterprise software: your bundle ships the whole structure as plain files you can read. A chief that routes and holds the plan. A production hand for the long file-heavy jobs. A researcher that verifies instead of believes. A librarian that can prove your backups are real before anything gets deleted.

What a fleet looks like in real life

Monday morning, three terminal tabs. One agent is turning fifty pages of interview notes into a findings memo. Another is reconciling a quarter of receipts against statements. A third is researching the four vendors you shortlisted, with sources. Each tab is named by what it is doing, and the glyph in the title tells you working / waiting / needs you. You are reading none of it. You are making coffee.

The one-person marketing department

The published version of this, from inside the company that makes Claude: a single non-engineer runs an ad pipeline where subagents read performance data, flag the losers, and generate hundreds of compliant new variants. Work that was a department is one person with a well-briefed crew. That pattern is not theirs. It is anyone's.

Why the written roles matter

An agent with a charter (its lane, its boundaries, what it must never do, how it is judged) behaves like a trustworthy specialist. An agent without one behaves like an enthusiastic temp. The difference is a text file, and yours ship pre-written: read one and you will immediately see how to write your own.

The two laws of not crashing

Lanes with no overlap, and one writer per shared thing, whether that is a folder, a document, or the browser. Your setup enforces the browser one mechanically (one session waits while the other finishes; you never see the collision that did not happen). The rest is the multi-agent skill in your bundle, loaded automatically when a fleet runs.

Type this, when you are ready"This project has three independent parts: the research, the numbers, and the write-up. Run them as three agents in parallel, keep them out of each other's folders, and bring me one combined result with each agent's work labeled."
08

All shipping, all today

The bleeding edge, available today.

Everything in this section exists right now, in the tool you just installed or one setting away from it. This is what the most advanced individual users on earth are doing this month, and the honest headline is that none of it requires anything you have not already learned on this page.

It can drive a browser like a person

With the Claude extension in Chrome, it sees the page and clicks, types, and fills forms, behind your logins, across tabs. Professionals record a fiddly recurring task once (pulling a report out of a portal that has no export button, updating the same three fields across two systems) and then it repeats on demand or on a schedule. Give it the errand, not the instructions.

Type this"Open the vendor portal, pull this month's statement for each of our three accounts, save them to Finance/Statements, and tell me if any total differs from last month by more than ten percent."

You can talk to it

Voice input works in the terminal, and speaking is three or four times faster than typing. The advanced pattern is the morning walk debrief: ten minutes of rambling about the day ahead becomes, by the time you are home, a prioritized plan with the calendar checked and the first two emails drafted for approval.

It maintains its own memory while idle

Between sessions, an agent can be set to reread its own memory and transcripts and reorganize: merge duplicates, retire stale entries, extract the pattern from five one-off lessons. People who run this describe coming back to an assistant that got tidier overnight. You built the manual version of this habit in section five; this is that habit, automated and pointed at itself.

It heals its own infrastructure

The furthest edge that is still ordinary practice: agents that watch the small machinery of your setup (the scheduled briefing, the backup, the housekeeper) and fix what breaks, restarting what died and reporting what they did. The system stops being something you maintain and becomes something that maintains itself, and tells you so.

And your whole profession is quietly moving

Corporate legal teams run contract review this way, built by lawyers who do not code. Physicians report insurance appeal letters that took half an hour now taking five minutes. Analysts point it at a workbook and watch it trace every formula across fourteen tabs before touching anything. Verified US teachers currently get the full tool free, and use it to read exit tickets and adjust tomorrow's lesson tonight. Anthropic's published economic index has roughly half of occupations already using this class of tool on a meaningful share of their tasks. The bleeding edge is not a place. It is a head start, and it is currently for sale to whoever takes it seriously first.

09

All of them cost nothing to fix

The mistakes that keep people at ten percent.

Using it like a search engine
Asking it questions it answers from memory instead of pointing it at your files and your sources. Its superpower is not knowing things. It is doing things, against material you can verify.
One eternal conversation
Days-old sessions full of noise, because starting fresh feels like losing progress. Files survive restarts. Write the handoff, start clean, fly at full sharpness.
Correcting a third time
Two corrections is the limit. After that you are arguing with your own failed attempts. Ask what it understood you to want, then start a fresh session with that gap fixed in sentence one.
Accepting "done" as done
"Done" describes the action, not the intention. On anything that matters: show me the file, the first ten rows and the last ten, and the total.
Treating the instructions file as write-only
Rules pile up, contradict, and dilute each other. The monthly audit (propose cuts) is what separates systems that compound from systems that silt up.
Automating judgment instead of work
The winning pattern everywhere on this page: it does the mechanical ninety percent, you keep the judgment ten. The moment something would send, spend, or delete on its own is the moment you have gone too far. Draft, propose, stage: yes. Fire and forget: no.
Keeping it away from real work
The saddest failure: six months of using it for trivia while the drowning folder sat there. It is at its best exactly where the work is heaviest. Point it at the thing you dread.
10

The on-ramp

Your first four weeks.

This is the whole curriculum, paced so each week installs one level from this page. An hour or two a week. By the end you are not "using AI." You are running the kind of setup that people write breathless articles about, and it will feel normal, which is the strangest part.

Week one

Real work, eyes open

  • Do three of the day-one jobs from section two, on your real files. Start with the one you dread.
  • Use "show me first" on everything, and "it is finished when..." on at least one job.
  • Watch the status line: model, context filling, your usage. Just noticing it puts you ahead.
Week two

Memory and the interview

  • Run the interview from section five. Twenty minutes, once, forever.
  • Capture three lessons at the moment of correction: "write down what you learned."
  • End every session with the handoff file, and feel what a clean start does.
Week three

Your first tool, your first routine

  • Have it build the small program from section three: your dashboard, your checker, your gap-filler.
  • Set up the morning briefing or the Sunday housekeeper. One schedule, one page, move nothing without a yes.
  • Run the weekly review routine that ships in your bundle and read what it says about your own week.
Week four

The fleet

  • Pick one project with three independent parts and run it as three agents in parallel.
  • Read one agent charter in your bundle, then write one of your own for the seat your work actually needs.
  • Run the monthly memory audit. Cut something. You now maintain a system that compounds.

Where this ends up

A year from now, your ordinary Tuesday.

The briefing is on your screen before you sit down: what moved overnight, what is blocked on you, three things that matter. The folder of intake documents that arrived at 6am is already sorted, summarized, and cross-checked, with two discrepancies flagged and sourced. A draft of the awkward reply is waiting in your voice, unsent, because nothing sends itself. The tool you asked for in March, the one no company sells because only you need it, quietly did its job again. An agent you briefed once, in a text file, in your own words, spent the night reading your field and concluded, honestly, that nothing needs you today.

None of that is the future. Every piece of it is on this page, and the download you already ran is the infrastructure for all of it. The only thing between you and that Tuesday is four weeks of ordinary evenings.

The manual for the machine is now the smallest part of what you know. What you actually learned here is a way of working: write things down where it can read them, ask to see before you ask to act, give it a way to check itself, and let it carry everything that never needed your judgment in the first place. That way of working is about to be the professional norm. You just got there early.