How we work
The tools we use, how we think about them, and what the words mean. Written down in case it saves you the time it cost us.
Start at the top. Everything under the tools is optional.
One thing up front
We reserve the right to get smarter tomorrow
Nobody is an expert at this yet, us included. Some of what is on this page will be wrong by the time you read it. We keep finding out we have been doing something badly for a month straight.
So this is what is working for us right now, not advice from anybody who has it figured out. If something here looks wrong to you, it might be. Tell us and we will change it.
Part one
Tools
Granola
Takes the notes in every meeting we are in so we can actually be in the meeting. The summaries are good enough to send to clients.
Wispr Flow
Voice dictation that keeps up with how people actually talk. Most of our emails and messages are spoken first now.
Mercury
Business banking that feels like software instead of a bank lobby. Every company we run banks here.
Raycast
The command bar for your Mac. Launcher, clipboard history, window management, and quick AI, all from the keyboard.
Grok Bot
Newest thing on this list and the one we are most excited about. We are early with it, so read this as enthusiasm rather than a verdict, but it has already changed how fast a question turns into an answer around here. It runs on a Cursor account, and the five-step path is further down this page.
Agents
Claude Code
Agentic coding tool that lives in your terminal. Describe what you want, it reads the codebase, writes the code, runs it, and iterates. Does most of the heavy lifting around here.
Codex
OpenAI's coding agent, same job as Claude Code. Running both means work happens in parallel and you get a second model's take on tricky problems.
Claude Design
A design agent with a visual canvas. You iterate on layout and look through chat instead of hand-placing everything, then pull the result into the build.
Cursor
The editor built around an agent. When we want to see the code while it changes, this is where we go.
What we build on
Next.js
The framework. React for the component model, Next.js for routing, server rendering, and the build pipeline.
shadcn/ui
UI components built on Radix and Tailwind that you copy into your repo, so you own the code and can change anything.
Vercel
Hosting and deploys. Push to GitHub and it is live in a minute, with a preview URL for every branch.
Worth reading
The Only Question Left
The cost to think and the cost to build are both heading to zero, which leaves one question: what is worth building. Not ours, but it makes the case better than we would have.
Part two
How to think about it
It amplifies, it does not replace
Technology has always been an amplifier. Diligence gets amplified, and so does laziness. If you hand it a shrug, you get slop back, and the whole internet is currently full of the evidence. The first question every time is what you are actually amplifying.
Agency is something you hand out on purpose
Give an intern the CFO's job and that is on you. Give the CFO the intern's job and that is also on you. Same call here: how much you let something take on should match how much you have taught it. Most horror stories are somebody handing their whole life to an assistant on day one.
Create first, then let it sharpen
For anything creative, you go first and it edits second. For anything administrative, flip it: let it take the first pass and you edit. Getting that backwards is how work stops sounding like the person who signed it.
Judgment is the part that got scarce
The Great Pyramid took roughly 100,000 workers. The Empire State Building took 3,400. One person with an agent can now ship what used to take a team of 50. When capability gets cheap, the premium moves to knowing what is worth building and when the answer in front of you is wrong.
Give it a job, not your life
The people getting real value gave their assistant one assignment and a role: schedule and inbox, or research and writing. Two narrow assistants beat one that is supposed to do everything, because you can tell when one of them is off.
Draft, do not send
A useful standing rule while you are learning what something is good at: it can write the reply, organize the inbox, and prepare the agenda, and a person presses send. Where it is wrong, you find out at no cost.
Own the context, not the product
The model you prefer will change, probably twice this year. What you write down about how you work, how you talk, and what you have already made is portable, and it moves to whatever comes next. Building that on one company's product is building on somebody else's land.
You can be blunt with it
Telling it the last thing it made was garbage and to never do that again is fair game, it works, and you cannot talk to a person that way. Asking it to roast you is also instructive, and usually accurate.
AI Vocab
Model
Does the thinkingThe thing that does the thinking. Trained once, then frozen. It has no memory of you between conversations and cannot do anything on its own. GPT, Claude and Gemini are models.
When people say AI got better, they mean a new model shipped. Past the top tier, which one you pick matters less than what you hand it.
Context
What it can see right nowEverything the model can see at the moment it answers. Your question, the files you attached, what was said earlier, anything it looked up. Context is a desk, not a memory. It gets cleared.
Biggest lever you have on the answer, and the one nearly everybody skips. A bad answer is usually a bad desk.
Tools
What it can reach and changeThings the model can use besides talking. Search the web, open a file, run a calculation, send an email, look something up in your system.
This is the line between something that answers and something that works. Whatever you connect, it can reach, so connect deliberately.
Harness
What keeps it working instead of only answeringEverything built around the model to turn answering into working. The loop that lets it try, check its own work and go again. The tools it is allowed to use. The rules it follows. What it remembers between runs. Model is the engine, harness is the rest of the car.
Same model, different harness, completely different results. When people argue about which tool is better, they are almost always arguing about the harness. Claude Code, Cowork and Codex are harnesses, and more than one of them runs on the same brain underneath.
Agent
All four of those, pointed at a goalA model, in a harness, with tools, allowed to take steps on its own. It decides what to do next, does it, looks at what happened, and keeps going until it is finished or stuck.
You ask a chatbot a question, you give an agent a job. An agent is only as safe as what you let it touch.
Every product you have heard of is some arrangement of those five. A chat app is a model and a desk. An agent adds the tools and the harness. Anything built for one company is all of it, aimed at a single job.
Four of the five are yours to keep. Which model you sit on top of will change, and one day it will be as dull a question as who your internet provider is.
Words you will also hear
Context window
How much fits on the desk at once. Your files, the conversation, its own notes. Run out and the older material falls away.
Tokens
The chunks text gets cut into, roughly three quarters of a word each. Context windows are measured in them and billing is counted in them, separately for what goes in and what comes out.
Knowledge cutoff
The date the training data stops. Ask about anything after it and the model is guessing unless it can search or read files. It is why a good agent looks things up instead of answering from memory.
Reasoning
The model working through a problem before it answers, sometimes for a while. It costs tokens and time, and on hard problems it moves quality more than anything else you can change.
MCP
Model Context Protocol. The standard plug that lets an agent talk to your other software: Slack, Notion, a database, a design tool. It is how an agent acts inside apps it was never specifically built for.
Frontier model
Whatever a lab currently ships as its most capable model. A moving label rather than a product name, and usually the most expensive per token.
Worth knowing early
“Train it on our data” means briefing it
Your documents go into context at the moment you ask, and nothing about the model itself changes. Actually altering a model, which is called fine-tuning, is rare, costly, and almost never what a business needs. This is the mix-up we correct more than any other.
What being wrong costs
A chatbot that is wrong costs you the minute it takes to read the answer. An agent that is wrong costs you whatever it touched. Match how far you let something reach to how much you would mind it being wrong there.
What it is reliable at, and what it is not
Reliable: drafting, summarizing, rewriting, pulling structure out of messy text, explaining something unfamiliar, and any first pass you were going to review anyway. Unreliable: arithmetic worked out in its head, facts nobody put in front of it, and any job where being confidently wrong is expensive. Confident invention is called hallucination, and it does not know it is doing it.
Why the same question gives two answers
There is deliberate randomness in how each next word gets picked, and the context is never quite identical twice. Repeatability is not what these are built for. When something has to come out the same way every time, write the rule down and have it follow the rule rather than remember it.
What a filling context window feels like
Replies get vaguer, a constraint you set at the top quietly gets dropped, and it asks you something you already answered. Nothing announces it. Start a fresh session and paste back the state that matters, or keep that state in a file it re-reads at the start of every run.
A bigger model is usually the wrong first move
Most disappointing answers come from a thin desk, not a weak brain. Before paying for the top tier, hand it the actual file and an example of what good looks like. We have skipped that step and bought the upgrade instead. It did not help.
Asking for the source costs nothing
When an answer matters, ask where it came from and open it. Something that searched will hand you links you can check; something answering from training will not, and the difference tells you which one you got.
What people bring up, and where we land
Where does what I paste actually go
Most common question by a distance, and a fair one. Whether your conversations train future models depends on your plan and a setting most people never open, and the default is not always the one you would pick. Check it rather than assume it. For regulated material, strip names and account numbers first, and remember that whatever an agent can reach, it can also send. How we set client work up is in the data section below.
The pull to always be producing
Being able to deploy a night of work while you sleep is genuinely useful and genuinely hard to put down. Several of us have caught ourselves getting one more job queued at midnight. Worth naming before it becomes a habit.
Knowing what is real
Generated images are past the point where you can tell. Picturing a deck on your own house is fine. Anything that goes to a customer or a congregation deserves a harder line, and it is worth deciding where yours is now rather than after.
Using it where a person should be
It will happily draft the pastoral reply, the condolence note, and the hard conversation. That it can does not mean it should. The line is yours to draw, and drawing it out loud with your team is the useful part.
Who makes what, and what it costs
Anthropic
Makes Claude. This is what we build on. Prices are per million tokens, input cost then output cost.
Claude Fable 5
most capableBuilt for agents that run a long time without losing the thread. The tier we reach for when a job is genuinely hard.
Claude Opus 5
our defaultWe default to whatever Opus is current, and right now that is this one: close to Fable's intelligence at half the price. Complex agentic coding and enterprise work.
Claude Sonnet 5
The balance of speed and intelligence for everyday work. What launched as introductory pricing is now the standard price.
Claude Haiku 4.5
fastestNear-frontier answers at the lowest cost, for volume and quick turns.
OpenAI
Makes ChatGPT and Codex. Shipped as three tiers rather than one model.
GPT-5.6 Sol
flagshipThe frontier tier, and what powers the reasoning modes on paid ChatGPT plans. When we go to OpenAI we go to whatever the top model is, which today is this.
GPT-5.6 Terra
The balanced everyday model, at well under half of Sol's price.
GPT-5.6 Luna
cheapestThe quickest of the three, and now cheap enough to run on volume work without thinking about it.
Where the tools sit
Each of these is a harness, and each can run on more than one brain underneath. Day to day we are in Claude Code, Codex, and Conductor, and the rest is a visit.
Claude Code
AnthropicThe harness for software. It reads a codebase, writes and runs code, and fixes what breaks. Lives in the terminal, the desktop app, the web, and inside VS Code and JetBrains.
Claude Cowork
AnthropicThe same approach pointed at everything that is not code: research, analysis, documents, multi-step admin. You give it a folder and plain instructions. Desktop, web, and mobile.
Codex
OpenAIOpenAI's coding agent. Searches repositories, runs terminal commands, and debugs. Runs on the GPT-5.6 tiers.
Conductor
Melty LabsA Mac app that runs several Claude Code and Codex agents at once, each in its own copy of the project so they cannot collide. The app is free and runs on the plans you already pay for. Mac only for now.
What just changed
The last two months, so the pace is visible.
Cowork moved to the cloud and onto web and mobile, so a task keeps running after you close the laptop. Anthropic's own usage data showed most Cowork users were not coding.
Chat and Cowork merged into one Claude. On web and desktop they share a single home screen, and projects and files carry across both. The wall between talking to it and handing it work is gone.
OpenAI shipped the GPT-5.6 family as three tiers instead of one model, and it landed in ChatGPT, Codex, and GitHub Copilot the same day.
Anthropic shipped Claude Opus 5 at the same price as Opus 4.8, close to Fable's intelligence for half the cost. It became the default on Max and the strongest model on Pro, and it is what we moved our own work to.
OpenAI cut Luna by 80 percent and Terra by 20 percent. The cheap tier is now cheap enough that cost stops being the reason to avoid running something on every record you have.
Anthropic made Claude Sonnet 5's introductory price permanent. The increase to $3 / $15 that was set for September 1 will not happen, so the everyday tier stays at $2 / $10.
Both labs ship about once a month. Everything above was accurate on August 31, 2026. If something looks off, it moved.
Part three
Where to start
Quick path
Get started with Grok Bot
Grok Bot is the always-on agent from xAI. It runs on a Cursor account, and the cheapest way in is not the obvious one, so here is the path we would take.
- 1
Make a Cursor account in a browser
Go to cursor.com and sign up. Grok Bot has no login of its own. It signs in with a Cursor account, so this is the account you will use everywhere.
- 2
Buy Pro on the web, not on your phone
Still in the browser, upgrade to Pro. It is the lowest plan that includes Grok Bot. It is better not to buy through the native iOS app.
Pro+ and Ultra add weekly usage, not features. Start on Pro and move up if you run out.
- 3
Download Grok Bot on your Mac
Go to x.ai/bot and choose Download for macOS. Pick Apple Silicon or Intel to match your Mac: open the Apple menu, choose About This Mac, and a Chip line means Apple Silicon while a Processor line means Intel. Open the disk image and drag Grok Bot into Applications.
- 4
Sign in with Cursor
Open the app, choose Get started, then Sign In with Cursor. It finishes in your browser with the login from step one. Come back to the app once the browser confirms.
- 5
Add the iPhone app afterwards if you want it
Install Grok Bot from the App Store and sign in with the same Cursor login. Mac and iPhone share one account and one usage allowance, so there is nothing to buy inside the phone app.
Checked Sept. 4, 2026. Share this section as ai-delivered.com/grok.
You want the tools, not a project
Put a note-taker in your meetings
This is the single biggest unlock for most people and it takes about a minute to set up. It listens, and afterwards you get what was said, what you committed to, and what you asked of other people. The real gain is that you stop taking notes and start listening, and people notice.
Photograph the paper and let it do the typing
The season schedule that came home in a backpack, the handout, the whiteboard. Take a picture, ask for calendar events, put your spouse on the invites. This is the one that converts skeptics.
Let it organize the inbox before it writes anything
Sorting and prioritizing is lower risk than replying, and it is most of the time you lose. Once that is working, ask it to study how you already write, then have it draft and leave the sending to you.
Ask it to show the work, not hand you the answer
Homework at the kitchen table is the clearest example. Do not give me the answer, show me how to solve it. The same request improves almost every other answer you get.
Give it your real thinking first
The difference between a useful answer and a generic one is almost always what you put in front of it. Your draft, your notes, your actual constraints. A thin brief gets you something anybody could have gotten.
You want an assistant of your own
Collect what you have already made
Talks, posts, proposals, past emails, anything that sounds like you. Put it in one folder. That folder is the thing worth having, and it is worth more than which model you use.
Have it read the folder and find the threads
Ask it to pull out the themes, the phrases you repeat, and how your tone changes by audience. What comes back is usually more accurate than what you would have written about yourself.
Split it into two assistants and name them
One handles schedule and inbox. One handles research and writing. Naming them is not a gimmick: it keeps you clear about which one you are talking to and what each is allowed to touch.
Write the rules down where it will re-read them
What it never does, how you like your week laid out, which day is protected. Rules in a file get followed. Rules said once in a conversation get forgotten when the conversation ends.
Ask for the morning email
Yesterday's meetings, what you committed to, what you are waiting on from other people. This is the first thing most people build that they keep, and it is the point where it stops being a novelty.
Part four
Where data lives
The apps are not wired to a model provider
The databases and tools we build don't call out to an AI model on their own. A model only sees client material if one of us deliberately brings it into a session, which means our own habits matter as much as the architecture.
The training settings get checked
Most AI tools have a setting controlling whether your conversations get used to train models, and the defaults are not always the ones you want. We keep a written record of where each tool we use stands, and we re-check it when providers change terms.
Databases and hosting
Postgres on Supabase, hosting and TLS on Vercel, both in US regions. Row-level security on, admin-only access policies, and service keys in environment configuration rather than sitting in the repo.
Private pages stay private
Anything sensitive sits behind a server-side gate before it goes out: encrypted passwords, rate-limited checks, signed cookies, share links on unguessable tokens, and no search indexing.
Regulated data stays put where we can manage it
Where PII or PHI is in play we would rather read from whatever system already holds it than copy protected records into a new one. Fewer copies tends to beat more encryption on more copies.
Sensitive documents can stay on our own hardware
For material that shouldn't touch a cloud model at all, we have run processing on a local model on a machine we control. It's slower and it's more setup, and sometimes that is the right trade.
You own it
The database and the code belong to the client. If an engagement ends, it stays with them and we step out of the way.
If you have a compliance requirement, an IT reviewer, or an attorney who needs to look at this, we would rather hear about it at the start than halfway through a build.
Part five
Starting to code
Habits
Live in the agent, not a browser
We open a browser to preview something and not much else. The rest happens in the conversation.
Point it at a repo, not a chat project
"Open my [name] repo with all the context" saves re-explaining yourself at the start of every session.
Give it house rules
A CLAUDE.md file holds how you write and what you never want it to do. Saying "add this to memory" works too, and it sticks.
Keep one repo as the source of truth
Ask it to interview you, something like "ask me 100 questions about me and my work," and it will organize what it learns. That becomes the thing you point everything else at.
Start from something that exists
Reusing a past build beats starting cold, and it saves teaching the same preferences twice.
Read before you accept
It is easy to rubber-stamp the seventh prompt in a row. We have done it and regretted it. A quick skim of what it is about to do is usually enough.
Accounts, a terminal, and the setup commands
We work on Macs, so that is the path we can vouch for line by line. Nearly all of this works the same on Windows. The accounts, the agent, and the hosting are identical, and only the terminal and the installer change. The one real gap is Conductor, which is Mac only for now; everything else here runs either way. Assume a Claude Pro or Max plan. Copy each command with the button, paste it in, press Enter.
Three accounts to make
GitHub
Where the files live. Version history in the cloud, free, and each project is a repo. Most of the internet runs on it.
Vercel
Hosting. It takes whatever is in GitHub and pushes it live. Push a change and the site updates in about a minute, and every branch gets its own preview link to share.
Supabase
laterThe database, for when a static page is not enough: dashboards, a library of something, a task manager. Worth skipping until you need it. One sentence to Claude spins it up.
In order
- 1
Make a GitHub account
Sign up at github.com with an email you plan to keep. That is enough for now. Once Claude Code is running you can tell it "make all my repos private" and it handles the rest.
- 2
Make a Vercel account with GitHub
At vercel.com, choose "Continue with GitHub." Install the GitHub app when it asks and give it access to all repos. Vercel can now see and deploy anything in GitHub.
- 3
Open Terminal and install Homebrew
Terminal is the app where you type commands. Find it with Spotlight: press Cmd and Space, type "Terminal," press Enter. Paste this in and press Enter. It asks for your Mac password once, and nothing appears on screen while you type it. That is expected.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"Homebrew is the tool that installs other Mac tools. We reach for it again later for things like ffmpeg, which is what lets Claude edit audio and video.
- 4
Install Claude Code
Still in Terminal, paste this and press Enter. When it finishes, type claude and press Enter to start it. Typing claude --version prints a version number, which confirms it landed.
brew install --cask claude-codeClaude Code runs on a Claude Pro or Max plan, the same login you already have. If you would rather stay out of the terminal, the Claude desktop app does the same work in a window.
- 5
Point it at a folder and talk
Make a folder anywhere (Documents is fine), open Terminal in it, and run claude. From there you describe what you want. It reads the folder, writes the code, runs it, and fixes what breaks.
- 6
Go live when you are ready
In Vercel, open the project's Domains tab and add your domain. Point the main site at something like yoursite.com. Later you can add a private dashboard at a subdomain like admin.yoursite.com to run projects from, which is where Supabase comes in. Nothing is public until you point the domain.
When you want more
Skills. Packaged workflows for jobs you repeat.
Loop. Have it wake up on an interval and keep improving something until it is done.
Artifacts. Turn words into a clean page, like a proposal with signatures built in.
Usage. Type /usage to see what you are actually spending.