# The Climb: 5 Steps to Running Your Business on AI


---

> **Rather skip the climb and have it built?** Our team builds done-for-you AI second brains: the brain, the hands, the bots and the schedules, set up around your business. **[Book a call with our team](/book)**

![The Climb: 5 steps to running your business on AI](/tools/climb/images/climb-poster-the-climb.png)

![Most people start at the top. Start at the bottom instead.](/tools/climb/images/climb-poster-start-here.png)

> **My AI isn't smarter than yours. It just has way more context.**
> Same models, same apps, same price as everyone else. The difference is that mine already knows every meeting, every person, every project and how we do things before I type a word. You can't buy that. You have to build it, in order.

---

## The One Rule

**The bots come last.**

Everybody wants to start at autonomy: the agents, the automations, the stuff that runs while you sleep. So they buy the bots first, drop them into a messy business, and the bots make things up, sound nothing like them, and get switched off in a week.

The bot was never the problem. It didn't know anything, and the business it landed in was a mess. So you climb in order. Fix the business. Give the AI a brain. Give it hands. Then hand it whole jobs. Then let it run. Then let it compound.

**The one number to watch:** how much of your week happens without you pressing go. Every step below moves that number. Nothing else really matters.

> The order matters more than the tool. You can run every step of this on Claude Code, Codex, Grok Bot, whatever. Tools change every month. The order doesn't.

---

## The Six Levels

| Level | Who you are | What the AI does | Your week looks like |
|---|---|---|---|
| **Base Camp** | The Juggler | Nothing yet (on purpose) | You're in every task. Stuff lives in six apps and your head. Fridays are catch-up. |
| **01 · Brain** | The Librarian | Answers from *your* notes | You still do the work, but you never re-explain your business. "What did we decide?" takes ten seconds. |
| **02 · Hands** | The Operator | Does tasks in your real tools | You review drafts instead of writing them. Email, calendar and Slack work lands ready to send. |
| **03 · Delegation** | The Dispatcher | Owns whole jobs, one bot per lane | You assign and approve. You message one bot and the work goes to the right specialist. |
| **04 · Autonomy** | The Architect | Runs on schedules and triggers | You read the log, not the inbox. Things happen at 11:30 PM without you. |
| **05 · Compound** | The Owner | Runs the loop toward a goal | You set the goal and approve output. Output keeps going up; your hours don't. |

## One ladder, four lenses

Same climb whether it's just you or a whole company. What each step looks like:

| Level | Solo | Small team | Agency | Bigger company |
|---|---|---|---|---|
| **Base Camp** | List your week, delete what nobody would miss | Every role writes down what they actually do | Map delivery per client type | Process inventory per department |
| **01 · Brain** | One vault: meetings, people, projects | One shared vault, everyone's calls land in it | A page per client, every call filed | A brain per department, one shared index |
| **02 · Hands** | AI connected to your email + calendar | Each person gets an agent in their tools | AI drafts client updates and reports | Agents on narrow logins per system |
| **03 · Delegation** | Bots for content, admin, life | Bots per function (sales, support, ops) | Bots per stage of delivery | Bots per workflow, with owners |
| **04 · Autonomy** | Nightly cleanup, a comms watcher | Daily standup summary, alerts | Weekly client reports on a schedule | Triggered workflows with audit logs |
| **05 · Compound** | Content machine | Lead machine | Delivery machine | Pick one loop per department |

---

## Base Camp · Fix the business before AI touches it

![Base Camp](/tools/climb/images/climb-0-base-camp.png)
*Caption: Fewer, cleaner processes in one place. AI multiplies whatever you give it, so give it something clean.*

**What this is.** The step that has nothing to do with AI. You write down every process the business actually runs, kill what you don't need, and simplify what's left. Before any AI touches it.

**Definitions.**
- **Process:** anything that repeats. Following up a lead, sending an invoice, posting a reel, paying a bill.
- **Source of truth:** the one place a thing lives. If it lives in two places, it lives in none.

**What tool to use.** A spreadsheet or a single note. Seriously. No AI yet.

**What I recommend.** One table, filled in honestly. Not what's *supposed* to happen. What actually happens.

**How to structure it.**
1. **Audit.** List every recurring process: who does it, how often, how long it takes, which tools it touches.
2. **Document.** Write your top 5 as plain steps: "When a lead comes in, we…"
3. **Delete.** Kill the steps, tools, subscriptions and reports nobody would miss. Delete before you simplify. Simplify before you automate.
4. **Simplify.** One place for files, one for leads, one for tasks. Pick the home your brain will live in (Step 1).

**Copy this worksheet:**

| Process | Who does it | How often | Time each | Tools it touches | Keep / Delete / Simplify | AI candidate? |
|---|---|---|---|---|---|---|
| e.g. new lead follow-up | me | daily | 20 min | email, CRM | Simplify | yes, low risk |
| | | | | | | |
| | | | | | | |

**How to read the diagram.** Left to right. Everything starts as a mess. The DELETE gate is the one people skip, and it's the one that saves the most time.

**Graduate when** the business is clean enough that AI would multiply it, not its mess.

**Proof you're here.** One page lists every recurring task and tool, and you can name the three most repetitive, lowest-risk jobs.

> **Remember:** automating a bad process just gets you a bad process faster.

---

## Step 1 · The Brain: everything in one place

![The Brain](/tools/climb/images/climb-1-brain.png)
*Caption: One folder the AI reads first. Meetings link to people, people link to projects.*

**What this is.** Everything about the business in one place: every meeting, every person, every project, and how you do things. Once that exists, any AI you plug in already knows how you work.

**Definitions.**
- **Vault:** a folder of plain notes. Obsidian calls it a vault. It's just files on your computer.
- **Context:** what the AI knows before you ask. More context beats a better prompt every time.
- **Ways of working:** your SOPs, how you talk to customers, pricing rules (what an AI can and can't quote), things you never say, and past decisions with the reason.

**What tool to use.** Obsidian (free, plain files, any AI can read it). Notion or Google Drive work too. Messy is fine. Missing is not.

**What I recommend.** Obsidian. I run my whole business out of one vault, and Grok Bot, Claude Code and Codex all read the same folder. If I switch AIs tomorrow, the brain comes with me.

**How to structure it.**
1. **Make the vault** with three main folders: Meetings, People, Projects.
2. **Meetings:** every call lands here with a summary, decisions and to-dos.
3. **People:** a page on everyone you work with, so the AI knows who's who before you ask. **Projects:** everything you're building and where it's at.
4. **Add one "How we work" page,** then point your AI at the folder and ask a business-only question: *"What did we decide with [person] and what's still open?"*

```
vault/
  Meetings/     <- every call: summary, decisions, to-dos
  People/       <- one page per person
  Projects/     <- one page per project + where it's at
  How We Work/  <- voice, SOPs, pricing rules, past decisions + why
  AGENTS.md     <- the rules every AI reads first
```

**How to read the diagram.** The vault sits in the middle because every later step reads from it. The links between notes are what let the AI go meeting → person → project in one question.

**Graduate when** the AI answers from *your* notes, not generic advice.

**Proof you're here.** That question gets an answer with the note it came from, and a new hire could learn the business from the brain.

> **Tip:** the full vault setup, folders, filing prompt and ask prompt, is in my VAULT guide (link at the bottom).

---

## Step 2 · Hands: stop chatting, start doing

![Hands](/tools/climb/images/climb-2-hands.png)
*Caption: Same brain, now connected to the tools you already use. Everything lands as a draft first.*

**What this is.** Connecting the AI to the tools you already use (email, calendar, Slack) so it actually does the work instead of just chatting about it.

**Definitions.**
- **Hands:** a connection that lets the AI read from and write to a real tool.
- **Draft checkpoint:** the AI prepares the email, the invite, the message, and it waits until you say go.
- **Narrow access:** the AI gets its own login with the fewest permissions it needs. Never your master account.

**What tool to use.** An agent that can actually run things: Claude Code, Codex, Grok Bot or similar, working inside the brain folder, with connectors to your tools.

**What I recommend.** Start with email + calendar. They're where most of your week goes, and drafts are easy to review.

**How to structure it.**
1. **Connect what you already use.** Email, calendar, Slack. No new software.
2. **Narrow the access.** Its own login, read-only wherever you can, write access only where it saves real time.
3. **Give it a working folder** next to the brain (below), with one file per repeatable task.
4. **Write the review rule down:** what it can do freely, and what waits for your yes.

```
business/
  vault/         <- the brain from Step 1
  AGENTS.md      <- rules every agent reads first (copy to CLAUDE.md for Claude Code)
  skills/        <- one file per repeatable task: "write the weekly report", "prep a call"
  approvals.md   <- what it can do alone vs what waits for your OK
  .env           <- keys live here. Never in notes, never in chat
  logs/          <- what it did and when
```

**Your first three "hands" jobs (pick one, run it this week):**
1. **Inbox triage:** *"Read today's unread email. Draft replies for anything routine, flag anything that needs me, and give me a 5-line summary. Don't send anything."*
2. **Call prep:** *"For each meeting on my calendar tomorrow, read the person's page and our last meeting in the vault, and write me a 5-bullet prep note."*
3. **After-call follow-up:** *"File this call transcript into the vault, then draft the follow-up email with the decisions and to-dos. Leave it in drafts."*

**How to read the diagram.** Every arm has a gate. Nothing reaches a real person or a real account without passing through the draft checkpoint.

**Graduate when** a whole task runs start to finish and you only check the result.

**Proof you're here.** It pulls from and writes to your real tools, and its output lands ready to send, not ready to rewrite.

> **Under the hood (mine):** one front door. I message one bot, Tofu, and it routes the work to the right specialist. They all read from the same brain. That's the setup for Step 3.

---

## Step 3 · Delegation: one bot, one job

![One Front Door](/tools/climb/images/climb-3-front-door.png)
*Caption: You talk to one bot. It routes. Every specialist reads the same brain.*

**What this is.** At Step 2 you're still pressing go on everything. Now you assign whole jobs. One bot per job, each with its own instructions, its own tools and its own lane.

**Definitions.**
- **Front door (orchestrator):** the one bot you talk to. It decides which specialist gets the work.
- **Lane:** one area of the business a bot owns end to end.
- **Brief:** the bot's job description: what it owns, what tools it has, what "done" looks like, what it must never do.

**What tool to use.** Anything that lets you run more than one agent with separate instructions: Grok Bot, Claude Code, Codex, whatever. The order matters more than the tool.

**What I recommend.** Don't build one giant bot that does everything. Build a front door and a few lanes, and add lanes only when one gets crowded.

**How to structure it.**
1. **One front door.** You message one bot. It routes.
2. **One bot per lane.** Mine: Client Delivery, Growth, Content, Life & Systems, and an Executioner that just does the builds. Eleven bots in total, counting the front door and the specialists inside those lanes.
3. **Same brain.** Every bot reads the vault before it acts, so none of them start from zero.
4. **Guardrails.** Anything that spends money or messages a client waits for my OK. No exceptions until a bot has earned trust.

**Write each bot a one-page brief:**

```
NAME: [lane] bot
YOU OWN: [the whole job, e.g. "weekly client updates"]
READ FIRST: AGENTS.md, then [the vault folders you need]
TOOLS: [only the ones this job needs]
DONE LOOKS LIKE: [the output, where it lands, by when]
ALWAYS ASK ME BEFORE: spending money, messaging a client, deleting anything
REPORT: one line in logs/ every time you run
```

**The life lane matters as much as the business lanes.** Mine runs my AI accountability coach (it checks on me morning and night and asks "do it, shrink it, or kill it?" when something keeps slipping), my to-do app and my routines. Organizing your life is half the point. The business runs better when you do.

**Where the DIY gets heavy:** briefs for every bot, routing rules, permissions per tool, approvals that actually hold. This is the step where most people stall. If you'd rather have it built around your business, that's what my team does. **[Book a call with my team](https://qwavelabs.io/book)**

**How to read the diagram.** One arrow in, five out. You never pick the specialist. The gold badges are the only places the system stops and waits for you.

**Graduate when** work gets routed to the right bot without you deciding each time.

**Proof you're here.** You send one message and the right bot does the job, in its lane, using the brain.

> **Remember:** one bot, one job. When a bot starts doing two jobs, split it.

---

## Step 4 · Autonomy: it runs without you

![The Clock](/tools/climb/images/climb-4-clock.png)
*Caption: Schedules and triggers instead of reminders. Every run leaves a receipt. (Pin positions are illustrative.)*

**What this is.** The part everyone wants, and everyone jumps straight to. It only works now because Steps 1-3 exist. The system runs on schedules and triggers, not on you remembering.

**Definitions.**
- **Schedule:** a job that runs at a set time. "Every night at 11:30."
- **Watcher:** a job that checks something on an interval and only bothers you when it matters.
- **Trigger:** an event that starts a workflow. A booking, a sale, an alert.
- **Receipt:** the log line every run leaves behind. If you can't see what it did, it isn't autonomous. It's unsupervised.

**What tool to use.** Scheduled tasks in your agent (Claude Code, Codex, Grok Bot), or plain cron/launchd, running on a computer that's always on.

**What I recommend.** One always-on machine. Mine is an M2 Mac that never sleeps and runs the nightly jobs. Start with one schedule and one watcher, then add.

**How to structure it.**
1. **Schedules.** Mine: three times a day, the brain updates itself from my email, calls and messages. Every night at 11:30, it cleans up the notes.
2. **Watchers.** Every ten minutes on weekdays, a bot checks my comms for anything that needs me, and only pings me if something does.
3. **Triggers.** When a low-balance alert hits one of my bills, a bot tops it up. And it asks me before it pays.
4. **Receipts.** Every run writes one line to the log: what it did, when, what it touched. I can scan a whole day in 30 seconds.

**Copy this schedule sheet before you automate anything:**

| Job | When | What it reads | What it does | Waits for my OK? | Log line |
|---|---|---|---|---|---|
| Brain update | 3× a day | email, call transcripts, messages | files them into the vault | no (only writes notes) | "filed 6 items: [[links]]" |
| Comms watch | every 10 min, weekdays | inbox + DMs | pings me only if something needs me | no (only notifies) | "2 flagged, 14 ignored" |
| Nightly cleanup | 11:30 PM | the vault | merges duplicates, fixes links, lists stale projects | yes, for merges | "3 fixes proposed" |
| [your trigger] | when [event] happens | [source] | [action] | **yes if it spends or sends** | "[what happened]" |

To set one up, tell your agent: *"Every night at 11:30 PM, run [job] from skills/[job].md. Write one line to logs/ with what you did. Never send, pay or delete without asking me."* Claude Code and Codex can run scheduled tasks, or use cron/launchd on an always-on Mac.

**How to read the diagram.** The clock is a normal day. Everything on it happens whether or not I'm at my desk. The trigger sits outside the clock because it fires on an event, not a time.

**Graduate when** the business runs a normal week and nothing breaks.

**Proof you're here.** Workflows run all week without you pressing go, and you hear about problems before your customers do.

> **Remember:** autonomy isn't "no rules". It's the same guardrails from Step 3, running while you sleep.

---

## Step 5 · Compound: it runs the loop

![The Loop](/tools/climb/images/climb-5-loop.png)
*Caption: Give it a goal, the data and your playbook. It runs the loop. You approve.*

**What this is.** Once the foundation exists, every new run costs almost nothing. You give the system a goal, the data and your playbook, and it runs the loop. Output compounds. Your effort stays flat.

**Definitions.**
- **Loop:** make → check → ship → measure → learn, then again, a bit better.
- **Playbook:** the rules you already wrote in Base Camp and the Brain. The loop follows them.
- **Approval point:** the one place you stay in the loop. You approve; it does the rest.

**What tool to use.** The same stack you already built. Compound isn't a new tool. It's the brain + hands + bots + schedules pointed at one goal.

**What I recommend.** Pick **one** loop that actually moves the business (content, leads, sales or delivery) and run it until it's boring before you start a second.

**How to structure it.**
1. **Pick the goal.** One output, one number.
2. **Hand it the data and the rules.** Your calls, your past posts, your playbook.
3. **Let it run the loop.** Mine is content: my calls turn into clips, the bots edit them, schedule them, and check they actually went live. Once a week, a research run tells me what's working in my niche so I know what to film.
4. **You approve, it improves.** I just approve. When something's wrong, I fix the brain (the rule, the example, the SOP), not just the prompt, so it's fixed for every run after.

**My content loop, step by step:**
1. **Make:** my calls turn into clips.
2. **Check:** the bots edit them (best take, captions, loudness) and check the frames.
3. **Ship:** I approve. Then a bot schedules them.
4. **Measure:** a bot checks each post actually went live, and I look at what landed.
5. **Learn:** once a week a research run ranks what's working in my niche (that's my REMAKE kit), and the playbook gets updated.

**How to read the diagram.** Inputs on the left never change much. The loop goes around every week. The only button with your name on it is APPROVE.

**Graduate when** you're approving output, not producing it.

**Proof you're here.** One core output runs as a loop you only approve, and it's getting better on its own week over week.

> **Remember:** most people buy the bots first. Build the foundation, and this is what your AI journey looks like.

---

## Where are you right now? (2-minute self-check)

| Question | Yes / No |
|---|---|
| I have a one-page list of every recurring task and tool | |
| My top processes are written down as steps | |
| AI can answer "what did we decide with ___?" from my own notes | |
| AI is connected to at least one real tool and finishes a task end to end | |
| Work routes to the right bot without me deciding each time | |
| At least one workflow runs on a schedule or trigger, with a log | |
| One core output runs as a loop I only approve | |

**Your level is the first "No."** Start there. Not at the bots.

---

## Where to start today

**In the next 24 hours**
- [ ] Fill in the Base Camp worksheet. Every recurring process, honestly.
- [ ] Delete three things nobody would miss: a step, a tool, a report.
- [ ] Pick the one home your brain will live in (I'd pick Obsidian).

**In the next 7 days**
- [ ] Build the brain: Meetings, People, Projects, plus one "How we work" page. File your last 5 calls.
- [ ] Connect your AI to email or calendar and run one real task end to end, as a draft.
- [ ] Write down your approval rule: what it does alone, what waits for you.

**Then**
- [ ] One front door, one bot per lane, one brief each.
- [ ] One schedule, one watcher, one log.
- [ ] One loop pointed at the goal that matters most.

---

## The stack (what I use, not what you have to use)

| Job | What I use | Swap for |
|---|---|---|
| The brain | Obsidian vault | Notion, Google Drive |
| Front door + bots | Grok Bot (Tofu routes to the lanes) | Any multi-agent setup |
| Building + coding | Claude Code, Codex | Cursor, any coding agent |
| Always-on machine | An M2 Mac that never sleeps | Any spare computer or a cloud server |
| Hands | Email, calendar, Slack connectors | Whatever tools you already use |
| Life lane | AI accountability coach, to-do app, routines | The free COACH setup below |

Tools change every month. The order doesn't.

---

## Go deeper

- **The brain:** our full Grok Bot + Obsidian vault setup, /tools/vault
- **The life lane:** the free AI accountability coach (3 notes, 1 prompt, 2 check-ins a day), /tools/coach
- **The content loop:** how to pick every reel to film (code + guide), /tools/remake

---

## Want us to build it for you?

You can climb this yourself. Everything you need is on this page. But if you'd rather have it done right the first time, that's what our team does.

**Who it's for:** founders, creators and small teams who are running something real, feel the AI wave, and want a system that actually knows their business, without spending months wiring it up.

**What you get:** our team builds your AI second brain with you:
- **Base Camp done with you:** we map your processes and cut what you don't need
- **The brain:** one organized home for your meetings, people, projects and how you work, with your existing calls and notes filed in
- **Hands:** your AI connected to the tools you already use, with approval rules so nothing spends or sends without your OK
- **Your first lanes and schedules:** the jobs that free up the most of your week, plus the daily accountability coach on top
- a walkthrough so you know exactly how to run it

**What happens on the call:** we look at how you work today, where things live and what eats your week, and work out which level you're at. We tell you what we'd build first. If it's a fit, we map out the build. If it's not, you still leave with the plan.

**[Book a call with our team](/book)**

**Rules we build with:** bots come last · one bot, one job · anything that spends or messages a client waits for our approval · when it's wrong, fix the brain · if you can't see what it did, it shouldn't be doing it alone.