AI Bootcamp · Week 2

The ways I actually use AI, with the prompts.

Not theory. Not a list I found online. Every one of these runs in my own business — here's what it does, why it's worth your time, and how to set it up.

What we actually did live in Week 2: opening escrow by voice off a real contract and two counters, training your voice profile and saving it as a skill, and setting up your memory layers and projects. The full recap is below, in order, with the prompts we used live. Everything else on this page is extra credit — the prompts all work today, and email triage is the centerpiece of Week 3.

You are not expected to do all of these. Pick one. Run it this week. Come back for another. If you try seventeen things you'll finish zero.

Copy the prompt, change what's in [brackets], and say it out loud instead of typing it. You'll give AI more useful context talking for twenty seconds than typing for five minutes.

Each one is tagged Run it now works today in the Claude app · Can automate can go on a schedule · Claude Code the deeper lane, Week 5.

Week 2 · Full session

Watch the Week 2 recording

Missed it, or want to run something back? The whole session, including the escrow demo start to finish.

▶  Watch the recording
📄 The slidesPDF · 20 slides 📂 The demo deal filesContract + both counters 📋 The recapWhat we covered, in order
📋 The Week 2 recap — what we covered, in order The whole session from the recording, with the prompts we used live. Follow it top to bottom and you're caught up. +

The through line

Week 1 taught AI who you are. Week 2 taught it your business — how to write in your voice, and how to hold your context so you stop re-explaining yourself — then pointed it at the messiest hour in a transaction to prove the setup was worth it.

1 · Show and tell — the website
  1. This site was built over a weekend by directing AI, not by writing code. If you want your own, the honest how-to is the website card on this page.
  2. Why bother with any of this: every email, text, and update you push off your own plate is brain power you get back for the work that actually pays — prospecting and being with clients.
2 · The escrow demo, by voice
  1. A demo purchase agreement plus two counters, loaded into a folder AI can read. Terms change across the counters — on purpose.
  2. One spoken command, and it reconciled the terms and drafted the email to every party — buyers, listing agent, TCs, escrow, title, lender, inspectors, home warranty, insurance — each scoped to only what that party needs, plus the texts, all saved to drafts. Nothing sends without you.
  3. The catch: the final price lives in only one of the three documents. If your AI gets it right, it read everything. The answer key is on the escrow card, with all three files so you can run it yourself.
  4. Roughly an hour of notifications became about fifteen minutes of reviewing and hitting send.

The command I gave it, more or less word for word:

We opened escrow. Draft emails and texts to all the parties involved — use the purchase agreement and both counters in the escrow folder.

That short version works because the context was already set up. The refined, works-for-anyone version is on the escrow card.

3 · The context sandwich
  1. Every good prompt carries three things: who this is for, what you need, and what good looks like.
  2. If it's visual — a flyer, a post — attach an example of the look you want. AI does far better looking at one than reading your description of it.
4 · The voice profile lab
  1. The prompt sends AI into your own Sent folder for about fifty substantive emails — skipping the one-liners and anything AI wrote — and builds a voice profile from how you actually write. Full prompt with the copy button is on the voice profile card.
  2. Credit where it's due: the original is Igor Pagani's, refined over about two years of teaching this. The version on the card pulls your emails automatically instead of making you paste samples.
  3. What comes back is a markdown (.md) file — just a plain text format AI reads easily. Download it.
  4. Upload it as a skill, in your Claude settings under Skills. It's a refining process: when a draft doesn't sound like you, edit the skill (open it, three dots, edit) and it keeps getting closer.
  5. Don't paste the whole profile into your Instructions field. Instructions get read on every single prompt — slower, and it burns usage. A skill loads only when it's needed.

The one line that does go in your Instructions field (Settings → Instructions for Claude, at the bottom):

Whenever you're writing anything for me — an email, a text, a Slack message, a social post, any communication — use my voice skill.

That line is the switch. Ask for an email and it pulls your voice skill; ask for anything else and it leaves it alone.

5 · Building your projects
  1. Claude reads context in layers: your account-wide Instructions, its own Memory, and then each project's instructions and memory on top. They stack — the full walkthrough is on the memory layers card.
  2. Everything you do should live in a project. New chat? First question is which project it belongs in. No project for it yet? Start one.
  3. Project memory builds itself as you work — buyers, escrows, decisions, it logs the thread without you writing anything.
  4. You can also point it at your history: "check our past chats about this buyer and everything that's happened this escrow so far." Five to ten chats of context is plenty; don't make it read everything you've ever typed.

The trick for a project you've already been working in — don't write the instructions yourself:

Based on all my chats in this project, write me project instructions. I'll paste them into the project's instructions field. Keep them focused on how you should help me in this area of my business.

Starting a brand-new, empty project instead? Use the interview prompt on the memory layers card — it asks you questions first, then writes the instructions. Update them every few weeks.

6 · Teased for Week 3
  1. Email triage — your inbox read, sorted by priority, with a reply drafted in your voice on everything that needs one. We set it up together in Week 3; the card with both prompts is already below if you want to run ahead.
  2. Homework: pick one thing from this page and actually run it. Keep talking to AI instead of typing — and if it tells you it can't do something it's done before, ask again.
📚 The rest of the library Everything else we didn't get to live — 26 use cases in all, browsable by topic. +

Start here

The setup work. Two things, and everything else on this page gets better once they're done. If you only do one thing off this page, make it the first one.

✍️ Build your voice profileWeek 1 recapRun it now Teach AI how you actually sound, using fifty emails you already wrote. +

What it does

Reads about fifty of your own real emails and writes down how you sound. How you open, how you close, your rhythm, your words, and the things your voice never does. You save that once, and from then on everything AI writes for you comes out sounding like you instead of like a press release.

Why it's worth it

This is the one that makes everything else on this page usable. Drafted replies, follow-ups, captions, none of it saves you a minute if you have to rewrite it. You already wrote the training material. It's sitting in your Sent folder right now.

The prompt

I want you to analyze my writing voice and create a reusable Voice Skill file I can use to make AI write like me.

## STEP 1 - GO FIND MY WRITING

Don't ask me to paste examples. Search my Gmail Sent folder and pull my actual writing.

What to look for:
- Roughly 50 of my LONGEST, most substantive sent emails. Real messages where I explained something, persuaded someone, handled a problem, or wrote at length.
- NOT one-line replies. Skip "sounds good," "thanks, see you then," "got it," and anything under about three sentences.
- Variety matters: client emails, problem-solving emails, intro and follow-up emails, anything longer and more personal.

CRITICAL - skip anything AI wrote:
- If an email reads like AI wrote it, exclude it. That's not my voice, that's ChatGPT's.
- Tells: em dashes as connectors, "I hope this email finds you well," "I wanted to reach out," "Let's dive in," perfectly balanced three-item lists, no contractions, no typos, uniform paragraph lengths.
- If I started using AI at some point, reach back to BEFORE that. Go back one, two, even three years. Older writing beats recent AI-assisted writing.
- Tell me what date range you used and why.

Before you analyze, report back: how many emails you found, the date range, and how many you excluded. If you found fewer than 15 substantive ones, STOP and tell me so I can paste some manually.

## STEP 2 - ANALYZE

For every email examine:

Structure: How do I open? How do I close? How long are my paragraphs? Headers, bullets, or flowing prose?

Rhythm: Short sentences, long, or mixed? Do I use fragments? Do I vary length for effect?

Word choices: Contractions? Slang? Shortcuts like "asap" or "thru"? Industry language? What's my register - friend, client, colleague?

Personality: Humor, empathy, authority, warmth, directness? How do I handle bad news, disagreement, or an upset client?

What I DON'T do: What punctuation do I avoid? What phrases would sound fake in my voice? What tonal registers never show up?

EVIDENCE RULE: Every pattern must be backed by a direct quote from my actual emails. In 2+ emails it's a rule. Only once, it's a tendency - say which. Never invent or paraphrase. Quote me.

## STEP 3 - BUILD THE FILE

Create this exact structure. It must work standing alone as a Skill file.

CRITICAL: the file MUST begin with the YAML frontmatter block below, exactly as shown, starting on line 1. Three dashes, the two fields, three dashes. Nothing before it, no code fences around it, no blank line above it. If this block is missing or malformed, the file will be rejected when I try to upload it as a Skill.

---
name: [my-first-name]-voice
description: Write in [My Name]'s voice. Use whenever writing an email, text message, Slack message, social post, or any other communication on my behalf.
---

# [My Name] - Voice Skill

You are writing as [My Name]. [One sentence about what I do and who I write for.]

Extracted from my actual sent email across [date range]. Internalize these as a whole portrait, not a checklist. The goal is writing I'd read and think, "yeah, that sounds like me."

## Core Identity
[One paragraph. Not personality adjectives. Describe the RELATIONSHIP between me and my reader. What should they FEEL?]

## Voice Principles
[5-8 principles. Each one:]

### [Number]. [Principle Name]
[2-3 sentences on what it means and why it matters.]
The rule: [One actionable sentence an AI can follow.]

| Instead of | Write |
|---|---|
| [Generic AI-default] | [How I actually write it - real quote] |
| [Another generic] | [My version] |

## Signature Patterns

### How I Open
[3+ actual opening lines. Then the pattern.]

### How I Close
[3+ actual closing lines, then the pattern, including how I sign off.]

### Transitions
[The actual transition words I use, pulled from my emails.]

### How I Handle Hard Things
[Bad news, disagreement, chasing someone gone quiet, an upset client. 2-3 real examples and the pattern. This is where voice matters most.]

## Anti-Patterns: What My Voice NEVER Does
[5-10 things that would make writing NOT sound like me. Each one:]

### No [Pattern Name]
[One sentence: why this breaks my voice.]

| Instead of | Write |
|---|---|
| [Wrong version] | [Correct version for my voice] |

Look for: punctuation I avoid (em dashes? exclamation points? semicolons?), filler that'd sound fake ("It's worth noting," "Let's dive in," "I hope this email finds you well"), structural patterns I never use, tonal registers that don't fit, and AI tells ("I'd be happy to help," "Great question!").

## Formatting & Language Rules
- Paragraph length: [typical sentence count]
- Sentence style: [short? long? mixed?]
- Contractions: [always/sometimes/never, which ones]
- Shortcuts: [do I write "thru," "asap," "ppl"?]
- Headers and bullets: [when I use them vs write it out]
- Bold/italic: [what gets emphasis and why]
- Emoji: [frequency, placement, which ones if any]
- Sign-off: [exactly how I end an email]

## Format Adaptation

| Format | What stays the same | What flexes |
|---|---|---|
| Text message | [core principles] | [length, formality] |
| Quick email reply | [core principles] | [what changes] |
| Long client email | [core principles] | [what changes] |
| Something public | [core principles] | [what changes] |

## The Voice Check
[3 specific questions to verify a piece sounds like me. Based on patterns you actually found, not generic.]

## RULES FOR YOUR ANALYSIS
- The "Instead of / Write" tables are the most important part. At least 15 across all sections.
- Pull ACTUAL phrases as evidence. Never paraphrase or invent.
- Anti-patterns matter as much as principles. AI defaults to generic unless told not to.
- Write it as instructions TO an AI: "Write as..." "Never..." "Always..." - not "The author tends to..."
- Output a reusable file that works immediately if saved as a skill.
- Start the file with the YAML frontmatter block. This is not optional. A file without it cannot be uploaded as a Skill.
- Give me the finished file as a downloadable .md file, not just text in the chat.

Now the part that actually matters

After it hands you the profile
  1. Open it and actually read it. Does it sound like you? Read a few lines out loud if you're not sure.
  2. Fix what's wrong. Edit it yourself, or just tell Claude "I'd never say that" and "you missed that I always ask about their family first." Keep going until it's right. If you run out of time today, finish it tonight. It's worth the second pass.
  3. Save it as a Skill. Not pasted into a chat you'll lose by Thursday. A Skill is a file it can pick up any time you need it.
  4. Then add one line to your account-wide "Instructions for Claude," the same field your Personal DNA from Week 1 lives in. Copy it below.

If the upload says it needs YAML frontmatter: your file is missing the header block. Open the .md file and make sure the very first line is three dashes, then a name line, then a description line, then three dashes again. Nothing above it. If it's missing, just say "add the YAML frontmatter block to the top of that file" and Claude will fix it.

Whenever you're writing anything for me — an email, a text, a Slack message, a social post, any communication — use my voice skill.

That line is the whole reason to make it a Skill instead of dumping the profile into your Instructions. Your voice profile loads only when it's actually relevant to writing something, instead of being read on every single chat. The rest of the time it stays out of the way.

🧠 Stop re-explaining yourself in every new chatRun it now Set up Claude's memory in layers so it already knows your business before you finish the sentence. +

What it does

Most people use AI like a stranger they meet every morning. New chat, same forty seconds of "I'm a realtor in the San Gabriel Valley, I work with first-time buyers, my brokerage is..." That's the tax you're paying, and you can stop paying it.

Claude reads context in three layers, and they stack. Before it answers you, it's already pulled from all three.

Layer 1 · Instructions for Claude
  1. Your account-wide field. This is the Personal DNA we wrote in Week 1. Who you are, what you do, who you serve, how you like to be talked to.
  2. Read on every single chat, forever. This is why Week 1 mattered. Everything else on this page assumes it's filled in.
Layer 2 · Memory
  1. Things Claude saved on its own out of your past conversations. Your lender's name, the fact that you never text clients before 9am, the address of the listing you've been working all month.
  2. You mostly don't manage this one. It builds itself as you work, and it gets pulled in only when it's relevant to what you just asked. Ask about a listing and it surfaces listing things, not your recruiting notes.
  3. You can steer it. "Remember that I never quote rates in writing" and it keeps that. "Forget what I said about the Wilson deal" and it drops it.
Layer 3 · Projects
  1. A Project is a walled-off workspace with its own instructions and its own files. Its instructions only load when you're working inside it.
  2. Think of it like a client binder instead of the whole office. When you're in your Listings project, Claude reads your listing playbook, your disclosure checklist, your marketing standards. When you leave, that context stops following you around.
  3. The chats you have inside a project stay together, so a month from now you can find the thread instead of scrolling your whole history.

Why it's worth it

The difference between an agent who says "AI is fine, I guess" and an agent who runs their business on it is almost never the prompting. It's whether the AI knows anything about them. A generic assistant gives generic answers. Nine minutes of setup here changes the output on every prompt on this page.

And the second layer matters more than people expect. Once your projects and your memory are real, you stop writing long prompts. "Draft the follow-up" starts being enough, because it already knows who, about what, and in what voice.

The projects to build first

Don't build twelve. Build the three where you repeat yourself the most. For most agents that's Listings & sellers, Buyers & transactions, and Marketing & content. Add a Recruiting or Farm project later if that's your business.

The prompt

Here's the part almost nobody does. Don't write the project instructions yourself. Make Claude interview you and write them. Start a chat inside the new empty project and paste this.

I'm setting up this Claude Project for one area of my real estate business: [LISTINGS AND SELLERS].

I want you to write the custom instructions for this project. But don't guess, and don't write them yet.

First, interview me. Ask me one question at a time, and use my answers to decide what to ask next. Dig into:
- What I actually do in this part of my business, start to finish
- Who the people are: my clients, my vendors, who I hand things off to
- What I repeat over and over that you could take off my plate
- My standards and my non-negotiables, the things I'd correct you on
- How I want things formatted when you hand them back to me
- What's off limits, what I never want you to say or promise on my behalf

Keep asking until you genuinely have enough. Don't stop early to be polite, and don't ask me things you could infer from what I already said.

When you're done, write the project instructions and show them to me. Write them as instructions to you, not as notes about me. Then tell me what files I should upload into this project to make you better at it.

How to run it

In the Claude app, about ten minutes
  1. Left sidebar, Projects, then new project. Name it after the area of your business, not the task.
  2. Open a chat inside it and run the interview prompt above. Answer out loud with the mic instead of typing. You'll give it far more in two minutes of talking than in ten of typing.
  3. Paste what it writes into the project's instructions field. Read it first. If a line isn't true about you, fix it now.
  4. Upload your files. Your listing presentation, your buyer packet, your disclosure checklist, your vendor list. This is the step people skip and it's where most of the value is.
  5. From now on, do that kind of work inside that project. That's the whole discipline. The layers only pay off if you actually go in the right room.

Revisit the instructions every few weeks. When you catch yourself correcting Claude on the same thing twice, that correction belongs in the project instructions, not in the chat.

One more thing: reference an old chat

You don't have to rebuild context you already built. In a new chat you can point Claude back at an old one and it'll pull the thread forward.

Look back at our earlier conversation about [the Arcadia listing / my open house follow-up system / the counter offer we reviewed] and pick up where we left off. Summarize what we decided so I know you've got the right one, then keep going.

Works best when the old chat lived in a project, because then Claude knows exactly which pile to look in. Another quiet argument for setting the projects up.

If you ever move to Claude Code
  1. Most of you won't for a while, and that's completely fine. Everything above is the right answer in the Claude app.
  2. But it's worth knowing where this goes. In Claude Code, which is the developer version I use for building, all of this happens automatically off a folder on my computer. It reads my instructions file, writes its own memory files, and keeps project context per folder, without me managing any of it.
  3. Same three layers. The app version is just the manual one. Set it up by hand now and you'll already understand the tool you graduate into.

Your work, day to day

The recurring stuff that eats your morning. We'll walk through these together in Week 3, but nothing is stopping you from running them tonight.

🧭 Hire a chief of staffRun it now One AI that knows your whole business, hands out the work, and reports back. +

What it does

One project that holds your whole world: your buyers, your listings, who you owe what and by when. You talk to it in plain English after every showing and every call. It files what you tell it, surfaces what is due, and tells you what to work on next. Every other card on this page is an employee. This is the manager.

Why it's worth it

This is how I actually run my business. The morning brief, the inbox triage, the follow-up radar on this page: mine all report to one chief of staff that knows my priorities, my people, and my open loops. The change isn't any single trick. The burden of remembering what's open moved out of my head and into a system that never forgets a commitment.

The setup prompt

You are going to be my chief of staff. Before you start, interview me. Ask one question at a time and wait for my answer:

1. What does my business look like right now? Active buyers, listings, escrows, anything mid-flight.
2. Who are the people I owe something to right now, and what do I owe them?
3. What are my top three priorities for the next 90 days?
4. What slips through the cracks most often in my week?
5. What would the perfect Monday morning update from you look like?

Then write Project Instructions for this project that make you my chief of staff:
- Track my open loops: every commitment I mention gets filed with a due date. Never let one disappear.
- Track my people: clients, leads, sphere, each with the last touch and the next one.
- When I tell you something happened, file it and confirm in one line.
- When I open this project, do not wait for me to ask. Tell me what you would work on first and why.
- Never make me re-explain my business. Start every answer from what you already know.

Give me the instructions block to paste into the project settings. Then start the job: ask me for today's open loops.

How to set it up

In Claude — everyone
  1. Create a Project called Chief of Staff. Paste the prompt above into a new chat inside it.
  2. Paste the instructions it writes into the project's custom instructions. That is the hire.
  3. Feed it by voice, not typing. After every showing or call, 30 seconds out loud: "Promised the buyers the disclosure package by Friday, need to call the inspector, the listing goes live Tuesday." It files each one.
  4. Learn the one phrase that makes it real: "log a commitment: who, what, when." Every "I'll send that over" you say at a showing is a promise you either keep or leak. The moment you say it, tell your chief of staff.
  5. Give it a rhythm: a scheduled morning brief ("what is due, who went quiet, what is my one thing") and an evening check-in ("what did I say I would do that I didn't, carry it forward"). Both cards are on this page; point them at this project.

The discipline that makes it work: anything worth remembering gets told to the chief of staff. A commitment that stays in your head is invisible to the system, and the system can only protect what it knows.

📥 Inbox triage with the replies already writtenCan automate Sorts your unread mail and leaves a draft in your voice on every email that needs one. +

What it does

Reads your unread email, tells you what each person actually wants in one line, sorts everything into buckets by urgency, and drafts replies for the ones that need one. Two prompts: a one-time setup that teaches it your world, then a daily one that uses what it learned.

Why it's worth it

Mine runs on its own, multiple times a day, and I never start it myself. It drafts a reply to every email that needs one, sitting in Gmail waiting on me. Because it's trained on my voice, the drafts are close enough that I skim and hit send.

The run that convinced me: 22 unread, 2 needed me that day, 4 that week, 16 didn't need me at all. The drafts are what changes your morning. You're editing instead of staring at a blank reply box.

Step 1 — teach it your world (once)

I want to set up an email triage system that runs every morning and drafts my replies.

Before we build it, interview me. Ask me these questions one at a time, and wait for my answer before moving to the next one:

1. Who are the people whose emails always matter, no matter what they say? (clients, your broker, escrow, lenders, your team)
2. What deals or addresses are live right now that I should be watching for?
3. What kinds of email do I get in a normal week? Walk me through the categories.
4. Which of those actually need me, and which are just noise I can ignore?
5. What kinds of email do I always want to answer myself, that AI should never draft for?
6. Is there anything that should get flagged as urgent even if it looks routine?

After I've answered everything, write me a "Triage Context" block that captures all of it — my VIPs, my live deals, my categories, my priority rules, and my hands-off list. Format it so I can paste it straight into a scheduled task.

Keep it tight. This is a reference card, not an essay.

Save what comes back. That's the part that makes this yours instead of a generic triage anyone could run.

Step 2 — the daily prompt (paste your Triage Context into the top)

[PASTE YOUR TRIAGE CONTEXT HERE]

---

Go through the last 24 hours of my inbox using the context above.

Sort everything into four buckets:
- NEEDS ME TODAY — someone is blocked or a deadline is real
- THIS WEEK — real, but not today
- WAITING ON THEM — I already replied, tracking so it doesn't go cold
- NO ACTION — newsletters, receipts, noise

For each one that matters: one line on who it's from and what they actually want. Not a summary of the email, the ask.

Then draft a reply to everything in the first two buckets, in my voice, saved to Gmail drafts as a reply on that thread. Include the ones that are just a quick yes — those are the ones that sit for three days.

If a draft needs a number or date only I would know, write it anyway and leave a [BRACKET] so I fill one blank instead of writing the whole thing.

Never send. Drafts only.

End with one line: how many needed me, how many didn't.

How to set it up

In Claude — everyone
  1. Connect Gmail first: Settings → Connectors → Gmail. Use the official one.
  2. Run Step 1 once, in a normal chat, and save your Triage Context.
  3. Build your voice profile first if you haven't. That's the difference between drafts you send and drafts you rewrite.
  4. Run Step 2 by hand a few times until what comes back is what you'd have done yourself.

It drafts only. Nothing sends without you clicking send — that's by design and it isn't going to change.

Then put it on a schedule — in Cowork
  1. Click Scheduled in the left sidebar, then New task in the upper right.
  2. Choose Set up manually. Give it a name, and paste your Step 2 prompt (with your Triage Context already in it) into the Prompt field.
  3. Set Frequency to Daily, or On weekdays.
  4. Leave the Folder field empty. That's what makes it run in the cloud — on its own schedule, even with your laptop closed. Only pick a folder if a task needs local files.
  5. Save. Read the first two or three runs before you trust it, then tune it out loud: "you're drafting too many," "that sender is never urgent," "you missed my escrow."

Run it more than once a day if your inbox moves fast — morning and midday covers most people. Each run only needs to look at what's new since the last one.

In Claude Code — Week 5
  1. Just tell Claude Code to set this up as a recurring routine, same two prompts, running on your own machine.
☀️ Your morning briefCan automate Rates, your farm, your calendar, the weather — one message before your feet hit the floor. +

What it does

Assembles everything you'd otherwise open five apps to find, and hands it to you in one place. It reads your calendar and tells you what your day actually looks like, and it pulls the weather for where you'll be.

Why it's worth it

I used to open my weather app every morning. Now it's just there. That's the small version of the point. The bigger one is that I stopped rebuilding the same picture every day. Mine grew by conversation: I said "add the weather for my two cities" and two minutes later it was permanent. Start simple and let it grow.

The prompt

Every weekday at [TIME], give me a short brief on my day.

1. My calendar -- what's on it, who I'm meeting, and anything I need to prep for
2. Anything on my calendar where I should know something about the person before I walk in
3. The weather in [CITY] -- high, low, and what it's doing midday
4. Anything from yesterday I said I'd follow up on and haven't

Under 3 minutes to read. Short headers, bullets, no filler. If a day is light, say so in one line instead of padding it.

How to set it up

In Claude — everyone
  1. Paste the prompt into a normal chat first and see what comes back.
  2. Tell it what to fix. "Too long." "I don't care about that." "Add my calendar." Keep going until you'd actually read it.
  3. Then say: "Run this every weekday at 6am and send it to me."
  4. Once it's on a schedule it doesn't have to sit waiting for you in the app. It can be delivered straight to your phone. Mine comes through Telegram, so it's already read before I'm at my desk.

Build it first, automate it second. Scheduling a brief you haven't tuned just means a bad brief arrives daily. Your laptop needs to be awake for it to run.

In Claude Code — Week 5
  1. Same prompt, running as a task on your own machine.
  2. The difference: it can pull from your own files and databases, and deliver to your phone instead of waiting in the app.
📰 A daily brief on whatever you need to watchCan automate Housing news, rates, your farm area, new construction. It reads, you skim. +

What it does

Goes and researches one subject you care about, every day, and hands you back what actually changed and where it got it.

Why it's worth it

Mine runs on SoCal housing and mortgage. Pick the thing you'd have to stay on top of anyway and hand it off. Your farm area. New construction going up near you. Rates. What the brokerage down the street is doing. The agent clients think of as the person who knows what's going on is the one who gets the referral.

The prompt

Every morning, research [TOPIC OR AREA] and give me a short brief.

Cover:
1. What actually changed in the last 24 hours, with the source
2. What it means for my clients, in plain English, not industry-speak
3. One thing in here I could use in a conversation, a text, or a post this week

Skip anything that's a rewrite of yesterday. If nothing real happened, say that in one line.

Once a week, add: one thing I should be watching that hasn't happened yet.

How to set it up

In Claude — everyone
  1. Be specific about the topic. "Real estate" gets you a wall of nothing. "New construction and rezoning in [your city]" gets you something you can use on Thursday.
  2. Run it by hand a few days and tell it what to cut. Every "too long" and "I don't care about that" makes the next one better.
  3. Then ask it in plain words to run on a schedule, and give it a time. Same as the triage below.

Check the source on anything you'd repeat to a client. A brief is a starting point, not a citation.

In Claude Code — Week 5
  1. Tell Claude Code to set it up as a recurring routine, and have it delivered to your phone instead of waiting for you in the app.
🤝 Know who you're meeting before you walk inCan automate A brief on every person on tomorrow's calendar. +

What it does

Reads tomorrow's calendar and hands you a page on each person — who they are, what they're doing publicly, what you've talked about, and what you promised them.

Why it's worth it

This is the one people react to hardest. You walk into coffee already knowing they just listed in Arcadia and that you last talked in March about the buyer who fell out. Fair warning: it will surface the follow-up you forgot.

The prompt

Look at my calendar for tomorrow. For every meeting with someone outside my company, give me a short brief:

- Who they are and what they do
- Anything public and recent worth knowing (their site, their listings, their socials)
- What we've emailed about before, and anything I said I'd follow up on
- Two things I could genuinely ask them about

Keep each person to under a page.

How to set it up

In Claude — everyone
  1. You need Google Calendar and Gmail connected, both from Week 1.
  2. Run it tonight for tomorrow. See what it finds.
  3. Once you trust it: "Run this every evening at 8pm for the next day."

If someone has almost no online presence it'll say so rather than invent things. That's what you want.

In Claude Code — Week 5
  1. Mine pairs each meeting with two automatic nudges — one before saying "hit record," one right after saying "voice dump while it's fresh."
  2. Those notes get filed somewhere searchable, so six months later I can find what we actually talked about.
🔍 Who went quiet on youCan automate Finds every dropped thread and writes the next message. +

What it does

Looks across your conversations, finds who never replied to you and who you never got back to, ranks them by what matters, and drafts the follow-up.

Why it's worth it

Follow-up is the thing every agent knows they should do and almost nobody does. My last run surfaced 19 open loops I'd let go cold — and caught a typo in an email address that had been silently bouncing for weeks.

The prompt

Go through my sent email from the last 30 days. Find every conversation where I was waiting on a reply and never got one, and every one where they replied and I never answered.

Rank them by how much they matter to my business. For the top 10, write the follow-up message — short, warm, no guilt, easy for them to answer. Put them in my drafts.

How to set it up

In Claude — everyone
  1. Gmail connector on.
  2. Run it Monday morning. Read the list before you read the drafts — the list alone is usually the wake-up call.
  3. Schedule it weekly once you've done it by hand twice.

Don't let this one auto-send. A follow-up that goes out unread is how you email a past client about a house they already bought.

🎤 Talk it out with a business partnerRun it now Brain dump out loud, get back priorities, drafts, and next steps. +

What it does

You say everything that's in your head out loud, and it hands you back an order to work in. Not a transcript. A ranked list of what actually matters, the emails already drafted, and the things you said you'd do written onto your calendar.

Why it's worth it

Nobody ever has talker's block. Agents who say they hate writing don't hate writing, they hate starting. Talking is the one thing you already do all day.

This is the one I use most. Monday mornings, when I've got twelve things on my plate and no idea which one to touch first, I talk for five or ten minutes and let it sort me out. Not as a note taker. As a coach who tells me what to do first and what can wait.

The prompt — when the week is on top of you

Act as my business coach. I'm going to talk for five to ten minutes and dump everything on my plate right now — deals, open loops, people I owe a reply to, and the stuff I keep pushing to tomorrow.

I have [how many hours] to work today.

Don't just organize it back at me. Tell me:

1. The 3 things that actually move my business this week, and why the rest can wait
2. What I should drop, delay, or hand to someone else
3. Anything you can hear me avoiding

Then help me knock out number one. If it's an email or a text, draft it in my voice.

The prompt — right after a showing or a meeting

I'm going to talk for two minutes about what just happened. Don't clean it up into corporate language. Turn it into:

1. A ranked to-do list of what I actually need to do
2. A short update email to my client, in my voice
3. Anything I said I'd do that I should put on my calendar

How to run it

Nothing to set up
  1. Open Claude on your phone, hit the mic, paste the prompt, then just talk.
  2. After a showing, do it before you drive off. Ten minutes later you've forgotten half of it.
  3. In the morning, do it before you open your inbox. Once you're in email you're working someone else's list instead of your own.
  4. Push back on it. "That's not actually urgent." "You're missing that this client closes Friday." It gets sharper the more you argue with it.

This one only works if you actually dictate. Typing it defeats the entire purpose. And if you've built your voice profile already, the emails it drafts come out sounding like you.

With clients

The ones that show up in front of the people paying you.

🗣️ Translate a conversation, live, in personRun it now Real-time back and forth while you're sitting across from someone. +

What it does

Your phone becomes a translator for a real conversation. You speak English, it speaks their language. They answer, it tells you what they said.

Why it's worth it

I used this in an actual face-to-face conversation with someone who didn't speak English. In our market that's not an edge case, it's Tuesday. The alternative is losing the conversation or waiting on the adult child to translate — and the parents are usually the ones with the money.

The prompt

I'm sitting with someone who speaks [language]. Be our translator.

I'll speak English, you say it in [language]. They'll answer, you tell me in English. Keep it natural and conversational, not formal. Don't add commentary, just translate.

How to run it

Nothing to set up
  1. Phone, voice mode, speaker on. Set it on the table between you.
  2. Tell them what it's for first so they're not confused or offended. "Do you mind if I use this? I want to make sure I understand you properly."
  3. Works with Mandarin, Cantonese, Korean, Vietnamese, Spanish, Tagalog and plenty more.

Practice once before you need it. Fumbling with your phone in front of a client is worse than not having it.

💳 A stack of business cards into 20 personal follow-upsRun it now Photograph every card, talk through who you met, get emails and texts drafted for all of them. +

What it does

You photograph the whole stack of cards at once and talk through who each person was. It pulls every name, email and phone into a clean list, then writes a personal follow-up email and a text to each one, in your voice, referencing the specific thing you said about them.

Why it's worth it

Doing this by hand is the reason most cards die in a jacket pocket. Typing in twenty names, twenty emails, then writing twenty messages that don't sound copy-pasted is an hour of work, easily more. So it doesn't happen, and the whole reason you went to the event evaporates.

By voice it's five or ten minutes. You talk, it writes, everything lands in your drafts, and you review and send. Twenty people who met you once now hear from you the same night, each one referencing something real.

The prompt

Here are photos of the business cards I collected at [EVENT] tonight. I'm going to talk through who each person was.

First, read every card and pull the details into a clean list: name, company, title, email, phone. Flag anything you can't read clearly so I can fix it instead of you guessing.

Then, for each person, using my voice skill:

1. Write a short follow-up EMAIL that references the specific thing I said about them. Warm, brief, no pitch. Save it to my Gmail drafts.
2. Write a short TEXT for the same person. Shorter and more casual than the email. Just the message, ready to send.

Rules:
- Every message references something real about that person. If I didn't say anything specific about someone, write a warm general note instead of inventing a detail.
- No two messages should open the same way.
- Don't send anything. Drafts only, and give me the texts in a list I can work through.

How to run it

Nothing to set up
  1. Photograph the cards. Lay them out and get several in one shot, or shoot them in batches. It reads all of them.
  2. Upload the photos, paste the prompt, then talk. Go person by person: "Jimmy, met his daughter, she just started at Cal Poly. Brian is deep into AI, we talked for twenty minutes. Clarence lives two streets over from my listing."
  3. Build your voice profile first. That's the difference between drafts you send and drafts you rewrite.
  4. Review every draft before it goes. Always.

The voice dump is the whole thing. Skip it and you get twenty identical "great meeting you" emails, which is worse than sending nothing at all.

Send the texts fast — Claude Code
  1. Instead of a list you copy one at a time, have it build you a tap-to-text page: every person as a button, their message already loaded.
  2. Open it on your phone, tap, your Messages app opens with the text ready, hit send, next. Twenty texts in about a minute, each one from your own number.
  3. Same mechanic as the texting tool in this library. Nothing sends automatically, you still tap every one.
🏠 Make an inspection report make senseRun it now Forty pages of jargon into a calm email your buyer understands. +

What it does

Sorts the report into what actually matters, what's worth asking about, and what's normal for a house that age — then writes the email that walks your buyer through it.

Why it's worth it

The inspection is where deals die from panic. Most of what scares a buyer is routine. Separating real problems from noise, in language they understand, is most of the job — and it's the moment they decide whether they trust you.

The prompt

Here's the inspection report for [address]. My buyer is a [first-time buyer / investor / family] and they're going to be overwhelmed by this.

Break it into:
1. Things that actually matter — safety, structure, water, systems at end of life
2. Things worth asking about but not alarming
3. Normal wear that every house this age has

Then write an email to them in my voice that walks through it calmly and tells them what I'd suggest we do next. No jargon.

How to run it

Nothing to set up
  1. Upload the PDF straight into the chat.
  2. Say who the buyer is. The same report needs a different email for a nervous first-timer than for an investor on their fifth property.
  3. You still read the report. This makes it explainable, not optional.
📊 Comps that look like a presentationRun it now Your comp data as a clean branded page instead of an MLS printout. +

What it does

Takes your exported comp data and lays it out as something you'd be proud to hand across a kitchen table.

Why it's worth it

Every agent at that listing appointment hands over the same grey spreadsheet. The data isn't your differentiator — how seriously it looks like you took it is.

The prompt

Here's my comp data for [address] as a spreadsheet. Build me a clean, professional one-page summary I can hand a client — my branding, easy to read, the comps laid out so the comparison is obvious at a glance. Give me a version I can save as a PDF.

Don't add any conclusion about what the home is worth. That's my conversation to have.

How to run it

Nothing to set up
  1. Export your comps to CSV or Excel and upload the file.
  2. Tell it your colors, drop in your logo.
  3. Iterate. "Make the photos bigger." "Put price per square foot next to each one."

Keep the valuation out of the document. The formatting is the product; the number is your professional judgment and belongs in your mouth, not in an AI-generated PDF. Standard "this is not an appraisal" language still applies.

📩 Open escrow and notify every party at onceRun it nowWeek 2 demo Drop in the contract and counters, get a scoped email and text drafted for all eleven parties. +

What it does

You upload the executed purchase agreement and every counter. It reconciles the final terms across all of them, then drafts an email to each party in the transaction, sending each one only what's relevant to them.

Why it's worth it

Opening escrow is the most chaotic hour in the deal. Buyers, listing agent, both TCs, escrow, title, lender, inspectors, home warranty, insurance, that's eleven people who each need something different, pulled from terms scattered across a contract and two counters. By hand it's an hour and it's where things get missed, usually the detail buried in the second counter. By voice it's a few minutes and everyone hears from you the same afternoon.

The moment that lands it: the final price often isn't in any single document. It only exists after you reconcile all three. When AI surfaces the right number, you can see it read everything, not just the last page.

The prompt

We just opened escrow on [ADDRESS]. Attached are the purchase agreement and both counters.

Read all three and reconcile the final terms, price, credits, dates, possession, everything that changed across the counters. Show me the reconciled terms in a short table first so I can confirm them.

Then, using my voice skill, draft an email to every party: buyers, listing agent, both transaction coordinators, escrow, title, lender, general inspector, termite inspector, home warranty, insurance.

Send each party only what's relevant to them. The lender needs loan amount and credit structure; title needs the parties and property; inspectors need access and the deadline; nobody outside the deal needs the buyer's financials.

Also give me a one-page escrow summary for my own file, and a short text I can send the buyers congratulating them.

Save the emails to my Gmail drafts. Don't send anything.

Practice on the demo deal

This is the exact transaction I ran live in Week 2. Fake deal, fake people, real paperwork structure. Download all three, upload them into one chat, and run the prompt above. You'll see the whole thing happen on a deal where nothing is at stake.

  • 01 Purchase Agreement 123 Main Street, Arcadia. The original offer at $1,268,000, with a $20,000 credit request.
  • 02 Seller Counter Offer No. 1 Seller pushes to $1,295,000, cuts the inspection window to 7 days, rejects the credit entirely.
  • 03 Buyer Counter Offer No. 2 · accepted Where the deal actually lands. New price, a $15,000 credit split into a rate buydown, a 7-day rent-back, and a rebuilt timeline.

Or grab the whole folder here. Every page is stamped as a training document. Don't use these as contract templates, they're built to teach, not to transact.

Check its work

Here's the point of the demo. The final purchase price is $1,285,000, and that number appears in only one of the three documents. Not the contract. Not the first counter. If you'd read the purchase agreement you'd be off by $17,000, and if you'd read the seller's counter you'd be off by $10,000 in the other direction.

Two more to check it on: the seller credit is $15,000, ten of it a permanent rate buydown, five toward closing costs. And possession is a week after close, not at close, because of the rent-back that got negotiated in the last counter. That rent-back detail is exactly the kind of thing that gets missed by hand and then blows up two days before move-in.

If your AI comes back with a table that gets those three right, it read all three documents. If it doesn't, tell it so and make it reconcile again. That habit of checking the summary before you trust the drafts is the actual skill here.

How to run it

Nothing to set up
  1. Upload the PDFs straight into the chat. Don't rely on a Drive link, the text extraction lags and isn't dependable when you need it live.
  2. Confirm the reconciled terms table before it drafts. That's your check that it read the counters correctly, not just the original contract.
  3. Build your voice profile first so the emails sound like you.
  4. Review every draft before sending. Drafts only, it never sends on its own.

The whole point is that each party gets only what they need. Double check the lender and title emails, those are the two where the wrong detail in the wrong inbox actually matters.

Tools I built for you

Not prompts this time. Actual software, already running, free for the agents in this room.

💬 The texting toolRun it now Say who and what in one sentence. It builds the list, writes a personal text for each person, and you tap-send from your own phone. +

What it does

All your contacts live in it, organized into lists. When you want to reach out, you just tell it who and what: "Text my open house list and the Chens, invite them to Saturday's broker open at 143 Maple." It figures out the audience, shows you exactly who it picked so you can add or drop anyone, writes a different message for each person, and you send them from your own phone one tap at a time. It is not a blast. Every message passes in front of you before it goes.

Why it's worth it

Nobody skips follow-up because they don't care. They skip it because fifty personal texts is a two hour job, and a little bit of an emotional one. This turns it into about five minutes, and you can talk instead of type. Say who and what out loud while the coffee brews, review the drafts in the Starbucks line, tap them out between sets at the gym. And because every text goes from your number, one at a time, it reads like you thought about the person. That's the only kind that gets answered. Load your contacts once, a spreadsheet, pasted names, or your phone's exported contacts, and it becomes the one place your sphere actually lives.

Where to go

Two ways to get it
  1. Use mine: log in at yourguidehome.co/login. I built it for the realtors I work with, and I'll set you up. Your own page lives at yourguidehome.co/firstname-lastname. Jimmy Tram's is yourguidehome.co/jimmy-tram.
  2. Or build your own: if you run Claude Code, you can build your own version of this. Here's the playbook.

Same rule as everything else on this page. These go to people who know you. Your sphere, past clients, someone who handed you a card. Don't point it at a purchased list.

📋 The inspection report analyzerRun it now Upload the PDF, get a plain-English breakdown and the client email already written. +

What it does

You upload the inspection PDF. It comes back with what actually matters, what's worth asking about, and what's normal for a house that age. Then it writes the email to your buyer, in their language.

Why it's worth it

This is the inspection prompt above, except built. It knows what an FHA or VA appraiser is going to flag before it kills your escrow, and it knows what a California insurer refuses to write. Those are the two things that quietly blow up deals and neither one jumps out at you in a seventy page report.

Where to go

Same login
  1. Log in at yourguidehome.co/login and upload a report.
  2. Try it on an old one first, from a deal that already closed. That's the fastest way to find out whether you trust it.
  3. The client email comes out in English, Chinese (traditional or simplified), Spanish or Korean. Need another language, tell me and I'll add it.

You still read the report. This makes it explainable, not optional. The summary tells your buyer to read the full report too, on purpose.

Don't have your own page yet? Text me at 858-216-6091 and I'll set you up.

Content & marketing

Where most agents quit, because they're starting from a blank page five times a week.

🎓 Build an entire presentation, then get coached on delivering itRun it now The exact loop I used to build this bootcamp — outline, slides, and a rehearsal coach. +

What it does

Takes a presentation from nothing to finished: outline, slide deck, and then the part almost nobody knows about — it listens to your rehearsal and tells you what to fix.

Why it's worth it

I built this bootcamp with it. I took Tony Robbins' AI bootcamp transcripts and slides, dropped them in a folder, voice-dumped my own context — who my audience is, what I actually wanted them to walk out with — and had it build a realtor version. Weeks of deck work became days.

But the rehearsal loop is the part that changed things. Practice used to have no feedback. You run through it, you feel okay about it, and you have no idea what actually landed. Now every run-through gets scored.

Your version of this: a homebuyer seminar. Build it once, record it, and it becomes an evergreen lead magnet you send to every buyer for the next two years. Same for a listing presentation or an investor talk.

Step 1 — the outline

I'm building a [90-minute homebuyer seminar / listing presentation / investor talk].

My audience is [who they are and what they already know]. What I want them to walk out with is [the one thing]. What I want to happen afterward is [book a consultation / sign a listing agreement / join my list].

I'm going to talk for a few minutes about everything I know on this and everything I want to cover. Don't clean it up yet — just listen.

Then build me an outline: what to open with, the order of the sections, how long each should run, and where the interactive moments go. Tell me honestly what I should cut.

Have material from someone else — a training you paid for, a webinar you liked? Upload the slides or transcript and say "use this as the structure, but rebuild it for my audience." That's exactly what I did.

Step 2 — the slides

Turn that outline into a slide-by-slide structure I can hand to a slide generator.

For each slide: the headline, two or three short supporting lines, and what visual belongs there. Keep every slide sparse — I'm doing the talking, the slide is just the anchor. Tell me the exact number of slides.

Give it to me as one block I can copy and paste.

Paste that into NotebookLM — it makes the best-looking decks of anything I've tried. Download as PowerPoint, then bring it back to Claude for feedback and iterate. Two or three rounds gets it sharp.

Step 3 — the rehearsal coach (this is the one)

Here's the transcript of me practicing this presentation out loud.

Be honest, not encouraging. Tell me:
- What ran long and what I should cut
- Anything I said that was inaccurate, overstated, or that I can't back up
- Where I lost energy or rambled
- Where the audience would tune out
- What actually landed and I should keep

Then give me the specific changes to make before I run it again.

How to run the whole loop

Nothing to set up
  1. Voice-dump your brain, or upload someone else's material. Get the outline.
  2. Generate the slide structure, paste into NotebookLM, download the deck.
  3. Practice out loud and record it as a voice memo on your phone.
  4. Paste the transcript back in. Get scored. Fix what it flags.
  5. Run it again. Three or four rounds. The third one is where it clicks.

Record on your phone, don't type notes. The transcript is the input, and you want it to catch how you actually sound — the rambling, the parts you rush — not a cleaned-up version you wrote down afterward.

Then make it evergreen

Once it's good, have AI build a resource page for it — the handouts, the links, the recording. Host it free on Vercel or your own site. Now anyone who missed it can catch up, and anyone you meet next year gets the link.

That's the difference between a seminar you gave once and an asset that keeps generating conversations. This page you're reading is that.

📱 Fix your Instagram bio without typing it forty timesRun it now Screenshot your profile. It already knows the character limits. +

What it does

You screenshot your profile, say what you're going for, and go back and forth until it's right. No counting characters, no guessing where the line breaks.

Why it's worth it

I stopped retyping mine into the app just to see how it wrapped. We went back and forth about six times in ten minutes and landed somewhere I wouldn't have reached alone. Screenshot, don't retype — that's the whole shortcut.

The prompt

Here's a screenshot of my Instagram profile. I'm a [what you do] in [market] and I want people to immediately understand [what you want them to get].

Give me 3 different versions of the bio. Respect the character limit and the fact that only about 4 lines show before it truncates. Make the first line the one that stops someone.

How to run it

Nothing to set up
  1. Screenshot your profile as it looks on your phone.
  2. Upload it, paste the prompt, then keep talking. Version one is a starting point, not an answer.
  3. Same move works for your Zillow profile, your email signature, your website header.
When should I actually post?Run it now It researches it live — then tells you if you asked the wrong question. +

What it does

Researches current posting-time data from real sources and splits the answer by content type, instead of repeating advice everyone's been copying since 2019.

Why it's worth it

I asked because my content partner and I disagreed. It came back with Reels 6–9pm, teaching content 7–9am — and then told me posting time was the fourth most important lever, and my real problem was the first five words of my hook. That was worth more than the answer I asked for.

The prompt

Research the current best posting times for real estate content on Instagram, using real sources, not general advice. Split it by content type — Reels vs carousels vs teaching posts. My audience is [who] in [market].

Then tell me honestly: how much does timing actually matter compared to everything else I could fix?

How to run it

Nothing to set up
  1. Paste it and let it search.
  2. That last line is the important part. Add "tell me if I'm asking the wrong question" to almost anything and it will.

This is the habit worth stealing from this whole page: ask AI to push back on your question, not just answer it.

One real moment becomes a week of contentRun it now One walkthrough or voice note into a caption, carousel, email and posts. +

What it does

Takes one piece of real raw material and turns it into everything you'd otherwise write separately.

Why it's worth it

The reason people stop posting isn't laziness, it's the blank page five times a week. Start from one real thing instead. One open house, one deal that almost fell apart, one thing a client said that stuck with you.

The prompt

Here's a transcript from [an open house walkthrough / a voice note about something that happened this week].

Turn it into:
1. An Instagram caption in my voice
2. A carousel outline, 5 slides, one idea per slide
3. A short email to my sphere
4. Three post ideas that come out of the same story

Same voice throughout. Don't make it sound like marketing.

How to run it

Nothing to set up
  1. Record a voice note walking a property, or just talking about your week.
  2. Upload it and paste the prompt.
  3. The raw material has to be real. A genuine story about a deal that nearly died beats anything invented, every time.
🎨 Use two tools on purposeRun it now Strategy and copy in one place, images in another. That's normal. +

What it does

This one's a workflow note, not a prompt. I write the strategy, the hook and the caption in Claude, then generate the actual carousel images in ChatGPT or Gemini, because right now they produce more cinematic visuals.

Why it's worth it

Anyone telling you one tool is best at everything is selling something. Using two on purpose isn't a failure, it's knowing what each is good at. And it changes — check again every few months.

The split

How I actually run it
  1. Claude: what the post is about, the hook, the caption, the voice.
  2. ChatGPT or Gemini: the image generation.
  3. Bring the image back and write the caption around what you actually got.

The website

The site you're reading this on. I'm not a developer and I don't write code by hand — I directed AI tools. Here's the honest version of how.

🌐 Build your own website by directing AIClaude Code A real, live, cinematic site. No code written by hand, about $30 for the domain. +

What it does

You describe your business and what you want people to feel, and AI designs and builds the site. Not a template with your headshot dropped in. joshuasun.co is the proof — the main site, plus /ai for the bootcamp and /rebuild for the Altadena project. They scroll like a film instead of a page.

Why it's worth it

Every agent has the same site as every other agent. This is the fastest way I know to not look like everyone else. Full disclosure: I'm not a developer. I got there by directing tools, and no AI one-shots a finished site — expect to regenerate and adjust. Budget time for that, not just money.

The prompt — the part that actually matters

I want to build my own website and I'm not a developer. Before you build anything, I'm going to give you everything about my business.

[Then talk for as long as it takes. Who you serve. What you actually do for them. What you want someone to feel in the first five seconds. What you want them to do next. Don't summarize — more context is better.]

Then:
1. Ask me whatever you still need to know
2. Show me the design direction before you build — layout, colors, fonts, the feel
3. Build it once we agree on that

Don't hand me a generic agent template.

This is the whole difference. I gave it one source of truth and fed it my thinking, not a one-line prompt. The site came out close to what I actually meant because I front-loaded the why, not the what.

The tools

What I actually used
  1. Claude, for the design. The visual style, layout, colors, fonts, the feel of the whole thing — decided before anything gets built.
  2. Higgsfield, for the video and images. This is what makes it feel like a film instead of stock photos. Buy credits, then direct it the way you'd direct a photographer: describe the scene, the mood, the lighting. Iterate.
  3. An AI coding assistant to wire it together. I use Claude Code (Week 5). A ChatGPT-based coding setup is a cheaper alternative and gets you close.
  4. Music: Uppbeat, Artlist, or Storyblocks. I search "cinematic emotional buildup."

Credit where it's due: I followed Jack Roberts' video "Claude Fable 5 + New Design Skill" for the scroll-driven approach.

The domain — unglamorous but important
  1. joshuasun.co cost about $30 for the year.
  2. Claude handled pointing the domain and getting it live. That's normally the part that trips people up, and it really was as easy as it sounds.

The full walkthrough

I wrote the whole thing up, tool by tool, including what it cost and where I got stuck: How I Built a Website With AI — The Playbook.

The deeper lane

This is Claude Code — it runs on your laptop and can build and fix things, not just write them. We install it together in Week 5. Nothing here is required and nobody is behind for skipping it. It's here so you know where this goes.

🚀 Mission Control: your whole business on one screenRun it nowClaude Code A dashboard AI builds you in an afternoon. Lead measures big, lag measures small. +

See a live example → A sample built for a fictional agent, so you can click through before you build your own. Sample data only — nothing here is a real business.

What it does

One page: your escrows with the next deadline on each, your listings and where each is in launch, this week's new leads, the follow-ups due, the conversations you have started. Built as a single HTML file you save and open in your browser. No hosting, no login, no monthly software.

Why it's worth it

Your CRM shows you your CRM. Nothing shows you your business. Mine started as one conversation and now it is the first tab I see every morning: what is running, what is stalled, what needs me. The design rule that makes it honest: the numbers you control get the big font (calls, doors, posts, open houses), the results render small (closings, volume). You stare at the inputs; the outputs follow.

The prompt

Build me a one-page business dashboard as a SINGLE self-contained HTML file I can save and open in my browser. No hosting, no accounts, everything in one file.

Interview me first, one question at a time:
1. If I could only see five things about my business every morning, what should they be? Help me think: active escrows with their next deadline, listings and their launch stage, new leads this week, follow-ups due, conversations started this week.
2. Which of my numbers are lead measures, the ones I control weekly (calls, doors knocked, posts, open houses hosted)? Those get the biggest type on the page.
3. Which are lag measures (closings, volume)? Those render smaller, on purpose.
4. My brand colors, and whether I want it bright or calm.
5. Anything money-related that should hide behind a "demo mode" button before I ever screen-share this.

Then build the file: clean, bold, readable from across the room. Include the demo-mode button. One honesty rule: never show a number I did not give you. If a section has no data yet, show "not wired yet" instead of inventing something.

I will tell you what to change and we will go back and forth until I actually enjoy looking at it.

How to set it up

In Claude — everyone
  1. Run the prompt, answer the interview, download the file it builds, double-click it. That is the whole install.
  2. Set it as your browser's new-tab or startup page so it greets you every morning.
  3. Iterate out loud, exactly like decorating an office: "make conversations the biggest number," "add a family tab," "that gold is too much."
  4. Update it by telling Claude what changed and downloading the new version. Weekly is plenty to start.
In Claude Code — Week 5
  1. Claude Code turns the same page into a living board: it refreshes the numbers itself on a schedule from your real files, and a watchdog (card below) tells you if a routine quietly stops.

The moment real dollars are on this page it is orange-zone: build the demo-mode button before you ever present or screen-share, not after. A screenshot travels further than you plan for.

🔁 The meeting loopClaude Code Four small automations that do what a CRM is supposed to do. +

What it does

A nudge to hit record before a meeting. A nudge to voice-dump right after, while it's fresh. The transcript filed somewhere searchable. And a watchdog that surfaces the person again when they go cold.

Why I built it

I kept having great conversations and then losing them. Not forgetting the person — forgetting what we actually said, which is the part that makes the next conversation good. Four small reminders chained together beat a CRM I'd never open.

Where to start

You don't need Claude Code for the first version. Start with the meeting brief above, and add a phone reminder after each meeting that just says "voice dump now." That's 80% of the value with none of the setup.

🎬 Content at scaleClaude Code Drop clips in a folder, give direction by voice, get finished content back. +

What it does

I put raw clips in a folder, describe what I want out loud, and get back edited, captioned, branded content. Revisions are spoken, not clicked.

Why it matters

It removed about 90% of the editing work. Not 90% of the thinking — I still decide what the piece is about and whether it's any good. But the part where you drag clips around at 11pm is gone. This is the single biggest reason I'd tell you to look at Claude Code.

Where to start

We'll go through it properly in Week 5. Until then, the version you can do today: record the footage, and use "one real moment becomes a week of content" above to handle everything except the edit.

🛡️ A watchdog for your automationsClaude Code Because the danger isn't that they crash — it's that they quietly don't. +

What it does

Keeps a registry of everything running on a schedule and checks whether each one is actually still doing its job.

Why it matters

Once you have five things running automatically, the failure to fear isn't a crash — it's silent degradation. Mine caught five automations that had all fired nine hours late, in one burst, with no error message anywhere. I'd never have noticed.

Where to start

Nowhere yet. This only matters once you have things running. It's here so that when you do, you know to check on them.