Build a Personal Agent That Knows Your Business Step-by-Step with This Tech Prodigy
The advanced playbook for personal agents that don’t break under pressure (not for beginners)
Have you ever met someone and instantly knew this person is smart? That happened to me a few months ago with Gray Mitchell. The rate of learning AI and GTM engineering with this guy is incredible. He’s just 18, but already building AI agents for business and I knew I had to get him on the show.
Now, before you give into the weekly workflow walkthrough, I’m going to warn you that this is on the advanced side. If this is your first time building a personal agent or your first time in n8n, I would not start here. Ok, with that caveat, let’s dive into this issue of Stack & Scale.
This week’s Stack:
1 video: Live-build a personal AI agent (advanced)
1 prompt: Your strategic sparring partner
1 tool: Scrape attendee lists (and the web) without the click-and-copy grind
3 resources: Spot AI-written text, AI growth loops, and why brand is the moat in AI
3 jobs: Some awesome director-level marketing roles at AI-first companies
Let’s dive in.
Advanced Workflow Walkthrough: Your Personal AI Agent
Meet Gray Mitchell. At 18, he’s already building agentic AI systems that swap brittle vector DBs for durable graph memory. In this episode, we live-build a personal agent that pulls from your actual company context (meetings, docs, CRM, calendar) so it answers like someone on your team, not a demo bot.
Again, if you didn’t catch the note in the intro, this is a more advanced AI build and a longer video.
Here’s what you’ll learn:
How to fork the LightRAG template, set env variables, and mount a volume on Railway so your graph persists.
Why graph memory + a binding LLM beats ranked-chunk RAG for questions that span multiple sources.
Model choices that won’t break the bank—like $10 on OpenRouter lasting months, and when to use smaller OSS models vs. big commercial ones.
Where ZEP’s graph memory fits in (and what’s changing fast).
How to tighten agent behavior with prompts and sub-agents so it doesn’t spin out when you add too many tools.
If you’ve been stuck watching “cool demos” on social that collapse under real workloads, this walkthrough is the practical path to an agent you can actually use day-to-day.
Grab the leave-behind:
Prompt of the Week: Thought Partner and Idea Sharpener
This week’s prompt comes one the back of a Forge newsletter written by one of my partners in The Forge, Sam Calhoun. In his newsletter, he shared the customGPT he leans on the most: a no-BS work confidante. The kind that pokes holes in your assumptions, forces clarity, and refuses to let you hide behind “I’ll circle back on that.” I took Sam’s idea and shaped it into a full prompt you can copy, paste, and run, so you’ve got your own sparring partner and strategy sharpener on call 24/7.
You are a sharp, no-nonsense strategic partner built for [insert your role: e.g., high-level marketer, founder, team lead].
Your job is to pressure-test ideas, force clarity, and hold me accountable so that my decisions drive real-world outcomes—not vanity wins. You are not here to flatter, agree, or “just generate content.” You are here to act as my trusted work confidante: the sparring partner who keeps me sharp, focused, and execution-ready.
Core Directives
1. Role Definition
Act as a work confidante: a coach, strategist, and sparring partner who challenges my thinking and keeps me aligned with [insert your ultimate goal: e.g., profitable growth, faster execution, stronger positioning].
2. Optimization Targets
“Winning” for me looks like:
A clear and well-defined [insert ICP / audience / customer profile].
A defensible [insert differentiator / USP / market edge].
Execution that ships [insert desired outcome: e.g., revenue, user adoption, retention].
3. Challenge Orientation
Do not flatter. Poke holes in assumptions, highlight contradictions, and force me to make tradeoffs. Pressure-test strategies as if [insert high-stakes context: e.g., I had $50k and one quarter to prove results].
4. Structured Action Bias
Ground every response in real-world actions. Favor ≤7-day tests, pass/fail thresholds, and [insert preferred artifact types: e.g., dashboards, templates, outreach scripts, SOPs].
5. Question-First Reflex
If my request is vague, do not guess. Ask up to 3 sharp clarifying questions before responding.
6. Multiple Thinking Modes
Be able to switch between:
CFO Mode: ROI, payback periods, downside risk.
Red Team Mode: Skeptical objections and failure modes.
Coach Mode: What can realistically ship by [insert next milestone date or cadence: e.g., end of week, end of sprint].
Operator Mode: SOPs, workflows, and repeatable processes.
7. Deliverable Style
Always respond with:
Bullets, numbered steps, or pros/cons lists.
No meandering walls of text.
Crisp, clear, structured output.
8. Guardrails
Do not invent results or make unsupported claims.
Do not sugarcoat or end with filler like “hope this helps.”
Flag gaps in my thinking and propose scrappy, low-cost fixes.
9. Accountability Hook
Call me out when I’m chasing [insert personal distraction patterns: e.g., short-term dopamine hits, shiny tools, vanity metrics] instead of [insert long-term compounding goal: e.g., sustainable revenue growth, durable brand equity, repeatable pipeline].
I’ve added this to my The Ultimate ChatGPT Prompt Library for B2B Marketing Leaders notion doc. Check it out for 60+ more prompts.
The Ultimate ChatGPT Prompt Library for Marketing Leaders
I remember the first time I used ChatGPT for marketing. It was late, I was up against a deadline, and I needed a competitive analysis that would have taken me and my PMM a full day (or more) to pull …
Know someone who might find this prompt useful? Share it with them!
Tool of the Week: Scape Website and Download Data with AI
Conference season means one thing: booking meetings before you’re on the ground. The problem? You never really know who’s going until you’re already there. At a recent event, the organizers dropped all the attendees into a Slack channel. Technically, you could click through every profile one by one…or you could let AI do the heavy lifting.
I used Thunderbit to scrape the list (shhh, don’t tell the organizers I cheated) and had a clean set of names and details ready in minutes. What surprised me was how good it is beyond conferences too. It’s basically a Swiss Army scraper for the web. Whether you’re prospecting, grabbing competitor intel, or just trying to avoid the click-and-copy grind, it’s a solid add to the stack.
Try it: thunderbit.com
AI Resource Roundup
SaaS Metrics Palooza 2025 (Benchmarkit)
I’ve always loved Palooza for the benchmarks, but this year I’m extra fired up, because it’s all about AI colliding with SaaS. More than half the sessions dive deep into how AI is changing the way we build, price, market, and sell software, and what that means for the numbers we live and die by.
Some of the biggest names will be presenting, like Craig Rosenberg (Scale Venture Partners), Dave Kellogg (Balderton Capital), Peter Walker (Carta), Gary Survis (Insight Partners), Jeff Epstein (Bessemer) Katherine Dunn (ICONIQ) to name a few. If you want a clear view into how AI is reshaping SaaS growth and efficiency, this is the one to catch.
How to Spot AI-Written Text (Wikipedia Guide)
AI copy has tells—odd phrasing, weird repetition, and “average-sounding” structure. This page is a handy reference for sharpening your detection radar, especially if you’re reviewing vendor decks, resumes, or intern applications that feel…off.
The GTM Moat Series: Brand in the AI Era (The GTM Newsletter)
In a market where tech advantages evaporate fast, brand is the moat. This essay breaks down why positioning and trust matter more than raw model access, and how to build a durable edge before the noise catches up.
Hot AI Jobs
Here are 3 awesome roles at some of the hottest AI companies. They might be a stretch, but these are all opportunities that will make your career.
Marketing Insights Research Manager at Anthropic
City / Remote: San Francisco, CA (Hybrid; ~25% in-office)
Pay Range: $200,000–$255,000 base
Enterprise Marketing (US) at ElevenLabs
City / Remote: Remote (US) or NYC/SF office
Pay Range: Not listed
Product Marketing Lead, AI at Snowflake
City / Remote: Menlo Park, CA (Hybrid)
Pay Range: $175,000–$249,900 base (bonus & equity eligible)
As I said last week, conference season is in full swing, and now I’m off to Pavilion’s GTM Summit in DC. If you’re there, ping me and let’s meet up. Then it’s wheels down back home and I’m grounding myself for at least a month.
Until next week, keep scaling!
– Brandon