30 Minutes of AI Knowledge Your Competitors Don't Have | Upside #102

Strategy

2 Sept

How to Build an AI Operating System for Your Business

Most people are still using AI like it's 2024: asking it a question, getting an answer, hanging up. In this episode, Al walks through the framework he teaches in his live talks for moving from "AI as a consultant on the phone" to "AI as a consultant sitting at the table with you," and the four pieces you need to build first so you're not rebuilding everything every time a new model launches.

In this episode: Ali and Leon cover the four building blocks of an AI operating system: agents, context files, connectors and skills, using the story of the standardised shipping container to explain why standardising your own information matters more than picking the "right" AI platform.

Key takeaways

Standardise your information once (in markdown files an AI can read), and you can move between ChatGPT, Claude and Gemini without starting over each time.
An agent isn't an entity, it's a job description: role, goal, tools, guardrails and output, written the same way you'd brief a new staff member.
Skills should be pulled in only when needed, not carried at all times, or they eat up the AI's context window unnecessarily.

Timestamps

00:00 Why this episode matters: using AI like it's 2024 vs 2026
01:53 The shipping container story: how standardisation cut costs by 97%
04:46 The four-stage AI journey, and where most business owners are stuck
06:56 What changed since February: Anthropic's Cowork and the OpenAI/Anthropic arms race
08:58 What an agent actually is (and isn't)
10:12 Writing your agent's "job description": role, goals, tools, guardrails, output
11:43 Building the context "boxes": MD files, Claude.md/Agents.md, memory files, context folders
19:33 Connectors: what to give AI access to, and what Al still won't connect
22:06 The cafe rule
22:59 The auto-send warning for recorded meetings
23:49 Skills: two ways to build them, and the /morning briefing example
27:15 Recap: agents, boxes, connectors, skills

The detail

The framework rests on one idea: don't keep re-teaching each new AI platform who you are. Al uses the story of Malcolm McLean's standardised shipping container (which dropped freight costs from over $5 a ton to 16 cents) to make the point that arguing over which AI to use misses the real lever, which is building a standardised set of information any AI can draw on.

That standardisation happens through three sets of files: an instruction file (Claude.md or Agents.md, since different platforms look for different filenames), a memory.md file for standing instructions you add over time, and a context folder holding files on your business, your ICP, your tone of voice and anything else the AI should know before it starts work. Al notes you can have the AI run its own 20 to 40 minute interview with you to build these out, faster still if you dictate your answers.

On agents specifically, Al reframes the term: you're not building an entity, you're writing a job description with five parts, role, goals, tools, guardrails and output, the same way you'd brief a new hire. Permissions scale the same way trust does with a new team member: start with read only access, move to write access as the agent proves itself, and use each platform's built in "allow once" or "allow always" controls rather than granting everything up front.

For what not to connect, Al's rule of thumb is the cafe rule: if you wouldn't say it out loud in a cafe, don't put it into an AI. He also flags a practical gap most people miss, turning off auto-send on meeting recording tools like Fathom, so an offhand comment at the end of a call doesn't go out to everyone on it.

On skills, Al describes two ways to build them: describe the goal and have the AI interview you to build the skill file, or have it watch you do the task and turn that into a skill. His own daily briefing skill (triggered by typing /morning) pulls from Slack, email and calendar every morning at 8am to summarise what needs attention that day.