The front tire washed out on a mountain bike berm, and Dan Martell was still writing code on his phone. That detail sits near the end of a dense, practical walkthrough Martell published on how to move from casual Claude user to full agent orchestrator, but it captures the philosophy running through the entire thing: the tools should work while you are not at your desk.
Martell says he has spent over 1,000 hours inside Claude and uses it every single day to build tools, run workflows, and launch companies. His claim is that most people are operating at roughly five percent of what Claude can actually do, treating it like a search engine rather than a system. He lays out six distinct user levels to explain the gap.
Where most people get stuck at level one
The amateur pattern, according to Martell, is one question in, one response back, then closing the tab. No memory carries over. Claude has no idea what the person is working on. Martell’s first fix is to change the initial prompt entirely. Instead of diving straight to the task, the user tells Claude: ‘Before you answer, ask me any questions that you need to perform this task properly.’ Claude then surfaces the context it needs before generating anything. The second fix is just as simple: after any output, type ‘check your work.’ Martell acknowledges it is frustrating that the model does not do this automatically, but the self-review pass consistently produces a better result.
Level two, which Martell calls ‘the regular,’ introduces Claude’s Projects feature. A project functions as a persistent workspace tied to a role, a client, or a workflow. The setup he recommends starts with naming the project after the role, then using Claude to build a master prompt. The prompt to get there is: ‘Interview me to build a master prompt for my role as a marketer.’ Claude asks questions, the user answers, and the output is a file containing everything about how that person works, what their team looks like, what tools they use, and what they are trying to accomplish. That file goes into the project alongside any company documents, sample data, or previous examples. Martell uses a project folder for every YouTube video he plans, loading it with his voice document, branding document, and past scripts he considers strong references.
A master prompt, he explains, is the ingredients. A system prompt is the recipe. He recommends asking Claude to interview you for the system prompt separately, then pasting the result into the custom instructions for that project. The same quality and format comes out every time after that.
Plugging Claude into the tools where work actually happens
Level three is what Martell calls ‘the integrator.’ This is where Claude connects to Gmail, Google Drive, Slack, Notion, and Calendar directly, so the user stops copying and pasting content between tabs. He runs a regular prompt telling Claude to scan Slack for the past week and give him a CEO-level summary of what is happening across the company. He also uses the Chrome extension so he can paste instructions generated in one chat directly into a live website and have Claude act on them in the browser.
Visualizations and interactive artifacts are features he flags specifically at this level. Claude can generate graphs, bar charts, and mockups inside the chat window. It can also build what Martell calls mini apps, outputs with clickable elements, sliders, and buttons that make working with data more immediate. He also points to Composer, a Google Docs-style editing surface inside the chat, where pasted text can be refined until the output matches what the user actually wants.
Level four is ‘the operator,’ where Martell says the shift moves from doing to directing. The operator sets up tasks that run on their own and only reviews or approves results. Three tools make this possible. First, system prompts defined for repeatable outputs. Martell’s view is that system prompts will become the core intellectual property of teams and companies. Second, skills: saved workflows triggered by a slash command. His rule is that if he is doing something more than three times a week, it becomes a skill. He built a company status skill that pulls analytics, reports, and team updates and returns a concise summary on demand. Third, scheduled tasks using a separate app called Cowork, which can take over a computer and run a job autonomously. Every night at 8:00, a scheduled task reviews his calendar and emails and delivers a briefing covering the following day.
He describes chaining skills as the key move at the operator level. A copywriting skill that writes in his voice feeds into an email-writing skill, and both get called by a broader inbox automation skill. Each skill is a self-contained unit, and linking them produces a full pipeline that runs without manual involvement at each step.
Level five introduces Claude Code. Martell ran a two-day internal hackathon where every team member learned to build with it. He says people operating at this level represent 0.04 percent of users. Inside Claude Code, he identifies three building categories: loops, which are recurring server-side jobs that can communicate with other agents and external APIs; tools, which are one-off builds for a specific project; and apps, meaning real production software. His house manager Betty, who is not a programmer, used Claude Code to build a system managing his personal life across cars, real estate, investments, and budgets. Martell’s line on that: ‘In the world of AI, English is the new programming language.’
Before writing any code, Martell always runs plan mode first, triggered by typing forward-slash followed by ‘pl.’ The feature takes in the idea, asks clarifying questions, and produces a complete plan for approval before any code is written. He attributes most complaints about AI app costs to skipping this step.
Level six is agent orchestration. This is where Claude stops being a tool and becomes infrastructure. Martell’s own system centers on a main orchestrator agent he named Kai. Kai does not execute tasks directly. He directs a set of specialized sub-agents, each of which owns a specific workflow. One sub-agent named Reese monitors real estate investment opportunities. Kai checks in with Reese daily and reports back. Martell connects Telegram so he can communicate with the agent system through his phone. He uses his own platform called Apex, available at apex.host, which he says makes the agent setup more secure and easier to interact with. Behind Apex, he notes, Claude is the main agent model.
One addition he flags for orchestration is a dedicated critique agent. Before any output is finalized, the main agent sends the work to the critique agent, which returns a list of notes. The main agent runs the task again using those notes. The loop continues until the output meets the standard.
The moment Betty’s system covered everything
Betty, Martell’s house manager, built a tool in Claude Code that tracks cars, real estate, investments, and budgets across every aspect of his personal life. She is not a programmer. The system runs in production.
Back on that mountain bike trail, Martell had his laptop elsewhere and his phone in reach. The berm cost him a washed-out tire. The code kept running anyway, which is the point he was trying to make before the fall interrupted it.
Source: Watch original
This article was reported in June 2026.
OHN Editorial Note: This article is based on publicly available sources. If you spot an error or have updated information, contact us at editorial@onlyhappynews.com. We correct mistakes promptly.
