Tech • AI • Robotics • Game
The AI iceberg: 6 levels, and most stop at the surface
5/10TL;DR
A six-level framework for AI use argues that productivity rises sharply when users move beyond simple chat and build structured workflows with project memory, connected tools, reusable skills and parallel agents.
KEY POINTS
From chat to operational use
The lowest level is basic chat-mode prompting, treated as a severe underuse of advanced AI systems. The framework compares it to driving a high-performance car in first gear: useful for questions, but far from the system’s potential to execute work such as drafting, planning or even assisting in media production.
Level 1 starts in code mode
The first upgrade is to work in a code interface tied to specific folders and projects rather than a generic conversation window. Recommended settings favor auto permissions for safer automation or bypass permissions for faster execution, while manual approval is discouraged because it slows down repetitive work and prevents the model from acting fluidly.
Project separation matters
A central rule is one task, one conversation. Separate folders and fresh chats reduce context pollution, improve output quality and help the model stay focused on a single deliverable, whether that is a quote, an image, a website draft or client work.
Context should stay lean
Large context windows are presented as powerful but easy to misuse. A working recommendation is to stay near 250,000 tokens rather than push toward 1 million, because bloated context can make sessions harder to manage and waste capacity on stale information.
Use AI before doing the task manually
The framework urges users to ask the AI first for nearly every task, from rewriting emails to generating ideas, preparing presentations or producing visual assets. Even image work can sometimes be generated through structured instructions and prior examples rather than relying only on heavy image-generation workflows.
Prompting improves through “grilling”
A key practice is an iterative questioning loop called grilling, where the AI asks clarifying questions before executing a complex or fuzzy request. The goal is to align the user’s intent with the model’s internal assumptions, reducing the common complaint that AI outputs are poor when the real issue is vague instructions.
Level 2 is a persistent project brain
The next step is to give the AI durable memory through a structured workspace containing files such as agent.md, about, roadmap, journal notes and decision logs. This lets the system remember who the user is, how they work, what tone to use, what constraints apply and where each project stands without requiring constant re-explanation.
Subscriptions should scale with actual limits
Spending is framed pragmatically. Users are advised to begin with the $20 plan, move to $100 only when they regularly hit usage limits, and then to $200 if the higher tier also proves insufficient.
Level 3 connects external tools
AI becomes far more useful when linked to operational systems such as Gmail, calendars, Notion and Stripe. With those connections, the model can move from drafting suggestions to taking real actions, such as handling customer messages, checking prior cases, issuing a refund and revoking access after approval.
Level 4 turns repeated work into skills
Recurrent processes should be saved as reusable skills: quote generation, weekly reviews, follow-up emails, accounting routines, writing styles or publishing checks. The logic is simple: if a task repeats, formalize it once, improve it when errors appear and reuse it through a single command.
Level 5 is parallelization with agents
The most advanced tier relies on sub-agents and multiple simultaneous sessions. Instead of running one long process, users can split research or production into parallel tracks, such as dividing a large slide deck into five parts or running five to ten sessions at once, then consolidating results into a single review stream.
CONCLUSION
The framework treats effective AI use as a progression from asking questions to orchestrating a structured digital workforce. The main competitive advantage comes not from the model alone, but from disciplined project setup, reusable processes and the ability to let multiple agents work in parallel.
Ask a question