Code, agents, skills. You start turning your own judgment into something the machine can run without you in the room.
Building used to mean coding. Now it mostly means describing - clearly, honestly, with judgment about what should exist. At Level 4 you turn pieces of your own judgment into things the machine can run: an instruction file, a skill, a small tool. You don’t need to be interested in code. Only in building.
Start where the builder’s power meets a normal window: the desktop apps. Claude’s desktop app includes Cowork - an agent you point at a folder of your real files, where it reads, edits, drafts, and organizes alongside you (most of its heaviest use is business operations, not code). ChatGPT’s desktop app runs its agent mode in a virtual computer - browsing, filling, assembling. Both are the real thing, without a terminal in sight.
Beyond them sits the deeper end: harnesses that live in a terminal instead of a window. Claude Code is the one to know, and it’s bigger than its name suggests - built with developers in mind, but really a general harness for real work: research, writing, analysis, operations, building, anything you can describe clearly enough to delegate. This whole practice runs on it; it’s our own primary harness, coding or not. Editors like Cursor stay closer to the code-editing end of the same spectrum. This page won’t walk you through those. Its job is to show you the terrain so you know what’s up here and what it’s for; the official guides below are current and genuinely good.
CLAUDE CODE / CURSOR
Install a terminal harness and give it one real, small task - not a toy, and not necessarily code. Watch how it works before you trust it with something bigger.
ANY TOOL
Save your Bearing File as a CLAUDE.md if you haven’t already. This is the level where it starts actually paying for itself.
ANY TOOL
Write your first skill: a judgment call you make often, turned into an explicit instruction set. Start with something small enough to finish this week.
The same four pieces of work recur at every level - and they are your WILD Intelligence, working. The name on each card is the job; the capacity behind it is what the reps build. The framework itself ↗
A skill is your judgment made executable - which forces the Wisdom question: what actually is your judgment here? Extraction is the deepest specify rep on the ladder. One honest warning: the grunt work you delegate is often the same work that built the judgment. Keep some reps for yourself.
Some work you automate because only the output matters. Some you keep by hand because the doing builds you - the stairs beside the elevator. Knowing which is which is Intention at Level 4, and nobody can make the call but you.
Not a demo for the drawer: one skill or agent you use weekly. The Leadership rep is trusting your own artifact enough to run it - and owning what it does.
Models shift under your builds. When one updates, your artifacts need re-testing, and your error log tells you where to look first. Builds are gardens, not statues.
Official guides worth your hour: getting started with Claude Cowork ↗ · ChatGPT agent mode ↗ · Claude Code, the general-purpose harness ↗
Level 4 is also where a guided first build starts beating a solo one - the walkthroughs end and judgment carries the weight. If you want a person alongside for yours, begin a conversation ↗.
A skill is the first thing you make that someone else can run. That changes the unit of work: not “I got AI to do this once,” but “our team encoded a judgment, and everyone inherits it.” Three moves make it real at a small company, none needing an IT department. Keep the shared pieces in one place - a skills folder in a drive or repo the team can reach, a shared bearing file so every agent starts from the same facts about how you work, and a short registry naming each standing skill or agent with an owner and a scope. The hardware bar is low: at this level, one capable laptop per builder, the agent reading and editing your real files locally, nothing on a server yet.
The WILD move underneath it: sharing a skill shares a judgment, so the team has to agree on the judgment before they automate it - a Wisdom move that happens out loud, in a room, before any code. And the deskilling guard scales up here too. Delegate the reps that only produce output; keep the reps that build the people. The grunt work you hand off is often the same work that trained the judgment you are now trying to encode.
You’re ready for Level 5 when you’re building the same kind of thing on a schedule, and you’re tired of being the one who has to remember to kick it off.