LEAN DOCUMENTATION IS THE NEW CODE

Open the obra/superpowers repository on GitHub. Look at the language bar at the top. GitHub will tell you it is 61.6% Shell, with a sprinkle of JavaScript and a pinch of Python.

It is not. The repository is, in substance, a pile of Markdown files — process descriptions, skill manuals, agent playbooks. The few shell scripts are wiring; the value is in the prose. GitHub's detector is honest: it counts bytes by extension. But it is measuring the wrong thing, because the meaning of "code" has quietly shifted.

Markdown used to be documentation about the program. Now, when handed to an AI agent, it is the program. The agent does not run the Markdown the way a CPU runs assembly — it reads it, infers intent, and acts. If the prose is loose, the agent improvises. If the prose is precise, the agent executes.

Which means the differentiating skill of the AI era is not prompt-craft, and it is not picking the right model. It is writing instructions a non-human reader will obey without inventing the missing pieces. Writing them at the right grain. Writing them so that operations, risks, mitigations, hand-offs, ownership and edge cases are all named and accounted for.

This is not a new discipline. It is exactly the discipline that came out of post-war Japan, that Toyota turned into a manufacturing religion, and that the West rebadged as lean. Standardised work. Andon. Poka-yoke. The whole point of the Toyota Production System was to write down a process so unambiguously that any operator on the line, having read it once, could perform it identically on the thousandth car as on the first. No improvisation. No drift. No silent variation.

That intellectual heritage flowed into Six Sigma, into Business Process Management, into the work that consultancies have been doing on factory floors and in back offices for forty years: map the process, name the actors, describe the inputs and outputs, identify the failure modes, define the controls. The output of that work has always been documents. Until now, those documents were read by humans. From now on, they will increasingly be read by machines.

The companies that will get the most out of AI are not the ones with the cleverest models. They are the ones who already know how to write a process down in a way that does not leak ambiguity. They have spent decades arguing with operators about whether a step is in-scope or out-of-scope, whether a risk is mitigated or merely noted, whether a decision is owned by role A or role B. That muscle, transferred to AI documentation, produces agents that behave deterministically. Without it, AI hallucinates because the prompt did.

At DOOGG that is the muscle we have been training the whole time, and it is what we now offer for AI deployments. We do not write prompts; we write operating standards for AI the same way a lean engineer writes them for a production line. Same discipline. Same vocabulary. New reader.

GitHub will eventually update its detector. The deeper change is already underway: Markdown is the new code, lean is the new compiler, and the rarest engineer in the room is no longer the one who knows the framework — it is the one who can write the standard.