Generative Specification · the open method

Write the spec. Let the machine derive the code. Check it from the outside.

AI assistants write code quickly, and every new session decides a thousand small things a little differently. Generative Specification is a discipline for writing down what a system must be, so that an assistant with no memory can derive it and a check outside the model can tell you whether it did. This site is where you learn it: free, open, and plain about what has and has not been tested.

Start with your first hour The free course See the evidence

Choose a way in

I'm new

Get the idea in two pages, then try it on a project of your own in one hour: a spec, a sentinel that routes the assistant, a gate that stops the build.

Two-page introductionYour first hour

I lead a team

Grade a project on seven properties, use checks that refuse a bad build, and decide what to enforce first.

The rubricQuality gatesLifecycle and debt

I want to see the evidence

Experiments, papers and equations, each with its status and its bound. The studies are small and written by the method's author, so they are published to be re-run.

Evidence and papersModels and equationsFindings and bounds

Three ways to reach the same lessons

By goal or roleWhat to read for your situation By course sequenceThirteen short lessons, in order By propertyOne of the seven ideas at a time
The free course Soon

Generative Specification in about 26 minutes: thirteen short lessons, English and Spanish.

One idea per lesson, from a spacecraft that failed because nobody verified a contract to the first hour on your own project. The videos are finished but not yet published. The lesson list is final, and no video is linked until it is live. Meanwhile, the same ideas are written up: start with your first hour.

See the 13 lessons

The core in five lines

  1. Derivability. The reader of the code is stateless. If the correctness of a program can’t be ratified from its spec, contracts and trace, it can’t be governed.
  2. The substrate. The spec (a contract root plus a cascade of derived documents), the sentinel that routes the assistant to them, hooks and gates, and verification. This is what you build. The harness is only the verification and enforcement layer of it.
  3. Three load-bearing mechanisms. The bridge between human language and code (with its read/write asymmetry). The sentinel, which bounds context. Phase collapse, where spec, design and code stop being separate phases.
  4. The rubric. Seven properties that make a project derivable. An instrument for measuring the substrate, not a mechanism.
  5. An honest bound. The value of the scaffolding is capacity-relative: highest where the model is weak, receding as it strengthens. What endures is the guarantee kept outside the model.

Everything else, one click away

Section What is there
The method The rubric, spec completeness, quality gates, the gate library, the working principles
Practice Paste-and-run guides: new project, existing project, join a codebase, migrate a stack
Models and equations Four short models and equations, each with the status it has earned
Recipes Step-by-step workflows for common project scenarios
Domain guides Fintech, machine learning, games, creative work, command-line tools
Evidence and papers The papers, the experiments, the raw evidence. For readers who want to check the work
Essays and ideas Longer reading on why the method exists
About Licence, citation, how to contribute, contact

Using this in a company? Governance, assurance and due diligence built on this method live at pragmaworks.dev. Here is the method itself, open.