Method

Premise

AI is a general-purpose technology. It is changing how work gets done in every industry: the workflows people follow, the processes companies run, how productive a person or a team can be, and how software and machines are operated, in the digital world and the physical one. People who understand the full range of techniques will make better decisions about where to use it. This site describes that range, shows how each technique works, and will measure what each one costs.

Principles

Fit the outcome and working experience. Prefer simplicity among approaches that satisfy the requested automation, quality, and human effort. Levels describe autonomy; they do not rank which solution is best.
Measure before claiming. A page may say a level helps only when a committed result file shows it. A page marked Measured prints its result, the model it ran on and the class of that model; every other number on the site is an illustration and says so.
Name real things. Every technique is tied to the models, products and tools that use it. Names come from a dated registry with sources.
Show how it works. Concept pages include diagrams and practical examples. Stepped traces are illustrations unless explicitly backed by a recorded run; runnable examples let readers inspect the mechanics.
State the costs and the failures. Each page lists the cost in tokens and time, the common failure modes, and how to test for them.
Stay current and say how current. Every page and every name carries the date it was last checked. Anything not checked in 90 days is flagged.
Build understanding gradually. The worked examples connect concepts to concrete tasks, evidence, and review decisions. Concept pages offer “Use it” and “Build it” reading lanes.
Independent. No rankings, no sponsorship, no affiliate links. Where several companies make something, the page names more than one.
Open. MIT licensed. The taxonomy, the registry and the result files are served as data. Corrections are welcome.
Private data stays private. Examples, traces and test documents are synthetic.

How a level is defined

A new level starts where the answer to “who decides the next step” changes: nobody, you, your code, the model for one action, the model for every step, several models, the models including when to start. Levels 1 to 3 all leave the decisions with the person or the code; they differ in how much has to be built around the model.

How a technique gets a page

It is described in primary sources from at least two independent makers or papers. It is placed at the lowest level where it applies. If it fits no level, the taxonomy gets a dated amendment.

How a page is written

Fixed anatomy, word budgets, sources for every factual claim, one production line: sources, draft, example and trace, eval, figure, review.

How results are measured

One task, run at every level, on three classes of model spanning small local to a frontier API. Result files are what the charts are built from; a number with no file cannot render.

How names are checked

A registry entry needs a primary source, a checked date and a verified flag before a measured page can cite it. Entries unchecked for 90 days are flagged.

How the site changes

Dated amendments to the taxonomy, a changelog per page, a review queue worked oldest first.

What the registry is, and is not

The registry lists models, developer tools and AI-first products, and records which technique each one demonstrates, with a primary source and the date it was checked. That is the whole of it. It is not a buyer’s guide: it carries no prices, no seat costs, no setup effort and no measure of quality, it ranks nothing, and it does not cover the ordinary business software a person already runs that has since grown an AI feature. Use it to work out what a product is doing underneath, not to choose which one to pay for.

If you want to buy rather than build, four things are worth asking about any product in any category. This site cannot answer them for you, because every answer depends on the plan and is the seller’s to state, but a seller who cannot answer them has told you something.

Which level it operates at. Does it answer one request at a time, run a fixed workflow somebody wrote down, or decide its own next step? That is the question this site’s eight levels ask, and the answer sets both what the product can do without you and how many ways it can go wrong.
Who else holds the text. The model’s maker is rarely the only company that receives what you type: the product in front of it keeps its own copy under its own terms. Who else holds the text says whose pages to read, in what order, and what the intermediaries usually do not say.
Whether a person approves before anything is sent or changed. Ask whether an approval step exists, whether it is on by default, and what it actually covers: a message sent, a record changed, a payment made.
Whether you can get the work back out. The documents, the records and the history, exported in a format another program can read, and what happens to them when you stop paying.

Where the site is today

The site is published and still being improved. Technique, topic, and recipe pages are sourced explanations: they provide primary references, but the label is not a guarantee that every sentence is correct. They have no recorded model run and scored result file behind them yet. A page becomes measured when it has both. Every cost-and-latency strip and stepped trace is an illustration, labeled as such. The evals page describes the measurement process. Worked examples are scripted teaching cases, not recorded model runs or certifications.

What the site does not do

It does not rank vendors, predict the future, or give legal, financial or safety-critical advice.

This page is also served as Markdown at /method.md, so a reader's own agent can quote the principles rather than read the markup around them.