Skip to content
AI Architect Academy

A cohort course, built in the open

For the engineer who was just handed the AI work.

The docs teach the APIs. Nobody teaches which decision comes first, which one you cannot undo, and what getting it wrong costs. That is the course.

You finish able to write the six decision records for a system you own, and defend them in a review you did not call.

It is not scheduled and it has no price. The people on the waitlist decide both.

Join the waitlistRead the free ADR first

What exists today

The material is public. The cohort is what is not.

Everything the course is built from is in an open repository you can read now: 16 design patterns, 3 hands-on labs, and 46 further modules, paths and governance documents. The counts are of the items listed on this site, read from the repository on 2026-09-10, not typed in.

The repository is the reference. The cohort is the part a document cannot do: your system, your constraints, and your decision records read by someone who will tell you which one you will regret before you ship it.

Design patterns
16
Hands-on labs
3
Micro-modules
21
Learning paths
9

See every item, with a link to the source

The shape of the cohort

Six decisions, in the order you have to make them.

A working outline, not a syllabus. It changes with what the waitlist tells us. Nothing below is a promise of a module.

  1. 01

    Scope the first system to one job

    With a failure you can measure before you build. Most AI projects die of ambition, not of models.

  2. 02

    Choose retrieval, fine-tuning, or neither, with a test

    Run it before anyone has an opinion. Then write the record that says why.

  3. 03

    Draw the agent boundary before production draws it for you

    What the model decides, what code decides, what a person decides.

  4. 04

    Write the eval before the feature

    The one you write first, the one you write when something breaks, and the one that stops a release.

  5. 05

    Budget the failure modes

    Latency, tokens, wrong answers, and who pays for each. Budgets, not hopes.

  6. 06

    Run governance that survives a real organisation

    A decision record, a review, a kill criterion. Three documents, kept.

Free today, in your own repository

An architecture team you can run today.

The AI Architect plugin runs architecture work as gated stages inside the coding agent you already use, on your own model keys. Each stage writes its evidence into docs/architecture/ in your repository. A red gate stops the run. No account, no hosted service, Apache-2.0.

It never writes your application code, and it stops at the gates a person owns: publishing, spending, credentials, anything destructive. The plugin produces the records. The cohort is where a person reads yours and tells you which one you will regret.

  1. frame00-frame.mdThe outcome, the non-goals, and the observation that would kill the project
  2. discover01-discovery.mdWhat is known, and what is marked unknown instead of assumed
  3. flow02-user-flows.mdEvery path through the system, and where a person steps in
  4. decideSYSTEM.md · adr/The decisions that are expensive to undo, each marked MADE or OPEN with its evidence
  5. cost04-roi.mdThe return model, and the prices it assumes
  6. secure05-trust-boundary.mdWhat the model and its tools are allowed to reach
  7. prove06-evals/The eval cases and the grading rubric
  8. operate07-runbook.mdThe workflow, the SOP, and the runbook
  9. verifyreceipts/An independent verifier re-checks each evidence pointer in a fresh context

Claude Code

The full team, as /architect

/plugin marketplace add frankxai/ai-architect
/plugin install ai-architect@frankx

Codex, Cursor, Gemini CLI

The team's skills, for other harnesses

npx skills add frankxai/ai-architect

Any skill host

The architect skills pack

npx skills add frankxai/skills

Running an AI practice rather than one system? The AI CoE operating model governs the organisation around the team: intake, review, trust tiers, and the weekly review that decides which skills stay in use. Source for the team: github.com/frankxai/ai-architect.

Join the cohort waitlist

Before you give an address

One decision, worked all the way through.

The unit of the course is the architecture decision record: the context, the choice, the options you rejected and why, and what it costs when it turns out to be wrong. The template and one complete worked record are on this site, free, with no form in front of them. If they are not obviously worth more than the docs, do not join the list.

Read the worked ADRSee a full team run, red gates included

Who is teaching

Frank Riemer

I am an AI architect. I have designed AI systems for large organisations and I teach the way I was taught best: by pairing, inside the codebase, with a question before an answer. The open repository already works this way if you open it in a coding agent.

This academy is independent work. It is not affiliated with, endorsed by, or sponsored by any employer, vendor, or client, and nothing here describes any organisation I have worked for. Every example is a composite written for teaching.

Waitlist, not checkout

Decide what this becomes.

One email to join. Three optional questions after, and they set the price and the order of the course. One email when Cohort 1 is scheduled. Two emails from this list, and nothing from any other list unless you join it.

The founding fifty get

  • The launch price, whatever it is set at, held for you for as long as the course runs.
  • Named in the course materials.
  • A say in the order the six decisions are taught.
  • Every later revision of the material, without paying again.

The first fifty keep the launch price and are named in the materials.

Straight answers

Is there a price?
Not yet. The price band you answer on the waitlist is the evidence that sets it. Nothing here is discounted, because there is no price to discount from. The founding fifty keep whatever it becomes.
When does Cohort 1 start?
No date. It is scheduled when there are enough people to run it well, and you get one email when that happens. There is no countdown and there will not be one.
What would it actually be?
The working intention: live sessions on the six decisions, applied to a system you bring, with your decision records read and reviewed between sessions, in a group small enough that every record gets read. Length, hours and size are not fixed. The waitlist answers set them.
What do I get today?
The ADR template and a worked decision record, free, on this site. And the open repository, which is the material the cohort is built from.
Who is this not for?
People who want a model tutorial. The docs already do that well. This is for the person who has to decide, and be accountable for the decision, before they feel ready.