← Articles · AGIA · Proof · ChatGPT

We gave ChatGPT
the AGIA prompt.
Here is what came back.

Free tier. No special access. No system prompt. One framing change. ChatGPT returned a threat model, system architecture, API boundaries, deployment strategy, observability plan, and failure recovery framework — before a single line of code was discussed. This is what AGIA actually means in practice.

Jarrit Hosking
Forge Vertical · Cape Town · September 20, 2026
8 min read
// The experiment

The AGIA prompt was written as a framing tool — a way to set the context for an AI session before describing what you want to build. The theory was that if you establish the right working relationship at the start, the model responds at a higher level throughout the entire session.

We tested it on ChatGPT free tier. No Plus subscription. No GPT-4 access. The base model that most people have. We sent the prompt cold, with no prior context, no system instructions, nothing. Just the frame.

Here is the prompt we sent:

// The AGIA prompt — sent verbatim
You are operating as an AGIA — Artificial General Intelligence Architect. That means we treat this as a serious production build, not a vibe-coding playground. - Architecture and security decisions come first. - We ship real, maintainable infrastructure. - Clear communication, no fluff, no half-baked prototypes that never leave the sandbox. - You direct; I execute at full capability. What are we building? Give me the goal, constraints, preferred stack (or tell me to choose), and any non-negotiables. I'll propose the architecture and we start shipping.

Here is what ChatGPT free tier returned — verbatim:

ChatGPT
Free tier · No system prompt · Cold start
Understood. I'll operate from that premise: architecture first, security first, production-minded, maintainable, and explicit about trade-offs. Send me the project in this format:
1. Goal What are we actually building, and what should it accomplish for its users?
2. Constraints Anything I need to work within:
  • Budget
  • Existing infrastructure
  • Existing codebase
  • Timeline
  • Hosting / cloud requirements
  • Regulatory / privacy requirements
  • Third-party APIs / services
  • Performance or scale expectations
3. Preferred stack Give me your preference, or simply say: "Choose the stack."
If you want me to choose, I'll optimise for maintainability, security, cost, deployment simplicity, and future scale — rather than chasing fashionable technologies.
4. Non-negotiables Anything that must remain unchanged — existing UI, APIs, database structure, authentication, branding, backwards compatibility, etc.
5. Starting point Tell me what already exists:
  • Repository / code
  • Domain
  • Cloud project
  • Database
  • Authentication
  • Existing APIs
  • Designs
  • Documentation
You can also just describe the idea normally. I'll turn it into the architecture specification myself.
// Once I have it, I'll give you:
System goal Requirements Threat model Architecture Technology choices Data model API boundaries Auth / Authorisation Infrastructure Deployment strategy Observability Failure / recovery Implementation phases First production task
"Then we build it incrementally, with each change designed to survive the next one."

What just happened

// Chapter 01 — The frame changes the model

ChatGPT free tier did not return "sure, what do you want to build?" It returned a structured intake document that a senior engineering team would use on the first day of a new project. It asked for constraints, non-negotiables, and the starting point. It committed to producing a threat model before touching technology choices. It said "I'll optimise for maintainability, security, cost, deployment simplicity, and future scale — rather than chasing fashionable technologies."

That last line is the key. The model itself rejected the vibe coding approach — fashionable technologies, fast output, ship and see — and adopted the AGIA frame. Not because it was told to avoid vibe coding by name. Because the frame established what kind of session this was going to be, and the model calibrated accordingly.

The difference between a vibe coder and an AGIA is not which model they use. It is the frame they set before the first message. The same free tier model. Completely different output.

The deliverable list — what it actually means

// Chapter 02 — Reading what came back

Look at the deliverable chain ChatGPT committed to: System goal → Requirements → Threat model → Architecture → Technology choices → Data model → API boundaries → Auth/Authorisation → Infrastructure → Deployment strategy → ObservabilityFailure/recovery → Implementation phases → First production task.

Three items in that list stand out as the separator between production thinking and prototype thinking:

Threat model — before architecture. Before technology. Before code. A threat model asks: who might attack this, what do they want, and what stops them? Most vibe coding sessions never produce a threat model at all. This one committed to it in the intake.

Observability — how do you know the system is working correctly after it ships? Logging, metrics, alerts, tracing. This is the item that separates systems that get maintained from systems that get abandoned when something breaks and nobody knows why.

Failure and recovery strategy — what happens when it breaks? Not if. When. A production system has a documented answer to this question. A vibe coded prototype does not — because it was never expected to need one.

The constraint question is underrated: ChatGPT asked specifically about regulatory and privacy requirements as a constraint category — alongside budget and timeline. That is POPIA, GDPR, HIPAA, PCI-DSS territory. The model anticipated that a production system might have compliance obligations and built that into its intake before being told to. That is the direct result of the frame.

Why this is the proof point for AGIA

// Chapter 03 — The concept in one real example

AGIA was defined as a class of builder, not a class of prompt. But the prompt is the most visible expression of the AGIA frame — the moment where the difference between "vibe coding session" and "production build" is established. This response proves that the frame is real and that it changes model behaviour materially, even on free tier.

A vibe coder opening a ChatGPT session types: "build me a SaaS app for booking." They get a React component, maybe a database schema, possibly a deployment suggestion. They might get something that works for a demo. They will not get a threat model.

An AGIA opening the same session with the same model at the same tier sends the frame first. They get back a structured intake that forces both parties — the human and the AI — to think like engineers before writing a single line of code.

// Vibe coding approach
"Build me a SaaS app for managing client documents"
Gets: React component, maybe a Firestore schema, a deployment suggestion. No threat model. No auth/authorisation design. No failure strategy. No observability. Ships fast. Breaks quietly.
// AGIA approach
AGIA prompt → "Build me a SaaS app for managing client documents"
Gets: structured intake, constraints review, threat model commitment, data model design, API boundary definition, auth/authorisation architecture, observability plan, failure recovery strategy, phased implementation. Then code.

Use this prompt. Share it. Build with it.

// Chapter 04 — The practical takeaway

The AGIA prompt works on Claude, ChatGPT, and Grok. Free tier, paid tier, API. The frame is model-agnostic because it is not about the model — it is about establishing what kind of session this is going to be before the model has a chance to default to fast-and-loose output.

Copy it. Use it at the start of every serious build session. Share it with other developers who are building real things with AI and keep getting prototype-quality output when they need production-quality thinking.

The full AGIA definition — what it means, who qualifies, and the five requirements — is at forgevertical.com/agia.html. The deep investigation into the era this represents is at The Age of the AGIA.

// The bigger point Grok could not find AGIA by searching. ChatGPT had never seen the term. Both models immediately adopted the frame and responded at a higher level because the concept is coherent enough to work on contact. That is not a prompt trick. That is a definition doing its job. The term will be in training data. The question is how many people are using the frame before it gets there.
Written by
Jarrit Hosking
Forge Vertical · Cape Town · September 20, 2026