A 25-term precision vocabulary for engineers who design, debug, and govern AI agent systems. Each term names a specific failure mode and the architectural pattern that prevents it. The terms are not independent; they form a layered system, from the substrate (Control Surface) through constraint design, system architecture, and validation.
A five-level framework for assessing and advancing a practitioner's or organisation's prompt architecture capability. Use it to diagnose where you are and identify the specific practices needed to advance.
| Level | Name & Description | Key Practice |
|---|---|---|
| L1 | Ad Hoc Prompts are individual, untested, and unshared. Each practitioner improvises independently. There are no shared templates, no validation, and no record of what has worked or failed. |
Recognising that a prompt is a designed artifact, not a casual instruction. |
| L2 | Managed Prompts are documented and stored, with basic validation applied to high-stakes use cases. A shared library exists in some form. Individuals can reproduce their own successes, but knowledge does not transfer systematically. |
Writing a Ground Truth Contract before deploying any production prompt. |
| L3 | Defined Organisation-wide standards exist: a style guide, a template library, a peer review workflow. All production prompts are reviewed before deployment. The Three-Constraint Rule and Constraint Architecture are applied consistently. |
Mandatory peer review of all prompts before production deployment. |
| L4 | Quantified Prompt quality is measured with structured benchmarks. Bias is audited on a regular schedule. ROI from AI interactions is tracked against defined targets. The Validation Suite is applied to all high-stakes outputs. |
Running a full Validation Suite (contract → tests → red-team → judge) on every production prompt. |
| L5 | Optimising Continuous improvement via automated red-teaming, prompt curriculum updates, and cross-organisational learning. The organisation can adapt to new models without rebuilding from scratch. Cognitive Infrastructure is the operating mode, not the goal. |
Automated red-teaming running continuously against a live benchmark corpus. |
The framework is backed by a structured curriculum: 4 modules, 12 lesson clusters, each expanded into 5 PMM-tiered sub-lessons. 60 lesson artifacts. 60 hours of seat time. Every sub-lesson targets a specific competency transition, from Novice to Organisational.
| Module | Lessons | Levels covered | Seat time |
|---|---|---|---|
| Module 1 The Translation Layer |
1.1 Translation Metaphor · 1.2 Three-Constraint Rule · 1.3 Semantic Drift | L1 to L5 | 15 hours |
| Module 2 Constraint Architecture |
2.1 Intent Decomposition · 2.2 Context Window Management · 2.3 Error Handling | L1 to L5 | 15 hours |
| Module 3 System Design |
3.1 Prompt Chains and State · 3.2 Multi-Agent Orchestration · 3.3 Cross-Modal Translation | L1 to L5 | 15 hours |
| Module 4 Validation and Scale |
4.1 Validation Frameworks · 4.2 Bias Detection · 4.3 Deployment and Governance | L1 to L5 | 15 hours |
Levels 1 to 2 · 24 hours
Foundation
Novice to Proficient. Core vocabulary, Three-Constraint Rule, Fallback design, Prompt Chain Integrity. For Persona A and engineers new to constraint architecture.
Levels 3 to 4 · 24 hours
Advanced
Proficient to Expert. Multi-agent orchestration, automated validation, LLM-as-Judge pipelines, Red-Team Protocol. For engineers building production systems.
Levels 1 to 5 · 60 hours
Full
Novice to Organisational. All 60 sub-lessons including the Level 5 governance tier: Cognitive Infrastructure deployment, governance-as-code, organisational competency systems.
The Agent Control Architecture Pack includes 12 deployable system prompts, 3 AGENTS.md templates, and 5 fully-worked BYOP diagnostic rebuilds.
Get ACAP ($89) → Try the free Prompt Diagnostic