Prompt engineering
The instructions the system runs on are written, versioned, and tested — not improvised in a chat window.
AI SYSTEMS
(Beyond the demo)
AI is not a chatbot added to your product.
It is a controlled system of instructions, context, workflows, tools, measurements, and safety boundaries.
A useful prototype proves that a model can produce an answer. A production system must also prove that it receives the right information, uses the right tools, stays inside its permissions, handles failure, and can be measured over time.
(Why demos fail in production)
The model is only one part of the system.
Most failures happen around the model: missing context, excessive permissions, unmeasured behaviour, and no defined way to recover when the expected path breaks.
- 01Missing or incorrect business context
- 02Unclear or untested instructions
- 03No controls on tools and actions
- 04No evaluation against real cases
- 05No fallback when the model is uncertain
- 06Permissions broader than the task requires
- 07Weak logging and observability
- 08No human approval for consequential actions
(Conceptual control artefact)
A sample approval-gate flow.
Conceptual example, not a client deployment or performance claim.
- 01
Receive
Accept only the data and tools required for the task.
- 02
Evaluate
Check inputs, expected output, risk level, and confidence.
- 03
Approve
Route consequential or uncertain actions to a named human owner.
- 04
Execute & record
Run the permitted action and retain a reviewable decision trail.
(Eight control layers)
Direction. Execution. Control.
These are engineering controls inside one production methodology, not eight separate products. The depth of each layer depends on the job, risk, and operating environment.
01
Direction
The system works on the right problem, with the right information.
Meta prompting
Those instructions are stress-tested and refined systematically, so quality doesn't depend on one person's phrasing.
Context engineering
The system sees the right business data at the right moment — and nothing it shouldn't.
02
Execution
The system does real work, in steps that can be checked.
Loop engineering
Multi-step work is broken into verifiable steps, so failures surface mid-task instead of in the result.
Harness engineering
The AI operates inside a controlled environment with defined tools and permissions — it can only do what it is allowed to do.
AI system design
Models, data, and business logic are architected as one system — not a chatbot bolted onto the side.
03
Control
You can measure it, limit it, and trust it in production.
Evals
AI behaviour is measured against test cases before and after every change — you see the numbers, not a demo.
Guardrails
Hard limits on what the system can say, access, and do — enforced in code, not in a policy document.
(Production decisions)
Every control should match the workflow's value and risk.
Models and providers
Select for the task, privacy boundary, latency, reliability, and cost. A larger model is not automatically the right model.
Data and permissions
Define which data the system can retrieve, what it can remember, who can use it, and where sensitive information is allowed to go.
Human approval gates
Require a person to review high-impact decisions, external messages, financial actions, destructive changes, or uncertain results.
Tool restrictions
Give the system only the APIs and operations its job requires. Sensitive actions get narrow scopes, explicit checks, and audit trails.
Evaluation datasets
Test against representative, difficult, and failure-prone examples so quality is measured against the work the system will actually see.
Logging and monitoring
Record useful operational events without unnecessarily storing sensitive content, then monitor quality, latency, failures, and spend.
Failure and fallback
Decide what happens when context is missing, a tool fails, confidence is low, or a provider is unavailable before those events occur.
Cost and security boundaries
Set rate limits, budgets, data boundaries, retention rules, and escalation paths that match the value and risk of the workflow.
Not every client needs every model, orchestration layer, approval gate, or monitoring system. The System Audit identifies the minimum controls needed to make the specific workflow useful, observable, and safe.
Start with the workflow, not the model.
Show us the work, the information it depends on, and what can go wrong. We will help determine whether AI belongs in the system and what controls it needs.
Start with a System Audit