The BourneForgeAI Principles for Reliable AI Practice
By Mark Bourne
A Practical Example
Consider a technical article similar to the one you are reading.
A simplistic workflow might be:
Prompt
Write a 5,000-word article about AI prompting.
Output
Publish.
A more mature workflow looks very different.
Stage 1 — Objective
Define the article’s argument, audience, purpose, and scope.
Stage 2 — Architecture
Design the major sections and narrative progression.
Stage 3 — Drafting
Write sections individually so each receives sufficient attention.
Stage 4 — Structural Review
Determine whether the sections build logically upon one another.
Stage 5 — Technical Review
Check factual and conceptual claims.
Stage 6 — Critical Review
Search deliberately for weak arguments, repetition, unsupported conclusions, and missing counterarguments.
Stage 7 — Editorial Pass
Improve readability, transitions, consistency, and tone.
Stage 8 — Publication Review
Verify headings, links, formatting, metadata, illustrations, and presentation.
The second approach requires more thought. But it also produces something fundamentally different.
It does not merely generate content.
It creates a controlled production process.
The Same Principle Applies to Software
Consider AI-assisted software development.
An immature workflow:
Build this application.
A more mature workflow:
- 1Requirements
- 2Architecture
- 3Risk analysis
- 4Implementation plan
- 5Code generation
- 6Automated tests
- 7Code review
- 8Security review
- 9Human approval
- 10Deployment
- 11Monitoring
- 12Feedback
The AI may participate in almost every stage. But responsibility does not disappear.
Instead, AI capability is placed inside an engineering structure designed to make that capability useful.
The Same Principle Applies to Research
Research can follow the same pattern.
- 1Question
- 2Scope
- 3Source collection
- 4Evidence extraction
- 5Contradiction analysis
- 6Synthesis
- 7Fact verification
- 8Human interpretation
- 9Publication
Again, no extraordinary prompt is required.
What matters is the architecture of the process.
The Emerging Discipline of AI Systems Engineering
This brings us to the broader conclusion of this article.
The AI industry currently talks extensively about models. That is understandable. Models are advancing rapidly and their capabilities are remarkable.
But production systems have never consisted solely of their most interesting component.
A web application is not merely its database.
A network is not merely its routers.
A cloud platform is not merely its servers.
And an AI system is not merely its model.
Reliable AI systems increasingly involve:
- models,
- prompts,
- context,
- retrieval,
- tools,
- permissions,
- workflows,
- evaluation,
- monitoring,
- security,
- humans.
The engineering challenge lies in how those components interact.
That is why the conversation is moving beyond prompt engineering.
The emerging discipline is AI systems engineering.
What Decades of Internet Engineering Can Teach AI
The Internet provides an important historical lesson.
It did not become dependable because individual components became perfect.
Routers still fail.
Servers still crash.
Packets still disappear.
Software still contains bugs.
Networks still experience congestion.
Security incidents still occur.
What made the Internet extraordinarily resilient was architecture.
- Redundancy.
- Protocols.
- Layering.
- Monitoring.
- Failure isolation.
- Standards.
- Fallback mechanisms.
- Distributed control.
- Continuous operational improvement.
The Internet works not because failure disappeared, but because engineers learned how to build systems that continue operating despite failure.
AI engineering should learn the same lesson.
We should not ask:
How do we make the model infallible?
We should ask:
How do we build systems that remain useful when the model inevitably makes mistakes?
That is a much more realistic engineering objective.
Beyond the Hype Cycle
Artificial intelligence is currently moving through an extraordinary period of experimentation.
New models appear. Benchmarks improve. Context windows expand. Agents become more capable. Tools become increasingly integrated.
Some technologies will prove transformational. Others will disappear. That is normal.
The Internet experienced the same cycle. Technologies arrived with enormous expectations. Some became foundational. Others became historical footnotes.
The enduring organisations were rarely those that chased every new technology. They were the ones capable of evaluating new technologies intelligently.
AI should be approached the same way.
Experiment enthusiastically.
Evaluate critically.
Adopt selectively.
Engineer carefully.
The Question That Matters
When comparing two prompts, ask:
What actually changed the result?
When comparing two models, ask:
What actually changed the result?
When introducing an agent, ask:
What actually changed the result?
When redesigning a workflow, ask:
What actually changed the result?
And then ask the more difficult questions:
- Was the improvement repeatable?
- Was it measurable?
- Did reliability improve?
- Did risk increase?
- Did human effort decrease?
- Did complexity increase?
- Was the trade-off worthwhile?
Those questions turn experimentation into knowledge.
Final Thoughts
The early era of generative AI has understandably focused on prompts.
Prompts gave ordinary users their first direct interface to extraordinarily capable models. Learning to communicate effectively with those models remains useful.
But prompts are only the beginning.
The deeper opportunity lies in understanding how AI fits into systems.
Systems containing people.
Systems containing information.
Systems containing tools.
Systems containing controls.
Systems that occasionally fail.
Systems that must nevertheless produce useful, dependable outcomes.
The practitioners who thrive in this environment will not necessarily be those who memorise the largest collections of prompts.
They will be those who learn how to:
experiment, measure, architect, verify, and improve.
That is the transition from prompt engineering to workflow engineering. And ultimately from workflow engineering to AI systems engineering.
The technology will continue changing. The models will become more capable. Today’s techniques will eventually be replaced.
But the fundamental engineering question will remain remarkably familiar:
Does this system solve the problem reliably?
That is the standard by which practical AI should ultimately be judged.
And that is the standard that should guide whatever comes next.
About This Book
This page is part of Prompt and Workflow Experiments: What Actually Improves AI Results?, an eight-chapter deep dive on reliable AI engineering.
More from the Notes
Short technical notes and observations, written up as experiments produce something worth documenting.
Was this useful?
Published