On this page
There is an interesting symmetry in the way determinIQ has been designed and the way I have developed it. The product uses AI where language and interpretation are useful, while keeping authority over analytical results elsewhere. During development I have followed much the same rule: AI can help me do substantially more work, but it does not decide what the product should be.
That distinction has been practical rather than philosophical. A plausible paragraph can be wrong. So can plausible code. Neither becomes correct simply because it arrived quickly or was expressed with confidence.
The speed is real
The acceleration from AI-assisted engineering is difficult to ignore. A task that would previously have required moving repeatedly between design, implementation, testing and documentation can now be developed through a much tighter cycle. I can bring an architectural question into focus, examine several implementation options and inspect a working change in far less time.
That does not mean there is no work. The work changes. There is more opportunity to explore, compare, question and improve, but also more material to assess. A rapidly produced implementation can contain incorrect assumptions, duplicate responsibilities, unsuitable interactions or a subtle regression elsewhere in the application.
The faster the output becomes, the more important it is to know what should be accepted.
Context helps, but it is not memory in the human sense
Over the course of determinIQ I have assembled a substantial working context: architectural rules, decisions, analytical terminology, governance boundaries, interface conventions, test expectations and current implementation details. This lets an AI assistant work with more of the product’s actual constraints instead of beginning each change from a generic example.
It is useful to be precise about what that means. I have not trained a separate model on determinIQ. The working context is a maintained body of instructions and source material that an assistant can consult during the development process. It needs to be corrected as the application changes. Outdated context can be worse than no context if it persuades the assistant to build against an architecture that no longer exists.
The current codebase therefore matters more than a confident recollection of an earlier design.
The acceptance test is outside the conversation
One of the habits I have found most useful is to demand that a proposed correction be followed through to a real observation. What failed? Where is the responsible authority? What evidence supports the proposed change? Did the code run? Did the interface behave correctly? What changed elsewhere?
For a WordPress application, that can mean reading the actual PHP, checking the JavaScript that invokes it, examining the generated HTML, reviewing logs and then returning to the running product. A screenshot can expose a problem that an apparently successful code change missed. A test can expose a mistaken assumption that a beautiful interface concealed.
The working loop is therefore closer to observe, trace, prove, change and test than to “prompt, generate, accept”. AI helps at several stages, but the resulting system still has to demonstrate that the work is correct.
Designing by removal
I have also found that rapid implementation creates its own temptation: another control, another option, another way of doing approximately the same thing. It is easy to accumulate features faster than they can be made coherent.
Some of the most valuable later changes to determinIQ have involved removing duplicate routes, consolidating navigation, clarifying status and giving a user a predictable way back to the question they were investigating. That is not a lesser form of engineering than adding capability. It requires understanding which parts of a product genuinely belong together and which only add another decision for the user.
A technically successful feature may still be the wrong feature. An apparently minor simplification may improve the whole experience.
A different meaning of being hands-on
AI-assisted development has not removed the architectural work that I value. I still need to decide what problem is worth solving, where responsibility belongs, which compromise is acceptable and when a result is good enough to use. What has changed is the amount of practical implementation I can examine while making those decisions.
The October 2026 determinIQ research paper describes this as human-led product development with AI as a force multiplier. That is an account of how I have worked, not a controlled productivity study or proof that the same approach will work equally well in every team.
I would not suggest that experience alone guarantees correctness, either. Experience helps me recognise risks and ask better questions, but the code and the running system can still contradict my expectations. The useful response is to change the implementation or reconsider the assumption, not to defend the original idea because it was mine.
What I want to preserve
The most valuable part of the acceleration is that it makes more investigation possible. I can go further into an architectural idea, follow a difficult edge case or test a different interaction without committing the same amount of time that this would once have demanded.
But that only has value if the quality of the judgement survives the speed. My aim is not to maximise generated code or the number of features delivered. It is to make useful things, understand how they behave and retain the ability to explain the decisions behind them.
There is a direct connection to determinIQ’s analytical philosophy: use AI for leverage, but keep evidence and authority somewhere that can be checked. The same rule has proved useful on both sides of the product.
Connected research
The flagship research programme sets out the original analytical question. The analytical-authority study examines how the application divides interpretation from calculation. This journal entry describes the parallel discipline in developing the application itself, where implementation speed does not replace review, verification or product judgement.
Research context
Research programme: From intention to implementation: human-directed AI-assisted engineering
Topic: AI-assisted engineering
Explore further
Supporting material is listed below.
Further reading is listed below.