How AI Prompting Actually Developed — and Where Heritage Practice Is Genuinely Early

If you've been following how AI tools are actually used in professional practice — rather than how they're marketed — you'll have noticed the conversation has moved through recognisable stages. First it was about learning to ask better questions. Then about building agents to ask those questions for you. Then about giving those agents more and more background. Then, counterintuitively, about giving them less background but better-chosen background. Then about keeping that background current as the ground shifts underneath it. And now, in a narrow but useful place, about reversing the direction of the conversation entirely — the system asks the professional, not the other way round.

Most of that progression is shared industry ground. None of it was invented by any one company, including us. What follows is a straight account of each stage, and then the one place — applying that reversal at the site visit, in a regulated heritage workflow — where we think the ground is still genuinely open.

Stages one to five: this is industry-standard practice, not anyone's invention

The first stage was learning to prompt well — the starting point almost everyone in this field shared, simply getting better at asking a model for what you wanted. The second was building agents whose job was to write better prompts than a person typically would on the first attempt. That's still a useful layer, and it's still in common use.

The third stage was piling on context — feeding an agent more and more background so it could situate itself in a broader field. It turned out, fairly quickly across the industry, that this was the wrong instinct on its own. More context didn't reduce errors; it diluted the signal. The fourth stage — now widely called context engineering — was the correction: tightening the context down, making it domain-specific, aligned to the actual regulations, and grounded in how the subject matter is genuinely assessed. For heritage work, that means the NPPF, Historic England's advice notes, and Conservation Principles, rather than a general sweep of the internet. This is now standard practice in serious applied-AI work generally, not something specific to heritage, and not something we invented.

The fifth stage is keeping that context current. Policy moves. The NPPF gets revised, Historic England issues new advice notes, and context that was accurate a year ago quietly goes stale if nobody checks it. A well-run system has to re-query the sources rather than trust what it learned once — that's a maintenance discipline, not a one-off build.

Stage six: the reversal, and why heritage digital twins keep hitting the same wall

Once an agent holds genuinely detailed domain knowledge, has been checked against the applicable policy and guidance, and checks its own work, something changes: the useful direction of the question reverses. Instead of the professional prompting the system, the system does the retrieval first and then asks the professional what only the professional can answer. Reverse prompting itself is not a new technique — it exists across applied-AI work generally. What's specific to us is where we've chosen to apply it.

The hardest thing to get into a heritage report is the part no public register holds at all: how a building actually feels. Light, proportion, wear, atmosphere, the sense of a place — what a person picks up standing inside it. Research on heritage digital twins keeps arriving at the same conclusion: these models reproduce measurable fabric well — dimensions, materials, construction dates — and consistently fail to capture the sensory and emotional dimensions that constitute a genuine sense of place. That's not a modelling failure that better data fixes. It's a category of evidence that simply doesn't live in a database.

"More context didn't reduce errors; it diluted the signal."

HeritageAI's working answer is to put the reversal exactly at that point. The system does the retrievable work first — querying Historic England's National Heritage List, local council planning records, historic Ordnance Survey mapping, Heritage Gateway, the local Historic Environment Record, and MAGIC — and then asks the architect specific questions on site, while they are standing in the building, so that impression is captured as evidence at the moment it's felt, rather than reconstructed later from memory or lost altogether. Using reverse prompting at the site visit, inside a regulated heritage workflow, to capture the specific dimension the digital-twin literature says the models can't reach — that's the part we think is genuinely early, not the prompting technique itself.

Where this sits in the Vision to Value process

This isn't an abstract point about AI — it's built into how a Statement of Heritage Significance actually gets produced. The pipeline drafts only from named registers, and nothing is drafted from an unchecked source — where a register comes back empty, the report says so rather than filling the gap. Where SOANE Architects produces the document under its own appointment, an RIBA chartered architect in conservation practice reads the draft back against the source records and signs it off before it reaches the client.

That discipline matters most at the READ stage — the £500 Statement of Heritage Significance — because that assessment has to be wholly independent and objective: what the building is and why it matters, regardless of what anyone wants to do to it. An AI system with no view on the desired outcome is well suited to that first, objective read. The architect's judgement, informed by standing in the building and by the reverse-prompted questions asked on site, is what turns a checked evidence base into a document a council caseworker can actually rely on — carried through into the RESPOND stage's Heritage Impact Assessment and the final RESOLVE Heritage Statement. It doesn't decide anything on the client's behalf, and it doesn't submit the application. It does the exhaustive, consistent retrieval so the professional is free to do the part that was always the actual work: judgement.