From Prompting to Systems: Using Agent Skills in Architectural AI Workflows
From Prompting to Systems: Using Agent Skills in Architectural AI Workflows
AI image and video tools are improving quickly, but the real value is not just in writing better prompts.
The real value is in building repeatable workflows.
For architectural visualisation, this matters because one good image is rarely enough. A project usually needs a coordinated set of outputs: site diagrams, exterior views, interiors, façade studies, sections, storyboards and video prompts. If every prompt is written manually from scratch, the design language quickly drifts.
The building changes. The material palette shifts. The camera logic becomes inconsistent. The output stops feeling like one project.
To solve this, I use Agent Skills .md files as part of my workflow.
The process is simple:
Your Prompt → Agent Skills → Generated Prompts → Video/Image Models → Video/Images
The .md file sits between the initial design intent and the final output. It acts like a reusable framework for how the AI agent should think, structure and generate prompts.
What the Agent Skills file does
An Agent Skills .md file is not just a long prompt.
It is closer to a creative specification.
It can define:
- project type
- design sequence
- visual style
- camera rules
- material palette
- negative constraints
- output format
- prompt structure
- video/image model requirements
This turns prompting from a one-off task into a repeatable system.
Instead of asking for “some nice images”, I can ask the agent to follow a defined architectural method, generate a structured prompt set, and maintain consistency across multiple outputs.
Example: The Verandah KL
I tested this workflow with a speculative apartment concept in Kampung Baru, Kuala Lumpur.
The project was called The Verandah KL.
The design intent was a boutique luxury apartment tower that responds to the contrast between Kampung Baru’s fine-grain urban fabric and the KL skyline.
The concept was not to create another generic glass tower.
The architectural language was defined around:
- shaded verandahs
- deep balconies
- bronze terracotta screens
- warm timber soffits
- pale stone surfaces
- tropical planting
- controlled 2-point perspectives
- black-and-white architectural linework
- coloured post-digital collage intervention
This became the project’s visual DNA.
Once that system was established, the agent could generate coordinated prompts for street views, podium spaces, apartment interiors, balconies, night shots and video sequences.
Why the system worked
The strongest result came from using a dual-style architectural language.
The existing context remained as crisp black-and-white linework. The proposed building and interiors were rendered in warm, material-rich post-digital collage.
This created a useful balance:
- the linework kept the image analytical
- the colour made the proposal feel inhabited
- the contrast clarified what was existing and what was proposed
- the style remained consistent across different shots
For interiors, the effect was particularly effective. The architectural shell stayed as a technical drawing, while the inhabited space became warm, atmospheric and spatial.
It felt like a drawing opening into a lived architectural experience.
Why this matters
Architectural image generation is not only about aesthetics.
It needs control.
A useful AI workflow should preserve:
- site logic
- design intent
- spatial hierarchy
- material consistency
- camera discipline
- narrative continuity
Agent Skills files help encode those requirements before image generation begins.
The designer still makes the key decisions, but the system reduces repetitive prompt writing and keeps the outputs aligned.
This is similar to how architects already work with briefs, drawing standards, material schedules and presentation rules.
The .md file simply extends that logic into generative media.
The practical benefit
This workflow allows me to move from:
one-off prompting
to:
structured visual production
The benefits are clear:
- faster prompt generation
- more consistent image sets
- reusable project logic
- easier image-to-video translation
- stronger design narrative
- less drift between outputs
- better control over architectural style
The important shift is this:
I am not just prompting for an image.
I am designing the process that generates the image.
For architectural visualisation, that is where AI becomes much more useful.