Is it possible to train AI tools like ChatGPT to build Ignition projects?
It can be useful for cleaning up existing code, but I wouldn't trust it much beyond that with Ignition. I've seen AI use functions that don't exist in example code that it spits out, and it wouldn't surprise me if it gets Vision and Perspective mixed up.
I would be cautious even with this. I find AI Ignition outputs are somewhat dirty both in code and comment.
Example:
Jython gives a developer access to Java and Python libraries, and Ignition has a long list of its own system functions that tend to play nicer with the gui and tend to clean up some of Jython's messy tendencies such as ASCII string outputs. Therefore, there are often at least three ways to get the same result in Ignition. My recommendation is to use Ignition system functions when available, and when not, use Java's methods, but the AI tools I've experimented with tend to take the opposite approach, often using obscure Python methods over Java, and they will almost never use anything provided by Ignition without intricate prompting.
I find Claude is quite good at it, even using system functions in ways I didn’t think of.
As long as it has some context and instructions/guidelines, it does a pretty good job.
Still gotta be careful, obviously, I wouldn’t use code without validating it first.
I think it could be used only when you're stuck, for example, but when you totally understand what you're doing, so you can check AI and understand if it generates correctly.
Indeed. Thanks to Claude, it’s looking like the next version of my Markdown Editor+ is going to have full-blown WYSIWYG (rich text editor) mode. It’s very good with JavaScript, HTML and CSS.
"Can you find my typo?" has worked pretty well for me when I've been at it for too long and, for whatever reason, just can't bring myself to call it a night.
I’ve used the chatgpt tailored for ignition a few times when I'm stuck on syntax. I’m very new to python, so sometimes the lists, vs tuples vs dictionaries, etc I'll convolute, and the chatgpt engine helps direct me a certain direction. But of course, I'll vet and test in the script console with plenty of prints to learn what exactly is going on. Just relying on AI without validation is probably not a good idea and you don’t learn going forward.
I ran into the same issue Justin is describing. The breakthrough for me was to stop treating the model like it “knows Ignition” and instead give it Ignition-specific skills.
By skills, I mean small context/instruction files that teach the agent the platform rules: Ignition 8.1 uses Jython 2.7.3, prefer native system.* functions, know the difference between Gateway scripts, Perspective events, tag events, transforms, etc., and don’t invent Perspective props or resource shapes.
I took it further and built a Web Dev API runner for a dev Ignition 8.1 Gateway so the AI agent can inspect the actual project instead of guessing. It can create/test Perspective pages, bindings, UDTs, tags, read values, check logs, dry-run changes, apply to a dev project, and validate what it did.
That process is actually how a lot of the skills improved: I let the agent work against a dev Gateway, build scripts and Perspective pages, hit errors, read the evidence, and update its own Ignition-specific instructions. I showed that workflow in a demo building Perspective pages with AI agents.
The skills I’ve built so far cover Perspective import zips, Perspective API workflows, button logging, Gateway log diagnostics, UDT creation, Jython scripting, SQL/Named Queries, HMI/SCADA styling, expressions, alarms, historian forensics, Perspective performance auditing, and the Web Dev API.
I’m building this as Ignition AI Skills. If anyone wants to see what I mean by “skills,” search “Ignition AI Skills” or go to IgnitionAISkills.com. The core idea is still simple: give the AI Ignition-specific rules, let it work only against a dev/staging Gateway, review the diff, and validate before trusting anything.
My thinking has changed on this since I wrote that and discovered how capable Claude is. I'd still say "trust, but verify," but I used Claude Code to create an entire project in the Exchange:
It's mostly JavaScript injected using the Markdown hack which is in danger of becoming extinct, but Claude understood Perspective enough to wire it up in a way that the data is accessible from within it.
You are a genius. May be you can coordinate between IA team and Claude team and explore the possibility of building a customized AI powered Ignition project builder. Thanks.
I don't know about that. As Claude himself wrote, "real genius is just old ideas getting it on".
This is something that I have been trying to build a module for. Hopefully it'll be done soon. It became a lot more work then I thought, but it has been getting much closer to a reality.
This is very close to what I’ve been working on independently, although my current focus is mainly Ignition 8.3 and Perspective.
I’m building an agent development harness that can inspect and edit actual project resources, run them against an isolated development Gateway, render and test Perspective views, inspect logs, compare screenshots and resource diffs, and iterate based on the evidence instead of assuming that the generated output is valid.
I’ve also reached the same conclusion about skills: the model itself does not need to “know Ignition” if the harness gives it the correct version-specific rules, tools, examples, and validation loop. to the point where i was building a Lovable.com but for ignition ,where u could extract components , views , even the project itself.
The end-to-end workflow is getting fairly close to being usable, so I’d be interested in comparing approaches. Your 8.1 skills and Web Dev runner seem potentially complementary to the 8.3 filesystem and visual-validation direction I’ve been working on.
Would you be open to a short call or demo exchange? It could be useful to compare where each approach is currently working, where it still fails, and whether there is an opportunity to collaborate rather than solving the same problems separately.
im runing some evals on gpt 5.6 sol right now as im writting this to see how much it improves over other models at doing perspective views with skill/without and with a design.md