Finally upgraded to Wappler 7 as it is now considered "stable", even though I have had it crash on me twice and bugs with the panels. Anyway, onto the AI assistant experience. I signed up to Github co pilot (the top tier plan) and this was my experience. To test it out, I created a blank page and gave a simple request: Create me a login page. What happened next? It created me a login page along with CSS styling. However, it overrode my entire style.css custom stylsheet that had over 1000 lines of css custom styling. Luckily I have backed up my data but this is absolutely insane. You guys need warn users of these things. Then, to make matters worse, I asked it if it can reverse any requests. the response was as follows: 400 prompt token count of 103488 exceeds the limit of 90000. Now keep in mind I have only given it 2 requests on a top plan paying 40USD per month so clearly there are still major issues with this.
I can only imagine the major head ache this would cause me if I asked it to do things with my DB tables..
Role-based access control with three distinct user levels
Digital certificate management with PKI infrastructure
Document workflow engine with approval chains
Real-time progress tracking and notifications
Comprehensive audit trails and compliance reporting
Mobile-responsive design for accessibility across devices
Yes, AI can get confused and run wild. My experience is, ask questions first and when satisfied with the suggestions, instruct AI to go ahead step by step.
When you do not supply boundaries in the instruction, as said, AI runs wild and you will be confronted with 400 prompt token count error messages. Using my method, I am able to work all day every day without these messages.
The project is still a work-in-progress; hence the above lists are far from complete.
With all due respect Ben, I have been using ChatGPT separately for coding assistance with Wappler and it's been amazing. But within the first 30 seconds of using Wapplers built in AI it bombed out on me. Luckily I resolved the issue by restoring my CSS file, but I am now extremely nervous in using it for other aspects, example on my DB. Don't get me wrong, I love Wappler and use it religiously, but users need to be very careful using a tool like the built in AI as it directly and very intrusively interacts with your codebase. Perhaps a well documented tutorial on the AI and what effects it has on your app is needed.
I recently tested the free Copilot and was quite impressed. When asked to create a dashboard SPA with a pie chart, it impressively integrated Chart.js without explicit instruction. However, the page became unusable after I requested multilingual functionality with UI and /language routing.
This experience suggests users must be cautious with prompts, understand the AI model's limitations, and employ a step-by-step prompting workflow rather than generic instructions.
I also discovered an 'undo and keep changes' option, but I'm unsure if it applies sitewide or only to the current page. I haven't used Wappler in a while due to other commitments and am excited to explore the new release.
I think the thing to remember is AI is still a child,it will make mistakes, think of it as a guide rather than a jedi master.
I confess i mainly use the external copilot in windows rather than the wappler version as wappler doesnt seem well trained on custom extensions which i spend most of my time doing.
As ben said, boundries are essential. Almost always i ask it to change something and say "change nothing else" or it can run riot.
I've had a number of those disasters. Had the backups so things were repaired fairly quickly. At this point I don't use Wappler AI. I use claude 4 and Gemini pro 2.5 inside github/copilot.
I have AI read the repository and generate the code and/or steps for implementation. This way I'm able to handle all code before it gets implemented. These conversations are stored providing reference documents for the future.
Overall, it's become a invaluable tool to the development process and I appreciate all the work put into AI by the Wappler team.
I think it would be a good idea to have a disclaimer when the AI Assistant is opened that Wappler is not responsible for errors generated by any LLM and a small warning to backup your files prior to proceeding wouldn't go a miss.
Do note that AI is not a magical being that performs wonders and can read your mind.
With the right questions and good models like Claude 3.7 and GPT 4.1, it performs well, but it can still make mistakes. There is a disclaimer that AI can make mistakes in the begin screen of the AI manger and also the local AI assistants per editor @Cheese
We supply instructions to the AI and specifically to not overwrite style.css and it follows those most of time, but again sometimes does not.
So always make sure you use git and review the changes before keeping them.
With the built-in Wappler specific knowledge, the AI can produce a lot of useful content and solutions that work great. But you always have to verify them and be on top of what is happening to prevent unwanted surprises.
Something has gone Pete Tong (wrong) with Claude recently* it seems to go off on one with too much reasoning with the Premium versions (if you drop the Modals with reasoning and use 3.5 it is not so bad). Where as GPT 4.1 seems to have become a lot more reliable.
*Despite comprehensive instruction sets and training data being provided.
I think it’s quite simple - after it has worked its magic, let it tell you what it has done and that you need to confirm or decline for it to apply. That way if it says something like “applied changes (or overidden etc) to style.css do you want to continue and apply the changes .. that way you know what it’s about to do and what files will be effected and in what capacity. I literally used it in the first 30 seconds of trialling new Wappler 7 with such a simple prompt and it overrode my entire style.css. That shouldn’t have ever happened guys
Btw the rest of Wappler 7 does look amazing so well done! I am very used to 6.8 and it’s taking a bit of time to adjust but I’m liking it so far. One thing I have noticed is it doesn’t feel as “snappy” and quick as 6.8 .. it feels more sluggish but does look nicer. Anyway just my 2 cents at this stage
Yes, as I mentioned in this feature request. I think it's inevitable that errors or 'misunderstandings' will occur and I wish it were easier to control (in relation to CSS in this case).