I am currently working on a Flutter application, and I would like to get the best tips so that the model can help me build it correctly and effectively. Are there any specific instructions I should give it, such as reviewing the official documentation of the technology, or something else?
Work as if you are a manager, not a programer, thats what i learned in the past weeks, have your tool work and think for you, what i mean is have it create its own test, give it guidance, dont tell it what to do unless necessary, brainstorm and plan with him and then once the plan seem good to you let him go, ensure he properly test his implementation.
Im using python myself and there is a nice testing module called pytest wich i just tell my agent to increasae the test coverage and then it test almost everything in my codebase and fix broken stuff by himself.
Just keep an eye as sometime they need baby sitting, just like real employee mahaha!!
Also do NOT allow them to simplify their work or you will end up with agent pluging a return true on his failing test to make it pass, yeah i saw that… mocking values just to have his test past without touching the proper application logic!
Refine, repeat, create workflow with him, when i want to include something smallish in my codebase i use the current prompt:
Your work will be to update the different phases in the integration_plan.
-Analyse task scope
-Analyse the code
-What do we have
-What do we needed
-How to accomplish
-Write temporaryTODO file
Execute the task
After the task is complete
-Validate against the spirit of the task
-Comprehensive testing
-Cleanup
-UPDATE THE PLAN/TODO
Or variation of this. In rules and even during testing i tell him he his NOT allowed to create file editing script, lost a whole codebase before i learned how to work with github, (yeah new programer here, Hiho!). He can create scripts that analyse, scan but every file editing must be done manually by him.
Right now i think its pretty much my tedtalk, thank you!
2 Likes
Thank you so much for sharing your detailed experience and workflow, it’s really helpful!
I have a couple of questions if you don’t mind:
-
Which AI model do you mainly use with this setup?
-
Do you work on Linux or Windows? I noticed that sometimes Windows has issues with PowerShell when running agents and scripts.
Thanks again for the insights, I really appreciate it!
To be honest? Ive been bouncing between the free model offers, right now im on grok, i also use the marketplace to find “agent” publisher, right now im sporting cline with free-grok. I find if your prompt is structured enough and you dont mind the few API error you can work with pretty much any model wich has a few brain cells, sometime it need more baby sitting but so far its something i can afford, and hey! working with dumb AI also build your prompt engineering skills for when you fall on more powerfull models so the experience is not lost. In fact i learned more when my AI was to dumb to fix something, or broke my codebase than using high end models. High-end isnt always the best, if i need high-end brainstorming i use my pro account with gpt-5 thinking, then code using dumb AI, only need to pay one place, GPT-5.
I can produce faster with grok-fast without the “deep” reasoning GPT-5 provide and if i have a deep issue, well i upload few file to GPT5, he trouble shoot it and i play with the dumbo of the day.
1 Like
Auto does a pretty good job. Claude 4 for architecture design and directory structure, I usually have a .md file to track progress for an implementation.
Make sure you have a solid plan for the app as it grows to allow you add more features. The only headache you’d suffer is when your Arch design is poor (including choice of state management and dependency injection, navigation system, and every other subsystem you may add e.g notifications).
1 Like
Haha, this is a solid approach! Treating your AI assistant like a “junior developer” rather than just a tool really makes a difference. Giving it guidance, letting it plan, and then validating everything ensures you don’t blindly trust its output.
I like how you highlighted the pitfalls—especially letting it shortcut tests or mock values just to pass. That’s such a common trap, and keeping manual oversight on actual code changes is key. Pytest is a great choice for Python automation, and having it focus on test coverage while you supervise the process sounds like a very efficient workflow.
Your phased plan—analyze, execute, validate, clean up, update—is basically how I try to structure AI-assisted development too. It balances autonomy with accountability perfectly. Definitely a good “ted talk” for anyone integrating AI into programming!
Thank you so much
Are you currently working with Flutter?
If so, I’d love to know what specific tips or practices you follow when building apps —
Do you usually rely on the official documentation, or do you use certain prompts with AI models?
And have you found this approach to be truly effective in your workflow?
I’d really appreciate hearing more about your personal experience with this
Thanks a lot, @Lance_Patchwork, really appreciate you sharing your experience! 
That was super helpful and I liked how you explained your workflow.
Do you use any specific MCP tools or setups that you find useful?
And if you have any other helpful tips or tricks like the ones you mentioned, I’d love to hear them!
Also, are you mainly working with Python, or do you use other languages too?
Thanks a lot, @Lance_Patchwork, really appreciate you sharing your experience!
That was super helpful and I liked how you explained your workflow.
Do you use any specific MCP tools or setups that you find useful?
And if you have any other helpful tips or tricks like the ones you mentioned, I’d love to hear them!
Also, are you mainly working with Python, or do you use other languages too?
Python seemed a good starting point because last time i tried doing programation 20year ago it was hell to setup a dev environment for C# and i knew python was kinda simple, i plan to expand but right now im more into learning with a few project i have on the sides, single player mud, a gurps bridge application to play with GPT or other browser in a type of play-by-mail and a full on text base spacesim as if you were in the bowel of a CIC, that mean instrucment navigation only, crew managing the various system.
The different scope of project allow me to learn a lot about how agents work and experiment with various provider, models, IDE quickly. I have a file ill s
scaffold_generator.7z (7.5 KB)
hare, its basicly a mini blueprint when i start a new project, built using python but you can open it with any text editor to see what it does, it basicly creates all the various .md pregenerated with placeholders for project specific information and everything else is data agnostics template and instruction for agents.
I installed the file_system MCP accross all the IDE’s since not all IDE’s offer the same subset of internal tools, the filesystem allow for multi-read and multi-edit wich not all IDE offers, for IDE addon i use a bunch of python based extension, git management and so on.
MCP used: Github, Context7, File System, Sequential Thinking (Give thinking to model that nativly doesnt “think”) Those are my goto’s.
I made a thread about tips i learned might want to check it out, i dont have the URL but follow my profile and you should find it quite fast, one last tip, work like elon musk, build fast, break fast, learn fast, build better faster. Create a few side project that force you to work differently, to get you out of your “safe-zone”. IE: One project hard rule, use only free model or only a specific provider or only non-thinking models it force your brain to rewire itself and find improvement you never though of, i 2 month i think i can consider myself well versed in agentic work. If i can do it, you can too, just force yourself to work in multiple different way to adapt.
Edit: Doh, i sent the wrong scaffold generator my bad.
1 Like
Thanks a lot, @Lance_Patchwork!
That was a really great and detailed explanation — I appreciate you taking the time to share all of this.
Your workflow and the way you use different MCP tools gave me a lot of useful ideas.
Seriously, thanks for the valuable insights! 