kept failing at building anything complex with AI until I tried this:

@d4m1n
Dan ⚡️@d4m1n
14 views May 13, 2025 ~3 min read
Advertisement
1
kept failing at building anything complex with AI

until I tried this:
Media image
2
if you like to watch a video instead, I got you
youtube.com/watch?v=dH4mc9… 🙌

Let me tell you how I did it:
3
I built an entire app with this.

imgxai.com/imgxai

It's a free, bring you own key ChatGPT image gen UI with a ton of advanced options: multiple image generation, a queue system, parallel jobs and detailed controlled from aspect ratio to quality to masking.
0:24
4
1. Write detailed requirements.
Spend 15-20m and think this through. The AI doesn't know what you want to build.

Be concise but lay out all details, including tech stack.
Media image
5
2. Create a PRD.txt / Product requirements document.
(Prompt in the blog post below)

Take the above and create requirements.
Double check these.
Media image
Media image
Media image
Media image
6
3. Use a task management system.
There are a few options:

1. Taskmaster AI task-master.dev
2. Roocode Boomerang Tasks docs.roocode.com/features/boome…
3. Shrimp Task Manager github.com/cjo4m06/mcp-sh…
4. Cline Task Tool docs.cline.bot/exploring-clin…

After trying them all, I prefer TaskMaster AI.
The MCP works great with Cursor, but you can also use the CLI directly.
7
But biggest reason to use TaskMaster AI:
It uses Perplexity to do web searches & seems a lot better at figuring out complexity because of it.

From what I've seen, it also identifies dependencies correctly & is good at implementing tasks in logical order.

Back to building.
8
4. Use a simple build loop that *only* works with tasks, no matter what.

Here's the prompt loop:

"Show tasks"
"What's the next task I should work on? Please consider dependencies and priorities."
"Implement task 2 and all of its subtasks."
"I'd like to add a new task that ..."

Some more examples in the blog post below, but there isn't a lot to it. Just don't go outside of the task management system, things will start to go wrong when your app gets complex.
9
Tips:
- Always start from a bootstrapped codebase. Don't ever start from scratch. Ideally, bootstrap it with a component library, but also consider form validation and database etc.

- Treat bugs like tasks. Don't just say “fix it” — describe the bug, what you expected, and how it's failing. Then create a new task and let the AI go at it systematically. If you let it band-aid bugs, you're just kicking the problem down the road.

- When prompting the AI to work on the next task, ensure to provide additional context on e.g. UI preferences, API docs etc. You can also attach images!

- Always start a new chat when implementing the next task

- Be flexible: deprecate outdated tasks and add new ones as you go, but remember to ask it to analyze.
10
All of these prompts and more are available in this blog post:

shipixen.com/tutorials/redu… 💪
11
It took me hours to build Imgx AI.

At best I could achieve this level in days instead, so I am quite impressed with what this can do even with current models.

This at a hands of a good developer is a fantastic combo!
What do you think of the app?
Isn't it kinda cool?

imgxai.com/imgx
12
sorry I am such an idiot, the link is: imgxai.com/imgx 🤦‍♂️
13
Here's the absolute best thing you can for continued success:

Break down files that are large than 500 lines.
AI is not great at handling large files.
It'll remove things and mess up the UI.

So if you have a file that is larger than 500 lines, break it down into smaller modules before working on a new task.

Here's a prompt you can use:
"Break down this file into logical modules so it's easier to ready.
Create directories if needed and move utils and interfaces to separate files, maintaining a domain-driven file structure."
Actions
What You Can Do
  • Export as PDF or Markdown
  • Batch Export to Notion
  • Bookmark & Highlight
  • Screenshot Tweet
Create Free Account

Includes 7-day Premium trial

Advertisement