TODO

• try Kilo Code and Roocode with Chutes or Z.ai API
  • Kilo code has $20 top up on 1st top up (extra credits expire after 60 days)
• cancel warp.dev and find alternative that is good for terminal tasks
• search for claude founder workflow tweet

example workflows

• ask perplexity for a map and explanations -> bolt to gen mvp of app -> cursor to program remaining details.
• another example workflow: find a template project -> claude code /init

Key concepts

AI Agents concepts
AI Development

tools

This guy ranks the performance of agent/models combos: https://www.youtube.com/watch?v=bp5TNTl3bZM

what I'm using and considering

Currently I'm using:

• Warp and nvim + sidekick (CLI coding agents)
• Cursor and when I run out of credits use CLI agents (This way I get good CLI agents + good unlimited autocomplete)
  • Kilo code + z.ai / chutes API
  • GeminiCLI / QWEN Code in the terminal inside cursor
• Antigravity has free high end models

It makes sense to cut out Warp ... but it works very well but with Cursor and Google subscription I barely use any credits.
it's good for automating terminal and operating system tasks.

misc

• perplexity.ai - AI powered search engine. More citation focused. Use when you need the latest documentation. Has the most up to date info.
• grok (xai) is good for market research bc it has access to real time tweets. Very valuable up to date content.

models

best coding models:

• Claude 4.5 Sonnet (best and expensive), and 4.5 Haiku (Good and Cheap)
• Gemini 3 Pro and fast (free/cheap)
• GPT5 is amazing at consistently following instructions and tool calling.
• grok code fast (cheapest and fast)

open source models. Very cheap or free

• GLM 5 - pair with Claude Code, Codex
• Qwen3-coder (free) - can pair with QWEN Code but also kilo code for better performance
• Deepseek R1

cheaper APIs

• novita.ai (many models)
• Chutes API (multiple models)
• Open Router is a service unified API gateway that gives you access to all LLMs and is a cost effective way to pay for good models.
  • it's an aggregator. you can access providers like Chutes via openrouter

This APIs that offer many models all have oficial documentation on how to use with many agents like Claude Code
Keep in mind that paying for API and subscriptions directly from model providers will most likely give me the best price. so a Codex subs will give me more value for using codex and GPT models.

• z.ai offers an API with their GLM 4.7. cheaper model and cheaper API
• moonshot.ai kimi 2.5 API
  note that GLM doesn't have vision so I cant paste and send images to it..

code editors

• Cursor code editor ⭐️ (vsc fork)
• Antigravity
• VSCode has their agentic code assistant features but most people lean towards forks or extensions that perform better.

vscode extensions that replace the AI functionalities built into the editor itself:

• kilo code
• augment. great feature where it holds in it's context with your entire updated code base. this saves you $ on models trying to gather context reading your files repeatedly.
• cline/roocode(fork)
• there is a google one (but most of the focus seems to be on gemini CLI)

nvim plugins:

• 2025-10-01: folke's sidekick.nvim - integrates all CLI agents
  • NES: predicts your next change somewhat like tab completion
    • finnicky
    • (doesn't replace inline suggestion like copilot.lua)

terminal / CLI agents

2025-09-05: some of these have VSCode extension for better integration with the editor (vsc) like knowing what code is highlighted in the editor (Gemini)

• Warp. Initially a terminal emulator with AI features but they have expanded focusing more on being a coding environment and AI agents
• Claude Code
• codebuff (apparently better than claude code)
• Codex (OpenAI)
  • $20 chatgpt subscription. doesn't proper limit
• opencode
  • crush controversial fork that was bought from the original by Glam
• Gemini CLI FREE
• QWEN Code (Gemini CLI fork that uses QWEN models) FREE

app generators

• https://bolt.new/ / bolt.diy (open source alternative) bolt.diy setup tutorial
• lovable (1st version of database)
• Firebase Studio
• reloom and webflow have AI builders that can gen a landing page (with animations) very fast that you can then edit.

autosuggestions

This feature started it all. These are tab autocomplete recommendations, available in all code editors.

• copilot
• supermaven

design

you could do a sketch in excalidraw or figjam if you have a clear idea of what you want the UI to look like. and pass this image to the following tools:

• figma "first draft" feature generates high fidelity designs. You can also add a figma MCP to cursor to give context of your designs and generate the code for them easier.
• v0 generates UI design & code (shadCN components)
• Relume to generate a design system / style guide.
  for assets generation (images and logos) see image and video AI generation
  design to code: - builder.io (NOT builder.ai)

other

• snyk: security scanning
• jam: debugging, filing bugs to send to devs: writing ticket title, description and images.
• gitingest.com: Turn any Git repository into a simple text digest of its codebase. This is useful for feeding a codebase into any LLM.
• https://deepwiki.org/: repo docs/wiki generator. This one could also be even more helpful to paste into an LLM.
• mintlify: another docs generator
  code reviews:
• coderabbit: code reviews. There's also one called 'graphite'
• gemini CLI custom /review command seems decent

running locally

• look into running models locally: jan.ai, ollama,
• LM Studio let's me download models and run a server that I can connect my coding tools to

prompts

coding prompts

tips and tricks

• use rules, skills and subagents

• commit often and use git as a checkpoint system. If you are using CLI agents this is the only checkpoint system, but even when using GUI that have checkpoints (like Cursor) I've found they are not 100% reliable.

• if context window is too large, start a new chat and reference the relevant files. Sometimes you can get away with compressing the conversation and continuing.

• avoid huge refactors.

• periodically prompt LLM to review the code and ensure it follows best practices. This is to try to keep things neat.

• comment a lot so ai models can better understand files.

• create a Product Requirements Document (PRD) and reference it always.

• roadmap.md this file should include your developer roadmap. what you are focusing right now, what would be the next step and other stuff to do in the future.

• have the AI tell you what to do

• after getting a response, ask AI to double check it's output and check for edge cases.

How I reduced 90% errors for my Cursor (+ any other AI IDE) great vid