My AI Toolkit
Sep 2026
This recipe is the result of weeks of experimentation with various AI tools, free and paid APIs, IDEs and agents. This is what has worked for me, given my unique requirements.
When I embarked on this journey, I intentionally chose to go back to the basics: Git, Linux, VS Code, and then Cursor, to fully understand what was happening under the hood of AI-assisted coding. The more I experimented with these tools, the better I discovered their limitations and strengths.
I also found myself drawing on the nearly 30 years spent in Enterprise technology, Architecture and System design, augmenting it with AI assistance as an efficient way to apply those fundamentals to my projects.
This repository chronicles my journey and shares some of those learnings in the spirit of open source. It will be a continuously evolving body of work, as I keep improving...
(If DIY is not your thing, buying a managed AI coding plan (Claude/Codex/etc.) and letting it handle all the decisions may just be the easiest way for you to get started.)
Guiding Principles
The following principles drive the decisions and recommendations you see in my toolkit:
- AI is a very rapidly evolving landscape; what works today may not work tomorrow
- The competitive environment for tool/model makers is fierce; my view is that it's best to de-risk by keeping your options open, and not have all your work hinged on a specific choice
- Current AI billing systems are highly complex and often opaque, and it's not easy to understand how your usage limits are being consumed; I think it is prudent to keep your costs under control any way you can
- When you work with these tools daily, your opinions keep evolving eg. Lately, I find ChatGPT tends to offer needlessly complicated solutions when simpler ones will do, which is all the more reason to be able to switch models/tools as needed
- Many AI Agents have a very aggressive approach; they will do what you ask for and then some, making significant edits for a small/precise ask; after a lot of trial and error, I have learnt which ones to avoid!
If you do not subscribe to the above philosophy, you can skip the rest of this page...
1. Tool Matrix
| Category | Primary | Secondary | Notes |
|---|---|---|---|
| AI Chats | Gemini/Deepseek/etc. (Planning & Analysis) | Open Web Models | I prefer multi-provider interfaces with low/no vendor lock-in |
| Coding IDE | Cursor (Subscription) | VS Code + Kilo | Cursor for daily work; VS Code + Kilo (BYOK) for when plan limits drop |
| APIs | OpenRouter (BYOK) | OpenAI/Gemini/etc. Direct Keys | OpenRouter for dynamic model routing; Direct keys for dedicated quotas |
| Models | Use Built-in when available | Model-Agnostic BYOK | Choose as per task to balance capability and cost, as needed |
2. What Each Is Good For
| Need | Default Approach |
|---|---|
| Everyday IDE | Cursor (Included plan allocation for primary workflows) |
| Fallback IDE | VS Code + Kilo Code + BYOK APIs (Zero monthly seat cost, pay per token) |
| High-Reasoning / Architecture | Deep-reasoning models via OpenRouter (BYOK) |
| Experiments / Testing | OpenRouter (BYOK) + Model of choice |
3. Model Routing & Execution (BYOK)
| Task / Mode | Minimum | Alternative | Notes |
|---|---|---|---|
| Plan | deepseek/deepseek-r1 |
anthropic/claude-3.5-sonnet |
Architect Mode: High reasoning. Generates blueprints and maps dependencies before touching files |
| Orchestrator | deepseek (V3)/deepseek-chat |
openai/gpt-4o |
Subagent Manager: Strong tool-calling precision; spawns tasks without runaway loops |
| Code | deepseek/deepseek-v4-flash-latest |
qwen/qwen-2.5-coder-32b-instruct |
Workhorse: High throughput, ultra-cheap file editing and refactoring |
| Debug | deepseek (V3)/deepseek-chat |
deepseek/deepseek-r1 |
Troubleshooter: Diagnostic analysis of error stacks and tracebacks |
| Ask | qwen/qwen-2.5-coder-32b-instruct |
google/gemini-2.5-flash |
Codebase Search: Answers context queries and inspects docs with zero risk of editing files |
Note: Model capabilities and costs are rapidly changing. Think of this as a starting point, and use your discretion to reassess as needed.
ABC Ladder
🟢 A — Default (Use First)
- Chatbot / IDE Agent → Initial planning and high-level strategy
- Cursor IDE → Primary coding IDE subscription with monthly plan allocation
- VS Code + Kilo → Everyday utility coding via BYOK
🟡 B — Better (Escalate)
Use when A isn't good enough / plan limits are throttled
- VS Code + Kilo + OpenRouter BYOK → Scale execution beyond subscription limits
- DeepSeek V3 / Claude 3.5 Sonnet / GPT-4o → For complex multi-file orchestration / refactoring
🔴 C — Specialist (As Needed)
- DeepSeek-R1 (OpenRouter) → High-difficulty logic bugs, system architecture and algorithmic design
- Deepseek/Gemini/Claude/etc. → Second opinion for validation and cross-checking reasoning
- OpenRouter Endpoint Switching → A/B test specialized models for specific niche stacks
eof
This project captures ideas, decisions, and lessons I’ve learned while building with AI. It aims to help non-technical users understand and use these tools effectively. I also use AI at times to improve or refine the content.
Readers should use the commands and instructions here with care. The creator accepts no responsibility for system damage, data loss or other consequences resulting from their use.