Why I use Codex

I want to be honest about something upfront: I'm not a developer. I'm a manufacturing business owner who taught himself to build software products because I had ideas I believed in and couldn't afford to outsource everything. So when I tell you why I use a tool, I'm telling you from that perspective — not a senior engineer's, not a startup CTO's. Someone who needs things to work without spending two days configuring them.

Codex is the tool I reach for when I want to delegate a task and come back to something finished.

---

## What Codex actually is

Codex is OpenAI's agentic coding tool. Not an autocomplete. Not a chat assistant you ping for code snippets. An agent you point at a task and let run — it reads your repo, figures out what needs to change, makes the changes, runs checks, and comes back with results.

The interface that changed how I use it is the desktop app. It runs parallel threads — you can have one agent working on a bug fix while another is building out a new feature and a third is running a code review. Each thread is isolated. They don't interfere with each other.

For someone building two projects simultaneously — I'm building an internal CRM for my manufacturing business and a booking marketplace at the same time — that parallel execution is genuinely valuable.

---

## Why I use it instead of just Claude Code

I use both, and they do different things well.

Codex is what developers describe as better for "fire and forget" tasks. You write a clear brief, send it to Codex, go do something else, come back to a finished result. The Inbox feature in the desktop app is built for this — completed tasks accumulate in a review queue while you're away, each one with a diff you can approve or reject.

The community consensus in 2026 is pretty clear on this: Claude Code has better code quality but hits usage limits too quickly to be a daily driver. Codex is slightly lower quality but actually usable. The smart move is to use both.

That matches my experience exactly. Claude Code is where I go when I need to reason through something complex or debug something that's broken in a non-obvious way. Codex is where I go when I have a clear, well-defined task and I want it done while I'm getting on with something else.

The process is different than writing code by hand — spending anywhere from 30 minutes to two hours writing prompts and generating context, then the task runs for 15-20 minutes while you context-switch to something else entirely. That's exactly the workflow I've built around Codex.

---

## What makes it work for a non-developer

The thing that isn't talked about enough is that Codex tends to produce high-quality results out of the box. It's generally more detailed in the UI by default — it shows you exactly what it's doing and usually explains it too.

That matters if you're not a developer. When Codex finishes a task, you can follow what happened. You're not staring at a diff wondering what changed and why.

The AGENTS.md file is the other piece. Drop it in the repo root, and Codex reads it at the start of every session. Your rules load automatically — no special prompts, no remembering to set context. That's the foundation of the TNDD workflow and it's what makes Codex genuinely reliable rather than occasionally helpful.

---

## The Inbox — the feature I didn't expect to care about

The Codex desktop app has an Inbox. When agents complete tasks, results accumulate there for your review — inline diffs, what changed, approval or rejection in one place.

I didn't think I'd use this much. Now it's how I start most mornings.

I'll write three or four tasks before I finish for the day, let them run overnight, and review the results in the morning. Each one comes back with a diff and what the agent did. I approve, reject, or push back. Nothing hits production until I've seen it.

For someone building in short windows of time — I run a manufacturing business, I don't get uninterrupted coding sessions — this asynchronous model is the right one. Lovable works in real-time. Codex works while you're away. Both have their place.

---

## Where it fits in my stack

Lovable handles new screens and UI scaffolding. I'm not leaving Lovable for that — it's genuinely excellent and nothing else comes close for visual iteration speed.

Codex handles everything else that can be scoped clearly and sent off: features, bug fixes, refactors, code reviews. Tasks where I can write a clear brief and trust the result.

Claude Code handles complex debugging and anything requiring deep reasoning about what's broken or how something should be architected.

Vercel and GitHub Actions handle all deployments. Codex never touches production directly.

---

## The honest limitations

Codex uses more tokens per task when you're running it hard. Pricing is complex — ChatGPT Plus at $20/month for limited usage, Pro at $200/month for unlimited. For most of what I do the Plus plan is fine, but heavy overnight runs add up.

The other limitation is that it works best on well-scoped tasks. Give Codex a vague brief and it'll make reasonable guesses — sometimes right, sometimes not. The TNDD pre-flight workflow (inspect, plan, declare scope, confirm what won't be touched) is what I built specifically to address this. Without that structure, Codex is capable but unpredictable.

With it, it's the closest thing I've found to having a developer who works through the night and hands you clean results in the morning.

---

*The TNDD AI Builder Workflow Toolkit includes the AGENTS.md file and install prompt for Codex — sets up the full workflow in any repo in under ten minutes.*

*[Download the Toolkit]*