Ten Income Lanes — and the Codex Workflow I Trust

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Two sets of notes changed how I plan paid work this year. One lists ten ways to let AI create monthly income. The other is what two hundred hours on Codex actually taught. Together they are a ladder: pick a lane, then use an agentic coding app without letting it burn production.

Lanes that sell outcomes, not chat logs

AI voice agents for local businesses, built with tools like Retell AI, are the clearest example. A five-person team selling at twenty to fifty thousand retainers quotes two months and now builds in two to five days — because clients buy answered phones and booked appointments, not model names.

Mercor pays people to evaluate and train models. That is income from judgment, not from shipping product.

Operations automation with Zapier, n8n, and Xero covers event-triggered work that a chat agent sitting in a browser cannot reliably catch — invoices, form submits, failed payments.

Content and voiceover work is real, with a sharp constraint: ElevenLabs needs a paid commercial plan for sellable clips. Free-plan audio stays in the lab.

Teaching ChatGPT or Claude one-on-one still works because most users ignore Projects and Skills. If you can install those habits for a team, you are selling leverage.

The software pricing ladder in the notes

Google AI Studio is free for learning. Lovable around twenty-five dollars a month supports charging five hundred to two thousand per project. Replit, Claude Code, or Codex around twenty dollars a month supports projects in the two to eight thousand range when you can deliver. That ladder matches how I quote: price the outcome and the risk, not the subscription you used.

Further lanes in the same list: Clay for lead lists, a vertical AI product, AI bookkeeping with Basis, and team AI workflows paid from training budgets. A master prompt that designs your thirty-day plan is only useful after you pick one lane. Spreading across ten is how nothing ships.

What two hundred hours on Codex changed

The useful framing: stop using ChatGPT like it is still 2025. The Codex app behaves like a full-power OpenAI workspace. The sidebar holds Projects, plugins (Slack, Notion, Figma, Canva, Calendar), and Automations.

One automation I copied immediately: every morning at nine, pull a newsletter, extract key points, build a short PPT, leave it for review. That is agent work with a human gate at the end.

Permission levels matter. Default pauses for review. Auto-review asks only for major actions. Full trust is for sandboxes. I live in Default for anything near client data.

Plan mode is how I start serious builds. Codex asks product-manager-style questions and plans before it codes. I answer those questions in writing so the plan becomes the contract.

Smaller tricks from the same notes: drawing screens in Paper through MCP, and switching to OpenRouter by leaving a comment on a code line when I want model flexibility.

What I skip on purpose

I skip lanes that require me to pretend I am a medical or legal expert. I skip anything that needs Full permission on day one. I skip “build ten automations” plans that never name the first buyer.

Codex plugins into Slack and Notion are tempting. I connect them only after the project has a written data policy: which channels are readable, which databases are writable, and how chat logs are retained. Automation without that document is how client text leaks into the wrong workspace.

The two-hundred-hour lesson is not “always use Codex.” It is that an agentic coding app with Plan mode, permission tiers, and scheduled automations beats pasting stack traces into a generic chat — as long as you stay the reviewer.

My Monday setup

Pick one income lane for the next thirty days. Stand up a Codex project with Default permissions. Write a plan-mode brief. Wire one automation that produces a reviewable artifact before nine-thirty. Keep purchases, sends, and deletes on manual approval.

Income follows repetition. The stack only matters if the calendar does.

If you only take one lane from the income list and one habit from Codex — Plan mode plus Default permissions — you already have a quarter’s worth of focused work.