My (Cheap) AI Workflows

Back in January, I figured: if AI can do it faster, why not?

To skip the fluff, here's my setup.

My Setup #

• Opencode Go (using oh-my-opencode-slim)
• Paseo

Notable AI tools/extensions #

• rtk
• fff-mcp
• grill-me
• handoff
• caveman

Cost #

• Server (for paseo CLI so I'm not opening my laptop that often) ($6)
• Opencode Go ($10)

My smoking budget is dead because of this. I call that a win-win.

Models that I use #

All cheap. I can't even burn through my $10 sub each month. I use MiniMax M3 for the orchestrator, DeepSeek v4 Pro/Flash for everything else.

How I use it? #

First, I set up my GitHub and GitLab API tokens (read-only) — just to read issues assigned to me. With Paseo, I schedule that as a daily task.

paseo schedule create \
  --cron "0 0 0 0" \
  --timezone UTC \
  --run-now \
  --name gitlab-issues-summarizer \
  --provider opencode/deepseek-v4-pro \
  --cwd ~/dev/my-app \
  "Fetch Issues that asigned to me using skills /gitlab-issues-summarizer"

It ranks my issues for the day and I pick them off in order. One-line issues get a worktree; the heavier ones need real context.

I skim it on my phone — there's a complexity/priority breakdown, so I pick the easy one first. While the AI is grinding, I read the PRD for more complex task and figure out the business logic in my head: what it should look like, which existing functions I can reuse. So my prompts stay mostly technical.

I always review the diff. Paseo has a decent phone diff viewer, but on desktop I stick to Neovim.

That's the day-to-day. I use AI to search the codebase and ramp up on the platform and the business logic — because sometimes business requirements don't translate cleanly into code. How do you tell a PM that product search hits a denormalized table instead of joining row-by-row?

Paseo also has per-project config (lifecycle hooks, etc.) — handy for spinning up worktrees without lifting a finger.

That's the whole setup. The killer feature: it runs on my phone. So I can "work" from anywhere with a signal — bus stops, queues, the dentist's chair.

Server Security #

How I keep my AI from doing something dumb and irreversible:

• Run the AI as a non-root user, no sudo
• Block destructive shell commands (rm -rf, force pushes) at the hook level
• Never expose the server publicly — I'm using Tailscale

I want to go cheaper #

Next: move from OpenCode Go to OpenCode Zen or OpenRouter's pay-as-you-go pricing.

Drawbacks #

AI start fresh every sessions, so I got a skills to gather the context before it executes fix/feature