I built a service on top of my own AI workflow. Here's what 14 days of shipping looks like.

I'll skip the hot take on AI replacing engineers. You've read it. I want to talk about the boring problem nobody is fixing: AI helps you draft code, but shipping it is still the bottleneck.

For most of last year I watched the same pattern across founder friends. AI spits out something that compiles, types check, looks right in a diff — then it breaks the moment a real user touches it. You spend the "savings" cleaning up the slop. Hiring is months of overhead before the senior eng you wanted is even productive. Agencies want discovery decks and weekly syncs. You wanted features shipped. You got a project plan.

Three failure modes, one outcome: your roadmap is moving slower than your AI hype cycle.

I had this exact problem on three different projects, and I got tired of solving it by working weekends.

I spent the last year building a workflow on top of Claude Code that actually ships to prod. Then I turned it into a service: Ondemandly. The pitch is one line — bring the backlog, I ship the work — and the proof is that I built the entire service on the same workflow I sell.

What 14 days looks like

To stand up Ondemandly itself: 18 merged PRs in 14 days. Solo. That's the landing site, Stripe Checkout (with a Brazilian PSP integration because I'm not based in the US), Cloudflare Workers, abuse defenses (rate-limit + Turnstile + a global cap), three legal pages, the brand system, the deploy pipeline. Production. Live.

Nothing about that pace is heroic. It's the workflow.

What's actually in the workflow

This isn't "I prompt Claude harder." It's a stack of constraints that force the AI to behave like a senior engineer instead of a junior one with a thesaurus.

Plan before you execute, never the same step. I have a hard rule codified in my repo and global instructions: writing a plan and starting to execute it are two separate steps. The model writes a plan, stops, and asks for approval. This single rule kills 80% of the "AI looks right, ships wrong" failure mode, because the failure mode is almost always the model rushing into code before anyone agreed on the approach.

Subagents in parallel, not one big agent in series. Specialized agents — Explore, Plan, code-reviewer, deployment-expert — dispatched per task instead of one model trying to hold the whole problem in its head. Each agent has a tight prompt and a narrow tool surface. Cheaper, faster, less drift.

Git worktrees for parallelism that doesn't collide. Four streams of work in flight at any time, each in its own worktree, each on its own dev port. You can't do this with one branch and one terminal — but you also can't do it without discipline, because four parallel streams of "vibes coding" is just four parallel disasters.

Durable project memory + a thoughts/ doc system. Decisions, ICP, won't-dos, in-flight features, post-incident notes — never re-derived from scratch in a new session. The model walks back into a project knowing what we already decided three weeks ago. This is the part most "AI workflows" skip and it's the reason they degrade after a few days.

TDD, code review, simplify, verification — built-in skills, not optional vibes. The skills enforce the steps. You can't claim done without verification. You can't merge without review. The discipline isn't a culture thing, it's a tool thing.

That's the whole workflow. There's no secret prompt. The leverage is in the constraints.

Why I packaged it as a service

Here's the part that took me a while to admit: most founders don't want a workflow. They want the output. The Notion full of tickets, gone. The "we'll get to it next sprint" list, cleared.

The shape is borrowed. Brett Williams at Designjoy proved productized design works as a flat-fee subscription with one active task at a time and async-only delivery. Ondemandly is that same shape pointed at engineering execution.

So Ondemandly is shaped around that:

• One active ticket at a time. Not because I can't do more in parallel — I do, internally — but because one ticket at a time is what makes it a dependable ticket eater instead of a freelancer with random availability.
• Async first. No standups. No weekly syncs. Visible progress on the board.
• Narrow scope on purpose. I don't do legacy rescue, open-ended discovery, 24/7 on-call, or custom design from scratch. Saying no is what protects the speed.
• Flat $4k/month. Pause anytime. Days carry over. No lock-in.

The math, if you care

A senior engineer at loaded cost (base + benefits + payroll tax + amortized recruiting and ramp) is roughly $19k/month, or ~$228k for year one. Ondemandly is $48k for year one. That's a $180k delta, with day-one productivity instead of a 2–3 month ramp, and no hire-to-fire cycle if priorities shift.

I'm not pretending it's a 1:1 substitute for a senior FTE. An FTE owns more. They sit in your design reviews, hold institutional context, get paged at 3am. Ondemandly does none of that on purpose. It's an execution layer for teams that already know what to build.

Who this is actually for

Three patterns I see over and over:

1. Solo founders with a Notion full of tickets and no one to clear them. You know what to build. You don't have time to build it. Hiring is months away.
2. Product teams whose engineers are heads-down on the hard problem. You've got a 0.5-FTE backlog of well-defined feature work that nobody on the core team should be touching, but it has to ship.
3. Operators wiring AI into a real product, not a demo. The integration work, the eval harness, the boring glue between the model and production.

If you're none of those, you're probably better off hiring or doing it yourself. I'd rather not waste your retainer.

The honest version

I'll save you the "founder journey" wrap-up. The actual story is: I had a workflow that worked for me, the math worked out for the kind of teams I wanted to work with, and the product is just exposing that workflow as a subscription. There's no growth playbook, no team, no funding round. It's me, the workflow, and a Stripe link.

If you've got a backlog and you're nodding at any of this, the next step is one concrete ticket. We refine it together, I ship it, and you decide if the rest of the backlog comes next.

Happy to answer questions in the comments. Especially the skeptical ones.