The waterfall is so back

I keep catching myself mid-week with a half-written plan for a side project.

Notes pile up. Constraints get named. A rough shape of the end state shows up before I open the agent. Then the weekend arrives and I implement. It feels a little like waterfalling. It also feels like just planning. The balance is hard to describe, but it is definitely beating vibing.

That is the joke that will not leave me alone lately: because generating software got so cheap, a lot of the good advice for AI builds starts to look like waterfall again. Not the multi-month enterprise version. A tiny one. Specify. Generate. Check. Ship. On a short loop.

The old model in miniature

Winston Royce's 1970 paper is the thing people point at when they say waterfall, even though the paper itself is more restless than the cartoon version that followed. Requirements, design, implementation, verification, operations. Phases. Hand-offs. The expensive part was building, so you tried hard not to build the wrong thing.

I can definitely see now that this cost structure flipped.

An agent can spit out a working slice in minutes. The expensive part moves upstream into clarity: what do I actually want, what is out of scope, how will I know it is done. So people keep saying: do a planning session. Write a PRD. Lock intent before you generate. I agree that is smart. But I also find it funny that it is pretty much waterfall, with code that is nearly free to produce.

Of course this is not a dunk. Planning is still a good idea. The insight is that the old model is back in miniature because the build step collapsed.

The weekend loop that beats vibing

On side projects I have lived this a lot of times. I create a plan throughout the week. Not a sacred document. Enough to know what I want at the end. Then on the weekend I start implementing with agents.

You can see the difference in the result. Some software looks like somebody really thought about the process and the destination. Other software looks like somebody vibe-coded until it stopped crashing. Both can ship. But only one feels finished in the way I care about.

The miniature waterfall helps when the end state matters and the cost of wandering is higher than the cost of writing the plan. A short upstream pass buys a cleaner generate-check-ship loop. That matches the best version of the "plan first" discourse I keep saving: research the codebase, write a concrete plan with verification steps, and do not touch code until the plan is ready. Simon Last's planning prompt is a clean example of that posture without turning into enterprise theater.

When a PRD is the wrong gear

Design work has been teaching me the other side.

For design projects over the past few months, big PRDs have not been what we needed. Vibing, spiking, and trying a lot of different ideas in an exploration phase has been way better. The mode shift matters more than the ceremony. You need the right tool for the right job.

That is not anti-planning. It is anti-fake-planning. If I do not know what good looks like yet, a polished requirements doc becomes theater. A spike teaches me faster than another paragraph of acceptance criteria. Karl Koch puts a related idea plainly: choose the tool from the needs of the task, not from a rigid workflow you perform for comfort.

The miniature waterfall quietly becomes over-spec when the work is still exploratory. Then the cheapness of codegen is wasted on waiting for a document that should have been a throwaway prototype.

"That is just think before you type"

The strongest pushback is fair: calling this waterfall is clickbait. Real waterfall was about irreversible phase gates and handoffs. A twenty-minute plan-then-generate loop is closer to agile with a better prompt.

I am keeping the punchy title anyway. A bit of clickbait is okay. The rhyme still does work for me, because it names the shape of the process when build cost collapses: clarity first, then a fast build, then a check. The difference from classic waterfall is the size of the loop and how cheaply you can reopen it. The similarity is that the scarce input is still human clarity, not typing speed.

That lands next to claims I have already been writing around. High agency, low authority is about where the commit point sits once agents can prepare end-to-end work. Compound judgment, not better artifacts is about what should improve inside me while generation gets cheaper. This piece is narrower. It is about process shape when the build step is almost free.

Pick the mode on purpose

I do not want a universal rule that every AI build starts with a PRD. I also do not want "just vibe" as the default because generation feels magical.

The useful move is noticing which gear I am in.

If the destination is clear and the cost of a fuzzy end state is real, a tiny waterfall helps: plan through the week, implement on the weekend, check against the intent you already wrote down. If the destination is still foggy, especially in design exploration, spike first. Let the prototype teach the plan. Switch gears when the job changes.

Bad waterfall was process that outlived the feedback. A miniature waterfall that you can reopen in an afternoon is something else. It is planning that finally matches how cheap the build step became.