Building an MVP: which decisions matter before you ship?

Before an MVP, decide the one workflow, how you will know it worked, what you can undo, and who owns the release after launch.

Discuss a project

By Keith Rull ·

Decide four things before you build an MVP: the one workflow a user must finish, how you will know it worked, what you can undo after launch, and who owns the release the week after. Everything else can wait. A startup that decides the stack, the brand, and the roadmap first usually ships a demo nobody operates.

An MVP is not a small version of the whole company. It is the smallest release that teaches you whether the workflow is real. The decisions that matter are the ones you cannot cheaply reverse once a user depends on them.

What has to be true before anyone writes code?

Write the workflow as a user would tell it. “A clinic admin invites a colleague, the colleague sets a password, and an expired invite shows a single next step.” That paragraph is the product. A list of features is not.

Then write how you will recognize success without a dashboard fantasy. For that invite, success is a person who can sign in, and a person who cannot when the invite is old. Count those two cases. Do not count pages designed.

Product development starts from that same cut: one workflow, acceptance examples, and an explicit out-of-scope list. If you cannot fill the out-of-scope list, the MVP is still the whole product with a shorter deadline.

Which early choices are expensive to undo?

Some choices feel technical and are actually promises to users.

Accounts and permissions. Who can invite whom is easy to widen and painful to narrow after people rely on the wider rule. Decide the smallest permission that lets the workflow finish.

The thing you store that you cannot casually delete. A message a user sent, a payment you captured, a file you accepted. The MVP should name which of those exist. If none exist yet, say so, and do not add them because the framework makes it easy.

The path you cannot roll back. A DNS change, a data migration, an email you already sent. How to keep an architecture decision reversible is the record for that path: the behavior you must restore, who can restore it, and what the rollback will not undo. An MVP that skips the record ships a one-way door with a friendly name.

The vendor you let sit in the middle of the workflow. A payments provider or an identity host is fine. A vendor who owns your only copy of the customer list is a different decision. Write down how you would leave.

This sketch is hypothetical. A team wants “social login, teams, billing, and an admin console” in the first release. The workflow that teaches them something is one admin inviting one user. Billing can be a manual invoice. Teams can be a single shared space. The admin console can be a script the founder runs. None of those cuts are permanent. They keep the first release inside a boundary you can operate.

How do you decide what waits?

Use a blunt sort.

Ship now if the workflow is false without it. Wait if a person can do the step by hand for the first ten users. Drop it if you added it because a template had the screen.

Decision Ship in the MVP Wait
The one user path, including the failure Yes
A way to undo or disable that path Yes
A named person who can release and recover Yes
A second workflow, reports, or a redesign of the brand Yes
A model call you have not evaluated on real cases Yes, or cut it

The last row is where AI MVPs go wrong. A chat box is not a workflow. If the feature drafts an answer a person must check, write the check into the path and decide what happens when the model is down. Why AI pilots stall before production is the longer version: name the task, cluster the failures, and count the review time before you call the demo a product.

If the only open question is one architecture call, and someone inside the company already owns the product, that can be a single consulting pass rather than a build. CTO consulting is a written recommendation and a stop, not a standing team.

What would I do in the first month?

I would freeze the workflow paragraph and the acceptance examples, then refuse new screens until those examples pass. I would name the release owner on the startup’s side before I named a stack. I would keep the previous build, even when the previous build is “the site is off,” so launch day has a back button.

I would staff from the same rule as any other release. If the startup already has a lead who can review, add the missing skill for the stretch. If it does not, do not hire three builders and hope direction appears. Staff augmentation or outsourcing is that choice. The security floor for the same release is small and specific: who can sign in, what is logged, and which dependency you actually run. That baseline is a practical security baseline for an early-stage startup.

You can make these cuts without outside help when a founder can describe the workflow and say no to the second one. Bring in help when the backlog is a tour of the future product and nobody can point at the path a user will finish this month.

FAQ

How small is small enough?

Small enough that you can demo the failure as well as the success in one sitting. If the happy path takes an afternoon to explain, it is more than one workflow.

Should the MVP be throwaway code?

No. Throwaway code becomes the production path the week a user depends on it. Build the thin path as if you will run it: a test for the acceptance example, a way to turn it off, and a person who can do that.

Do we pick the stack in week one?

Pick a stack the release owner can operate. Do not pick one because it won a comparison chart. The expensive mistake is a platform you cannot reverse, not a framework you can replace behind one workflow.

What do we show investors?

Show the workflow, the acceptance examples, and what you learned from a real attempt. A roadmap of unbuilt modules is not evidence. If a buyer will later ask whether the architecture can be undone, the reversible-decision record is the artifact, not a slide of logos.

If you’re cutting an MVP down to the release you can run, talk to us.

Build your startup’s MVP, and be able to run it

Discuss a project