Take a Bolt App to Production Without Another Prompt
Bolt generates a full-stack app in the browser. Production is a second user, secrets off the client, and a deploy on your cloud with a way back.
By Muhammad Saif and Aashan Khan · Techtribe co-founders
Bolt will build you a full-stack app in the browser. You describe the product. You get screens, a server, and a data store you can click through. You can edit the files. That is a real start, and it is a better start than a slide. It is not production. Production is a second person signing up, a secret that is not in the client, and a deploy you can undo without asking the same prompt box to remember what it did.
The build guide for this move is take your Bolt app to production. This article is the decision around that guide: when another prompt is the wrong tool, what is worth keeping from the generated repo, and when the honest sentence is a migration. If the app was not Bolt, the siblings are v0 and Replit. The verdict that applies to all three, repair or rewrite, is when to repair a vibe-coded app.
What Bolt already did, and what it did not
Bolt generates the stack, not only the picture. That is the fact that changes the work. A tool that only draws the interface leaves auth and data as an obvious gap. Bolt fills the gap with files. The files look like a product. They compile. A friend can click the preview and believe you have launched.
What the preview does not prove is that a second account is a second account. It does not prove a payment in Stripe became a row you can reconcile. It does not prove the key that calls a model, or a database, is absent from the bundle a stranger can read. It does not prove you can deploy that same commit to a cloud account you own and roll it back on a Thursday.
The habit that hides those gaps is the next prompt. "Make it production ready" is not a change a generator can apply. It is a list. The list is an inventory of every service the app touches, a schema you can reproduce, access rules tested with two accounts, authorization on the server, secrets moved and the old ones revoked, staging that is not the production database, and a deploy with a way back. Some of that list is already in the Bolt repo. Some of it is a fiction the preview implies. The work is telling those apart.
When prompting stops being the fix
Stop when a repair breaks a different feature. You ask Bolt to fix checkout. Settings stop saving. You repair settings. Login drifts. The tool is rewriting more than the bug, because the bug was a rule and the tool can see files. The last working version lived in a chat. Freeze the chat. Export the files. Read the path.
Stop when the same error returns after three prompts and the credit balance moves anyway. You are paying to sample the same failure. Capture the repo. Have a person read the function that should have rejected the write. If they cannot find a function, that absence is the finding. It is not a prompt to "add validation everywhere."
Stop when account B opens account A's row, or a logged-out request still returns data. That is access control. It is the class of hole that shows up again in anything else you generate on top, including a chat. Another prompt that hides the row in the interface leaves the request in place. The next feature will call the request.
Stop when money is silent. The processor captured a payment. Your table did not. Or it captured two, because a webhook retried and nothing was idempotent. The article for a webhook that never arrives is Stripe webhook not triggering. Read it for the single endpoint. If "paid" only exists inside Bolt's preview and the processor's dashboard, you are not debugging a webhook. You are missing a ledger. That is a scope.
Stop when you cannot explain the write. If you cannot point at the check, you cannot tell an investor how a customer's data is protected, and you cannot safely let Bolt edit that file again. Explanation is the gate. Fluency is not.
What to keep from the generated app
Keep the flow. If a person can finish the job in the preview, film it and treat the film as the spec. A migration that begins from a blank description of "like the Bolt app, but better" will spend a week re-deciding buttons. The buttons were the progress.
Keep the files that match the flow and are readable. A page that renders a list, a form that posts a shape you can name, a type that matches the table. These are ordinary code. Ordinary code does not need to be punished because a model wrote it. The technical-debt note, AI generated code technical debt, is about the debt that sticks, not about a moral objection to generated files.
Keep the product decisions: who the user is, what they are allowed to finish, which empty state you argued about. Write those down outside the prompt. The prompt history will not be what a hire reads.
Do not keep a client that holds the service-role key. Do not keep a query that returns every row and filters in the browser. Do not keep a schema you cannot draw. Do not keep secrets that were pasted into a file the preview bundled. Revoke them. Rotation is part of the move, not a follow-up ticket you will forget after the domain switches.
Patch, or migrate, in that order
Patch when the core flow is sound and the gap is narrow. One missing server check. One webhook signature. One environment split. You can do that in the repo Bolt already created, if the repo is actually yours and you can run it without Bolt. If you can only run it inside the tool, you do not have a patch. You have a suggestion the tool may overwrite on the next prompt.
Migrate when the data model or the auth cannot be trusted, or when the only deploy path is the tool. The safe order is boring. Freeze a version you have clicked through. Put that version, unmodified, on a host you control. Prove the core path there. Only then change the dangerous parts, on that copy, with a rollback to the unmodified deploy. Point the domain last. Bolt can stay published for a few days as a fallback, knowing that any row written to either copy after the cutoff has to be carried across by hand. Two live writers is how you lose an order.
A rebuild of the core flow is the right offer when the generated one cannot be migrated. That is not a slogan. It is what you say after reading the schema and the auth, not before. The repair or rewrite note is the test. Use it before you let a proposal throw the repo away because the folders feel generated.
Your cloud, your repo, a way back
The code lives in your repository from the moment someone is paid to touch it. A ZIP on a laptop is not that. A project inside Bolt that one login can open is not that, once the login is a founder who might be on a plane. Your repo means a second person can be added without sharing a password, and a commit can be reverted without a prompt.
Your cloud means the database, the host, and the model key are in accounts your company can open. A preview URL is a convenience. It is also a single vendor, a single session, and a bill you may not see until it is interesting. Prefer accounts in your name. Put a spend alert on any key that calls a model before you send the URL to a customer.
A way back means you have deployed twice: forward, and back to the previous release, on purpose, before the day you need it. A rollback you have only discussed is a hope. Staging that points at the production database is not staging. It is a second door into the real rows. The production package includes a deploy on your cloud and the operational pieces around it because those pieces are the product a second user meets. They are listed on take your Bolt app to production: auth that holds for a second person, the schema in the repo, the core flow checked against a real record, a review of who can read and write every row, secrets off the browser, and staging that is not production.
What the call is allowed to decide
A single broken path, with the rest of the preview still the right product, can stay a repair. Say if you think that is you. Inflating a webhook into a new app helps nobody. Bring the path. We will tell you if it is really one path.
A short build fits when you are done with Bolt as a brief and you need one flow, one AI feature, and a live URL in your repo. It does not fit a second user, billing, and a pile of integrations. Bolt will have sketched all of those. The sketch is not a commitment that a short engagement includes the sketch. The two-week LLM MVP says this about generated drafts in general. It applies here without a special exception for Bolt. The call is where we refuse the sketch that does not fit.
The usual shape of "take this Bolt app to production," when a customer will depend on it, is a longer fixed build. Your brand if the audience is a customer, auth, billing hooks, a few integrations, an eval if there is a model in the path, and monitoring. If Bolt already wired five services, name the ones the first release needs. The rest wait. Pretending five is three is how the date slips. We will count them with you on the call.
A team that stays is what you book when production is the start of a roadmap, not the end of a migration. Do not book a team to finish a cutover that should have been a fixed build. Hire a dedicated AI team is the test for that choice. The test happens in the meeting, not in the prompt.
A boundary you can hand the next prompt
After the move, Bolt can still be useful. Low-risk UI, on a branch, reviewed by a person before it merges. That boundary should be written down: Bolt does not edit auth, payments, or the policies on a table. A tool with permission to "improve" those files will improve them back into the bug you just closed. The Lovable version of this boundary is in hire a developer to fix a Lovable app. The same discipline applies when the generator is Bolt. The logo on the tool is not the risk. The open-ended edit is the risk.
If nobody on your side can review a diff, do not leave the generator as a co-author of production. The handoff problem is an AI product your team can maintain. One person who can change a safe feature, and an engineer on auth and payments. If that person does not exist, the scope includes finding the shape of the handoff, not a fantasy that the prompts were the documentation.
A week that ends in one writer
The first days are the inventory and the freeze. We list every service the preview calls, revoke any secret that sat in the client, and mark a commit you can still click through. If that commit cannot be run outside Bolt, the first deliverable is making it run, not making it prettier. A prettier preview you cannot leave is the same product you arrived with.
The middle is the copy on your cloud, unmodified, then the dangerous paths changed on that copy. Two accounts. One real record or one test payment. Staging that has its own database. You click the rollback once, back to the unmodified deploy, then forward again. That drill is dull and it is the difference between a migration and a hope.
The last step is the domain, with Bolt still up only as a read-only fallback if you still need one. New writes go to one place. A week later you delete the fallback or you keep it dark on purpose. You do not "leave it for now" while both accept signups. The date of the cutoff is a sentence in the handover, not a memory.
What to bring
The Bolt URL and the export or the repo. The one flow you are showing people. The list of services the preview calls, including the ones you forgot: email, a model, a payment, a database, analytics. Two test accounts if you have them, and a description of what the second must not see. Any key that has ever been in the client, so it can be revoked rather than reused.
Leave the prompt that says "make it scalable." Bring the number of users you actually expect in the next month. How to scale a Lovable app is about the moment more users show up, and the lesson is the same for Bolt: scale is a second problem. A second user who cannot see the first user's data is the first problem. Do them in that order.
If this is your app, send the repo.
A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.
How Techtribe takes a Bolt app to production
We inventory the services, freeze a version that works, and put an unmodified copy on your cloud before we rewrite the dangerous paths. Secrets move off the browser and the old ones are revoked. Two accounts prove the data rule. The core flow is checked against a real record or a real payment. Bolt may stay up briefly as a fallback. It does not stay as the only deploy.
A narrow gap stays a repair when that is the honest size. A customer-facing cutover is a fixed build with an end. Work that continues is a team. We say which one after we have seen the preview, and we will say if the preview should not be saved. The build guide is Bolt to production. The decision if you are unsure the repo should survive is repair or rewrite. The service page is AI MVP development.
The scope call is 30 minutes with a founder. Bring the preview URL and the one flow. That meeting is where we say what the work is. This page is the reason to book it.
Questions
When should we stop prompting a Bolt app and hire someone?
Stop when the same bug returns after three prompts, a fix to one screen breaks another, or a second user can see data that belonged to the first. Another prompt rearranges the repo. It does not add a rule you can explain.
Can we keep the Bolt project and still go to production?
Keep the flow and the files that match it. Move secrets off the browser, put the database rules on the server, and deploy a copy you can roll back. Bolt can stay as a fallback for a few days. It should not stay as the only place the app can run.
Is a Bolt app a frontend, or a whole product?
Bolt generates a full-stack app in the browser from a prompt, and you can edit the files it wrote. The preview is not a second user, a real payment, or your cloud. Those are the production step, and they are a scope, not a prompt.
How do we know if Bolt needs a repair or a move?
Book the scope call and bring the preview. A narrow repair of one path stays a repair. A second user, your cloud, and the core flow checked against a real record is a production move. More product after launch is a team that stays. We say which one on the call, after we have seen the files.