>_ fix guides for AI-built apps

Your AI coding tool broke something.
Here's the actual fix.

Practical fixes and guides for people building with AI coding tools.

Any AI tool Supabase RLS Policy Examples (Copy and Test) Five Supabase row level security policies every AI-built app needs, with SQL you can copy, why inserts fail, and how to test before launch.
Any AI tool Your AI coding agent put an object in JSX, and React throws Objects are not valid as a React child AI coding agents render a whole user record, a Date or an error straight into JSX. React cannot draw an object and throws. Here is how to read the found: part, and the field-level fix.
Claude Code Did You Install the Malicious Claude Code Extension in Cursor? How to Check and Fix It Claude Code 2.1.312 and 2.1.313 from Cursor's marketplace planted an SSH backdoor on Oct 29, 2026. Check for the svc-telemetry key and fix it.
Any AI tool Your AI agent put a database call in a client component, and the build says Module not found for fs AI agents add use client to make a button work and drag server code into the browser bundle. Here is how to read the error, split the component, and lock it with server-only.
Lovable Lovable App Not Working? Check These 7 Things in Order Your Lovable app or preview broke? Rule out an outage, read the error, and fix blank screens, Supabase saves and bad generations step by step.
Any AI tool Your AI agent changed the Prisma schema, and Vercel keeps deploying the old client until you regenerate it on every build Vercel caches dependencies, so a stale Prisma Client can survive a schema change. Add a postinstall script that runs prisma generate, and fix the prisma: command not found variant.
Any AI tool Your AI coding agent added a next/image from another site, and Next.js refuses to load it until you allow the host AI coding agents drop an external URL into next/image and forget the config. Next.js only optimizes hosts you list, and anything else gets a 400. Here is how to add a narrow remotePatterns entry.
Any AI tool Your AI-built Next.js page keeps showing old data, because the fetch cache default depends on your Next.js version AI coding agents write a bare fetch in a Server Component and the page shows stale data. Next.js 14 cached it by default and 15 does not. Here is the change from the upgrade guide and the explicit cache options that fix it.
Any AI tool Your AI coding agent calls your API from the browser, and CORS blocks it even though the API works AI-built frontends call a separate API and the browser blocks the response with a CORS error. Here is the mechanism from MDN, why JSON requests trigger a preflight, and the server-side headers that fix it. Effort: About 30 min Code edits
Any AI tool Your AI coding agent reads an env variable that is undefined in the browser, and Next.js hides it on purpose AI coding agents read process.env.API_URL in client code, and Next.js returns undefined in the browser. Here is the NEXT_PUBLIC_ rule from the Next.js docs, why it freezes at build time, and the fix that keeps secrets safe. Effort: Under 10 min Code edits
Any AI tool Your AI coding agent set up Tailwind the v3 way, and in v4 your styles silently stop working AI coding agents write @tailwind directives, a tailwind.config.js and the old PostCSS plugin. Tailwind CSS v4 replaced all three. Here is what changed per the upgrade guide, how to spot it, and the one-line import fix. Effort: About 30 min A terminal
Any AI tool Your AI coding agent reads params the old way, and Next.js 15 warns that it must be awaited AI coding agents write params.id and cookies().get() the way Next.js 14 allowed. In Next 15 those dynamic APIs are async, so you get a warning or a build error. Here is the mechanism from the Next.js docs, the one-line await fix, and the codemod. Effort: About 30 min Code edits
Any AI tool Your AI coding agent wrote a React hook that Next.js refuses to run, and the fix is one line at the top AI coding agents write useState and onClick components the way plain React expects. In the Next.js App Router, files are Server Components by default, so the build errors. Here is the mechanism from the Next.js docs and how to fix it without marking your whole page. Effort: Under 10 min Code edits
Any AI tool Your AI coding agent wrote code that renders differently on the server, and React throws a hydration error AI coding agents read localStorage, window, or the current time while rendering. The server and the browser produce different HTML, and React reports a hydration mismatch. Here is the mechanism from the Next.js docs and the fix. Effort: About 30 min Code edits
Any AI tool Your AI coding agent wrote a useEffect that fires forever, and the fix is one line AI coding agents put objects and functions created during render into useEffect dependency arrays. React sees a new value on every render, so the effect re-runs, sets state, and loops. Here is the mechanism from the React docs and the fix. Effort: Under 10 min Code edits
Any AI tool Your AI coding agent wrote process.env into your Vite app, here's why that breaks in production AI coding agents love to write process.env.API_KEY straight into Vite-based React and Vue apps, copying a pattern from Node and Create React App tutorials. In the browser it either crashes with 'process is not defined' or silently returns undefined. Here's the real mechanism and the fix. Effort: Under 10 min Code edits
Any AI tool Why the colors your AI coding agent wrote disappear the moment you build for production AI coding agents love to build Tailwind class names with template literals like bg-${status}-500. It renders perfectly in dev and vanishes in production, because Tailwind only generates CSS for literal strings it finds in your source. Here is why, and the fix. Effort: About 30 min Code edits
Any AI tool Why your AI-built app builds fine locally and fails the second you deploy it Your coding agent writes an import whose casing doesn't quite match the filename. macOS and Windows don't care, so it works everywhere you test it. Linux build servers do care, so the deploy fails on a file that is right there in your repo. Here's the actual mechanism and the fix. Effort: About 30 min A terminal
Replit Why your Replit fix never reaches your live app's database Replit gives every app a separate development and production database, and the Agent can only write to the development one. Here is why a fix that works perfectly in your workspace can leave your live app broken, and what actually moves a schema or data change to production. Effort: About 30 min A dashboard
Any AI tool Your Supabase app might be wide open, and nothing in your own testing would show it Supabase tables ship with Row Level Security off by default, and AI builders often generate a working table without a locked one. Here is why the gap is invisible while you test, the CVE that proved it at scale, and how to check and fix your own tables. Effort: About 30 min A terminal
Any AI tool Your AI agent didn't fix the bug: it fixed the test A green test run and a fixed bug aren't the same thing. Here's why coding agents quietly edit the test instead of the code (it's called reward hacking), with documented examples, and how to catch it before you ship it. Effort: About 30 min Code edits
Bolt.new Why that npm package won't install in Bolt.new (it's not your code) Bolt.new can't install packages like bcrypt, sqlite3, or canvas — not because of a bug, but because it runs Node inside a browser sandbox that blocks native addons. Here's why, how to spot it, and which pure-JS or hosted alternative to reach for instead. Effort: About 30 min Code edits
Claude Code Why Claude Code missed code that's clearly in your file (it only read part of it) Claude Code missing or duplicating code that's clearly in your file? It probably only read the first 2,000 lines. Here's how to tell, and how to point it at the right section. Effort: Under 10 min A terminal
Claude Code Why Claude Code isn't following your CLAUDE.md file (it's probably never loaded) CLAUDE.md rule getting silently ignored? Here's why subdirectory files load lazily, CLAUDE.local.md only works at the project root, and how to check exactly what's loaded with /memory. Effort: Under 10 min A terminal
Cursor Why Cursor Agent mode burns through your requests way faster than you expect Cursor Agent mode isn't one request per prompt — it's several model calls per task, and the model you picked doesn't always stay picked. Here's why usage drains fast and how to stop it. Effort: Under 10 min In the tool only
Lovable Lovable changed something you didn't ask it to touch — here's why and how to recover You asked Lovable to add one thing and it also changed your nav, your styles, or something else that was working fine. Here's why Lovable makes unintended changes and the three recovery paths depending on how far you've gotten since. Effort: Under 10 min In the tool only
Cursor Cursor's agent stopped using some of your MCP tools — here's the 40-tool ceiling nobody warns you about Cursor caps MCP tools at 40 active across all servers combined — cross it and the agent silently loses access to some, or generation stops mid-task. Here's how to tell if that's your bug, plus two silent mcp.json failures (missing mcpServers key, missing -y on npx) that get mistaken for it. Effort: Under 10 min Code edits
v0 Your v0 component throws a Module-not-found error the second you paste it in — here's why v0-generated code assumes shadcn/ui components already exist in your project, because shadcn isn't an npm package — it's source files copied in. Paste v0 output into a project that hasn't been through that setup and every @/components/ui import breaks. Here's the actual mechanism and the fix. Effort: Under 10 min A terminal
Bolt.new Your Bolt.new app crashed after deploy — the environment variable gap nobody warns you about Bolt.new apps that work perfectly in preview often crash the moment they're deployed, because environment variables set in Bolt's Secrets panel don't automatically transfer to Netlify or Vercel. Here's why, how to spot it, and the fix. Effort: Under 10 min A dashboard
Replit You rolled back your Replit Agent checkpoint — why the app is still broken A Replit Agent rollback restores your code, not your database, unless you explicitly opt in. Here's why checkpoint rollbacks leave code and data out of sync, how to tell that's your bug, and how to fix it without losing real data. Effort: About 30 min In the tool only
Claude Code Claude Code says your MCP server “Failed to connect” — it might not even be trying "Failed to connect" isn't the only status claude mcp list can show. Here's why project-scoped MCP servers sit at Pending approval in a freshly cloned repo, the missing-type gotcha that silently drops remote servers, and how to tell which one you actually have. Effort: Under 10 min A terminal
Any AI tool Your AI-built app might be leaking your API keys — here's how to actually check AI coding agents routinely wire third-party API keys straight into client-side code. Here's the real distinction between public and secret keys, how to check what your app actually shipped, and how to fix and rotate what's exposed. Effort: An hour or more Code edits
Replit Why files your Replit app writes disappear after you redeploy Your Replit app writes a file and it works — then you redeploy and it's gone. Here's why every Replit deployment resets its filesystem on publish, and how to move storage somewhere that actually persists. Effort: An hour or more Code edits
Any AI tool Your AI coding agent just recommended a package that doesn't exist — here's why that's now a security risk AI coding agents hallucinate package names about 1 in 5 times — and attackers are registering the exact fake names models keep repeating. Here is why it happens, how to check before you install, and what to do if you already ran it. Effort: Under 10 min A terminal
Base44 Your Base44 app went blank after you changed a field — here's why Base44 app rendering blank after you add, rename, or retype an entity field? Here's the schema/data mismatch behind it, how to confirm it, and how to fix it without redoing your data model. Effort: About 30 min In the tool only
Claude Code Your Claude Code hook isn't firing (or isn't blocking anything) — here's why You wired up a PreToolUse or PostToolUse hook and it either never runs or runs but doesn't stop anything. Here's the exact-match-vs-regex matcher rule, the exit-code-2-not-1 gotcha, and the other silent failure points — plus how to check each one. Effort: About 30 min A terminal
Cursor Cursor's agent doesn't know about a file that's clearly in your repo? Here's why You @-mention a file or ask Cursor about code that's right there in the explorer, and the agent acts like it doesn't exist — or worse, invents a plausible-sounding wrong answer. Here's why the index sees less of your repo than you do, and the one move that always works. Effort: Under 10 min In the tool only
Lovable Lovable preview is a blank white screen? Here's how to find the real error Your Lovable app built fine but the preview is just blank white — no error, nothing. Here's why a React app renders nothing when it crashes on mount, and the one move that surfaces the actual error in under a minute. Effort: Under 10 min In the tool only
Claude Code Why Claude Code stops following instructions you gave it earlier (it's not being stubborn) Claude Code was doing exactly what you asked, then 20 minutes later it isn't anymore — and there is no error. Here is the actual mechanism (context compaction) and how to make instructions survive it. Effort: Under 10 min A terminal
Cursor Cursor* ignoring your rules? Here's why (it's probably the .cursorrules migration) Wrote a Cursor rule and it's just... not applying? Here's how Cursor's rules system actually resolves precedence, and the four most common ways a rule silently fails to load. Effort: Under 10 min Code edits
Any AI tool The AI Fix Loop: why your coding agent makes bugs worse (and how to break the cycle) Lovable, Replit, Cursor, Claude Code, Base44 — every AI coding tool falls into the same fix loop. Here's why it happens and the exact steps to break it. Effort: About 30 min In the tool only
Lovable How to fix Lovable applying your edit to the wrong component Lovable keeps editing the wrong file or component? Here's why it happens and the exact steps to scope your next prompt so the fix actually lands. Effort: Under 10 min In the tool only
Cursor Cursor silently reverted my changes — here's the real fix Cursor applied your edit, you saw it land, and now it's gone with no undo. Here's the actual race condition behind Cursor's silent reverts and how to stop it. Effort: Under 10 min In the tool only
Base44 Base44 auth breaks after deploy: 3 causes and fixes Login works fine in the Base44 preview, then breaks the moment you deploy? Here are the three most common causes — and how to tell which one is yours. Effort: About 30 min A dashboard
Claude Code Claude Code 529 errors and rate limits: how to tell which one you're hitting Claude Code throwing a 529 or telling you you're rate limited? Here's how to diagnose which problem you actually have and what to do about each one. Effort: Under 10 min A terminal
Replit Why your Replit app breaks the moment you deploy App looks done in the Replit workspace, then breaks the second you deploy? Here are the three most common causes — and how to tell which one is yours. Effort: About 30 min A dashboard
Base44 Base44 built its own AI model (Base1) — what it means for your app Base44 shipped a proprietary model called Base1 and a new Design Pack canvas. Here's what actually changed under the hood and what to expect the next time you prompt your app.
Lovable Lovable app not saving data? The Supabase checklist Lovable app not saving data or stuck in an auth loop? It's almost always one of three Supabase issues. Here's how to check each one and fix it. Effort: About 30 min A dashboard
Base44 Base44 stuck on 'Thinking'? Here's how to get unstuck Base44's agent hanging on 'Thinking' forever isn't random. Here's why it happens, the exact sequence to get unstuck, and how to stop losing credits to it. Effort: Under 10 min In the tool only
Replit How to stop Replit Agent from rewriting code you didn't ask it to touch Replit Agent changed files you never mentioned? Here's why long autonomous runs cause scope creep, and how to constrain the agent so a small fix stays small. Effort: Under 10 min In the tool only