The term Vibe Coding was popularized by Andrej Karpathy¹ to describe a conversational, high-level way of programming with AI models, where you specify the intent, style, or 'vibe' of what you want and let the model draft most of the implementation details of a Web app.

No code, just words.

At Being Guided, Vibe Coding is the third of three, interrelated elements of digital innovation: Think; Simplify; Vibe.

As illustrated above, Vibe Coding here is powered by Salesforce, Google or Microsoft technology and driven by Design Thinking based on the Stanford d.school² method. This is five iterative stages to create a solid Think foundation for digital innovation. See Journal post: Think.

As readers probably know, there are a bewildering number of startups that have entered the Vibe Coding marketspace: Base44⁸, Bolt⁹ and Lovable¹⁰, among many. In turn, these new startup challengers, although in receipt of significant funding valuations (Lovable valued at $6billion¹¹), rely on Supabase¹² as their infrastructure.

At the time of writing this Journal post (January 6, 2026), I have prioritised Vibe Coding on the platforms delivered by the global market leaders: Salesforce, Google and Microsoft. This is due to a limitation with Vibe Coding startups, where their 'Stochastic Models' exhibit the inconsistency of Large Language Models (LLMs). Although these tools work well as Prototypes, they need a further 'Optimisation' step to become Production-ready 'Deterministic Models^{13, 14, 15}.

Salesforce, Google and Microsoft have all combined a broader set of technologies and mature infrastructures to enable a successful transition from Vibe Coding a Prototype to Optimising the Web apps in a Build phase for secure, scalable Production outcomes.

Vibe Coding with Google Firebase Studio.

Today, Vibe Coding is an effective way to Prototype new Web apps. It is literally no code, just words. Prompts entered into the right hand panel of Firebase Studio result in a Preview of the Prototype Web app in the left hand panel.

Citrizen CRM Vibe Coded app, Optimised as a Production Web app.

Build.

The Build process takes a Vibed Coded Prototype and turn this into a Production class Web app on eith Salesforce, Google or Microsoft technology.

Salesforce Vibe.

With Salesforce, Vibe Coding is offered on both Salesforce Lightning Platform (Vibe Codey) and on Heroku (Heroku Vibes), both providing the high-level intent and natural language description, as the AI handles the technical boilerplate (scaffolding, metadata, and code).

Vibe Codey.
This is the autonomous coding agent within the Agentforce Vibes platform. It acts as a pair programmer that understands your specific Salesforce metadata. If you describe a workflow, Vibe Codey generates the Apex classes, Lightning Web Components, and tests while adhering to your Salesforce Org existing security guardrails.

Heroku Vibes.
This extends vibe coding to the broader app ecosystem. It allows developers to describe an entire application's requirements in plain language, which Heroku then auto-provisions—creating the database, setting up the runtime, and deploying the app without the developer manually configuring the infrastructure.

Google Vibe.

In moving from Vibe to Build lets look at Google tech there is an intermediate step: Optimise. This is working with the combination of Firebase Studio and Google Cloud Platform (GCP).

Optimisation Introduced.

The core objective of Optimisation is to act as the 'Production Hardening' layer in Vibe Code apps conceptualised with Google Firebase Studio and then moved to a Production Web app running on Google Cloud Platform (GCP).

The role of Optimisation is also to ruthlessly eliminate variable costs to ensure economic viability. It is, in the jargon, a human-in-the-loop taking a 'Stochastic Model' inherent in a Vibe Coded Web app and turning this into a robust 'Deterministic Model' required for demanding Software-as-a-Service (SaaS) use cases, such as Customer Relationship Management (CRM) Web apps.

The Solution: Optimisation systematically replaces expensive, probabilistic AI inference calls with cheap, deterministic cloud-native functions (Firestore Aggregations, Edge Caching, and Indexing). This will stabilise the Monthly Active User (MAU) cost at approximately £0.10 per user (based on the £100/month for 1,000 users model), ensuring the solution avoids what might be called ‘the Stochastic Model Trap’ of Vibe Coding leaduing to ballooning consumption costs.

Costs Reduced.

As noted above, when delivering the Production Web app on Google Cloud Platform, with an an intermediary Optimise step, The cost differences here are significant.

Vibe Coding (via Firebase Studio) defaults to ‘Reasoning’ (using Gemini 3 Pro) to solve data problems because it is the fastest way to build. However, at runtime, this incurs a massive ‘AI Tax’ (could be as high as $700/month for 1,000 users) and introduces latency.

Optimisation systematically replaces expensive, probabilistic AI inference calls with cheap, deterministic cloud-native functions (Firestore Aggregations, Edge Caching, and Indexing). This will stabilise the Monthly Active User (MAU) cost at approximately £0.10 per user (based on the £100/month for 1,000 users model).

Performance Enhanced.

In the Prototype Web app, a user asks "How many tickets are open?" Gemini reads 50 records and counts them, so costs are high if used in Production. So, Optimise replaces the natural language query handler with Firestore Aggregation Queries, resulting in cost optimisation.

Optimise identifies Summary Components in the Vibe Code. Refactor to use count(), sum(), and average() serverside calculations. This reduces 50 Read operations + 1 AI Call \to 1 Aggregation Query cost (approx. 1/1000th of a penny).

If say, 1,000 users loading the same Departmental Updates page generates 1,000 Reads every time the page refreshes. Optimisation provides a Smart Cache (Read Reduction). Optimisation reduces Bandwidth and Read costs by up to 90% for high-traffic, low-change data.

Optimise applies a 'deleteAt timestamp' to all temporary data (eg. Audit logs, completed task notifications). Firestore native TTL policy automatically deletes these records at T+X days without incurring Delete operation costs. This significantly reduces Web app storage costs with higher user numbers. This keeps costs within the GCP Free Tier for storage for as long as possible.

Security Enforced.

From a security perspective, Optimise injects the following security posture:

Firestore Security Rules Engine: This is creating granular, role-based rules and ensures strict multi-tenancy isolation (Data Sovereignty).

Input Validation: Optimise wraps all API endpoints with Zod validation libraries. This prevents 'Garbage In' and protects against injection attacks, ensuring that even if Gemini hallucinates a code snippet, the data entering the DB is strictly typed. (Zod Schemas)

Solution Delivered.

Prototype: Use Gemini 3 Pro to build the features and UI fast. Get stakeholder buy-in.
Audit: Run the Vibe Code through Optimise: Identify every generateText() call.
Refactor: replace generateText('Summarize this list') with db.collection('list').count().
Move: logic from 'Clientside AI' to 'Serverside Typescript'.
Deploy: Push to Firebase Hosting with the optimised backend.

Microsoft Vibe.

Organisations that have extensive investments in Microsoft technology will find Vibe Coding with Power Apps⁵ attractive. At Being Guided we are creating CRM Prototypes with Power Apps, AI Builder⁶ and Dataverse⁷.

For a modern Web app, Microsoft Vibe has three pillars functioning as a single, cohesive unit:

Dataverse (The Foundation).
It is not just a database; it is a smart data service. It stores the tables, logic, and security rules that the application relies on. When you Vibe Code, using natural language to build, Dataverse automatically structures the metadata based on your prompts.

AI Builder (The Brain).
This component adds 'intelligence' to the Web app. It can extract text from documents, predict outcomes, or categorise sentiments. In Vibe Coding, AI Builder models are often invoked as functions directly within the app's logic.

Power Apps (The Canvas).
This is where the user interacts with the Web app. It consumes the data from Dataverse and triggers the AI models from AI Builder. Through Copilot, Power Apps allows developers to describe a UI and it automatically binds the UI to the underlying Dataverse tables.

References

1. Karpathy, A. [@karpathy]. (2025, February 2). There’s a new kind of coding I call “vibe coding”, where you fully give in to vibes, embrace exponentials, and forget that code even exists. X. https://x.com/karpathy/status/1753472166197080428
2. The Hasso Plattner Institute of Design. (2004) Stanford d.school. https://dschool.stanford.edu/about
3. Kilpatrick, L. (2025, November 18). Start building with Gemini 3. Google. https://blog.google/technology/developers/gemini-3-developers/
4. Google. (2025). Get started with Firebase Studio. Firebase. https://firebase.google.com/docs/studio/get-started
5. Microsoft. (2025, June 24). Embedding AI in Power Apps: Advanced customization. https://www.alphabold.com/ai-integration-in-power-apps-for-enhanced-decision-making/
6. Microsoft. (2025, August 14). AI Builder documentation. Microsoft Learn. https://learn.microsoft.com/en-us/ai-builder/
7. Microsoft. (2026, January 9). What is Microsoft Dataverse? Microsoft Learn. https://learn.microsoft.com/en-us/power-apps/maker/data-platform/data-platform-intro
8. Base44. (2025, November 3). A simple guide to how it works. Base44 Blog.
   https://base44.com/blog/vibe-coding
9. StackBlitz, Inc. (2025, October 2). The vibe coding revolution enters its next phase: Introducing Bolt v2. Bolt Blog.
   https://bolt.new/blog/bolt-v2
10. Lovable. (2025, December 4). GPT Engineer and Lovable: The evolution. https://lovable.dev/gpt-engineer
11. Hope, G. (2025, December 22). AI coding startup Lovable now valued at $6.6B. AI Business. https://aibusiness.com/generative-ai/ai-coding-startup-lovable-series-b
12. Supabase. (2026, January 9). Database: Getting started with PostgreSQL. Supabase Docs. https://supabase.com/docs/guides/database/overview
13. AICerts. (2025, December 26). Salesforce deterministic AI pivot boosts Agentforce reliability. AI CERTs News.
    https://www.aicerts.ai/news/salesforce-deterministic-ai-pivot-boosts-agentforce-reliability/
14. GitGuardian. (2025, June 16). Prevent vibe coding security vulnerabilities with automated guardrails. GitGuardian Blog.
    https://blog.gitguardian.com/automated-guard-rails-for-vibe-coding/
15. Microsoft Learn. Sarkar, A., & Drosos, I. (2025, November 13). Vibe coding and other ways AI is changing who can build apps and how. Microsoft Source. https://news.microsoft.com/source/features/ai/vibe-coding-and-other-ways-ai-is-changing-who-can-build-apps-and-how/