Connect Claude to a Google Ads API and ask “what are my conversions this month?”
You’ll get a number. It’ll look right. You’ll feel smart.
Ask again tomorrow. Different number. Ask from a different prompt thread. Different number again. Change one word in the question. Wildly different number.
That is a slot machine, not a reporting system.
Vibecoding a dashboard is easy. Every founder with a Cursor subscription has done it by now. The actual job – the one nobody talks about – is vibecoding a dashboard that gives you the same trustworthy answer on Tuesday that it gave you on Monday. One you can hand to your CFO without a disclaimer.
This is the playbook Vision Labs uses in production. Five layers. Each one exists because skipping it breaks the system in a way you won’t catch until the wrong number hits a board deck.
What Are the Five Layers Every Vibecoded Dashboard Needs?
Every reliable reporting system – vibecoded or not – runs on the same stack:
- Raw – where the data comes from
- Storage – where you own and keep it
- Organization – the semantic layer that defines what words mean
- Visualization – how humans see it
- Action – what happens because of it
Most vibecoders jump straight from Raw to Visualization. That’s why their dashboards break. Here’s how to build each layer so yours doesn’t.
Layer 1: What Raw Data Sources Should Your Dashboard Pull From?
Every business runs on four categories of source data. You don’t need all of them on day one, but you need to know which ones feed your dashboard.
Ad platforms: Meta, Google Ads, LinkedIn, TikTok, AppLovin. These give you spend, impressions, clicks, and platform-reported conversions. Every platform counts conversions differently, so you can’t compare them raw.
CRM: HubSpot, Salesforce, GoHighLevel. This is where leads, deals, and pipeline live. CRM data tells you what happened after the click.
Product usage: Logins, feature adoption, session counts, activation milestones. If you’re SaaS, this is the data that tells you whether your customers are actually using what they’re paying for.
Sales and revenue: Subscriptions, one-off purchases, renewals, churn events. This is the money layer. If it doesn’t connect to the other three, your dashboard is a vanity project.
The mistake: vibecoding a dashboard that only pulls from one category. A Google Ads dashboard without CRM data can tell you cost-per-click. It can’t tell you cost-per-customer. Those are different questions with very different answers.
Layer 2: Why Should You Store Your Data in a Warehouse Instead of Querying APIs Directly?
Here’s the #1 vibecoding mistake we see: pointing Claude or Cursor directly at a source API and asking it to pull numbers.
Three things go wrong immediately:
- Different answers every run. APIs paginate, rate-limit, and time out. Run the same query twice and you get two different row counts. You won’t notice until someone else pulls the “same” report and gets a different total.
- AI invents its own definitions. Ask Claude “what’s my conversion rate?” without a schema, and it will decide what “conversion” means on the fly. It might use purchases in one thread and form fills in the next. Both answers look plausible. Neither is repeatable.
- No historical record. APIs serve current or recent data. Delete something from your CRM? It disappears from your reports retroactively. No warehouse means no audit trail.
The fix: store a copy of your data in a warehouse you control.
BigQuery is the default recommendation. It’s cheap (most companies under $50M spend under $100/month on storage), it speaks SQL so every tool connects to it, and it handles the volume. Snowflake and Redshift work too – but BigQuery’s free tier and native connection to Looker Studio make it the fastest path for a vibecoded stack.
The rule is simple: AI reads from your warehouse, never directly from the source API. Same query, same data, same answer. Every time.
Layer 3: What Is a Semantic Layer and Why Does Your Dashboard Need One?
This is the layer most vibecoders skip entirely. It’s also the layer that determines whether your dashboard survives contact with reality.
A semantic layer is a set of agreed-upon definitions that sit between your raw data and your reports. It answers questions like:
- What counts as a “lead”?
- What counts as an “active” user?
- When does an “impression” mean a Meta impression vs. a Google impression vs. a blended total?
- What does “revenue” include – gross bookings, net of refunds, or recognized only?
Without these definitions locked in, every person (and every AI agent) that touches your data will make up their own answers.
Real example from production: A client defined “lead” as “anyone who DM’d us on Instagram.” Six months in, they changed it to “anyone with an email or phone number on file.” Without a semantic layer, that change means rebuilding every report, every dashboard, every Slack alert that references “leads.”
With a semantic layer, you change the definition in one place. Every downstream report updates automatically. The dashboard from three months ago still works – it just uses the new definition going forward.
This is where the power is. A semantic layer turns your vibecoded dashboard from a fragile prototype into infrastructure.
How to build it: Start with a shared definitions doc. List every metric and dimension your dashboard uses. Write one plain-English sentence for each. Then encode those definitions as SQL views or dbt models in your warehouse. Every query, every dashboard, every AI agent reads from the views – never from the raw tables.
Layer 4: Which Tool Should You Use to Visualize Your Vibecoded Dashboard?
Once your data is stored and organized, you need to display it. Three options worth considering:
Looker Studio (Off-the-Shelf)
Connects natively to BigQuery. Drag-and-drop. Free. Your ops team can maintain it without touching code. Best for teams that want dashboards live this week and don’t need heavy customization.
Downside: limited interactivity. Complex layouts require workarounds. But for 80% of reporting needs, it’s the fastest path from warehouse to screen.
DIY HTML/CSS/JS (Vibecoded Custom Reports)
This is where AI shines. Claude and Cursor are excellent at generating clean HTML, CSS, and chart.js visualizations. You can vibecode a custom report page in an afternoon, host it as a static site, and wire it to your warehouse via a lightweight API layer.
These rarely break because HTML doesn’t have dependency chains. No npm updates. No framework migrations. A vibecoded HTML report from six months ago still renders exactly the same today.
Best for teams with a developer (or a technically comfortable founder) who wants full design control.
Hex and Agentic Reporting Tools
Hex, Count, and similar tools sit on top of your warehouse and let you mix SQL, Python, and narrative in a single notebook-style report. Some now include AI agents that can answer ad-hoc questions against your data.
Best for data teams that want exploration and presentation in one tool. Overkill if you just need a weekly scorecard.
The recommendation: Pick based on who maintains it. Non-technical team → Looker Studio. Technical founder → DIY. Data team → Hex. Don’t choose the fanciest tool. Choose the one that gets updated when the numbers look wrong.
Layer 5: How Do You Turn a Dashboard Into Actual Business Decisions?
A dashboard nobody checks is a dashboard that doesn’t exist. The visualization layer only works if it connects to action. Four modes, ranked by maturity:
Pull (Dashboards)
The human opens a dashboard and looks at it. This is where most teams stop. It works until the CEO forgets to check on Friday and misses a 40% spend spike over the weekend.
Push (Alerts)
The system sends a Slack message or email when something crosses a threshold. “Daily spend exceeded $5,000.” “Lead volume dropped 30% vs. last week.” No human has to remember to look. The system tells them.
Push is where most teams should start after building their first dashboard. One well-configured Slack alert prevents more bad decisions than a 12-tab dashboard.
AI Ask + Answer
An AI agent sits on top of your warehouse (through the semantic layer) and answers natural-language questions. “What was our blended CAC last week?” “Which campaign drove the most SQLs in Q3?” The agent queries the structured data, not the raw API, so you get the same answer every time.
FYI / Agentic Action (Where This Is Heading)
The system doesn’t just inform – it acts. It reallocates budget between campaigns, pauses an underperforming ad set, or drafts a report for the board. Then it pings a human: “I moved $2,000 from Campaign A to Campaign B because CPA crossed your threshold. Approve / deny / adjust.”
This is the frontier. Not fully autonomous AI, but AI that does the work and asks for a yes or no. The human stays in the loop without doing the manual work.
Where Is AI-Powered Reporting Headed Next?
The dream: raw data → Claude → human decision.
The reality: that setup fails without the semantic layer in between.
Point Claude at raw API data and it will hallucinate definitions, miss edge cases, and produce numbers that look right but aren’t. Point Claude at a clean warehouse with a semantic layer, and it becomes a reliable analyst that speaks your company’s language.
The rule: AI on top of structured data. Not AI on top of raw APIs.
Every team that skips the middle layers – storage, semantic definitions – ends up rebuilding from scratch within three months. The teams that invest in the stack first vibecode reliable reports for years.
When Should You Build the Full Five-Layer Reporting Stack?
Not every company needs the full five-layer stack today.
Below ~$10M ARR (or pre-scale): A Looker Studio dashboard connected to a Google Sheet is fine. Seriously. If your data fits in a spreadsheet and one person owns the numbers, you don’t need a warehouse yet. Build that when the spreadsheet starts lying to you.
Above ~$10M ARR, or venture-backed and growing: You need the full stack. Multiple ad platforms, a CRM with thousands of records, product data that changes daily – spreadsheets can’t hold this, and a single API integration can’t unify it. Build the warehouse, define the semantic layer, and wire your dashboards to both.
The in-between: If you’re spending $50K+/month on ads across more than two platforms and your team argues about which number is right, you’re already past the threshold. The cost of wrong data is higher than the cost of building the system.
How Do You Start Building a Reliable Dashboard Today?
Act on this today:
- List your sources. Write down every platform you pull data from: ad platforms, CRM, product analytics, payment processor.
- Pick a warehouse. BigQuery free tier is live in minutes. Create a project. No data yet – just the container.
- Define five terms. Write one sentence each for: lead, customer, conversion, revenue, active user. These are your first semantic layer.
- Load one source. Pick your highest-spend ad platform. Use Fivetran, Airbyte, or a simple scheduled script to load raw data into BigQuery daily.
- Build one dashboard. Connect Looker Studio to BigQuery. Build a single chart: daily spend vs. conversions (using your definition). Ship it.
- Set one alert. Configure a Slack or email notification for one threshold: daily spend exceeds X, or lead volume drops below Y.
Six steps. You can finish the first three in an afternoon. The rest take a few days. After that, you have a reporting system, not a slot machine.
Frequently Asked Questions
Can I vibecode a dashboard without a data warehouse?
You can, and it will work on the day you build it. The problem is next week. Without a warehouse, your dashboard pulls from live APIs every time it loads. API responses change, rate limits hit, and historical data disappears. A warehouse gives you a frozen, auditable copy of every data point. That’s what makes the numbers repeatable.
What’s a semantic layer in plain English?
A shared dictionary for your data. It answers: “When we say ‘lead,’ what exactly do we mean?” Without it, every person and every AI agent makes up their own definition. With it, “lead” means the same thing in your dashboard, your Slack alert, and your Claude conversation. You define it once, and everything downstream uses that definition.
Why does Claude give different answers when I ask the same question twice?
Two reasons. First, if Claude hits a live API, the underlying data may have changed between requests (new rows, deleted records, different pagination). Second, without a schema or semantic layer, Claude decides how to interpret your question on each run. “Conversions” might mean purchase events in one thread and form submissions in the next. A structured warehouse with defined views removes both problems.
Looker Studio vs. Hex vs. DIY – which should I pick?
Looker Studio if your ops or marketing team maintains reports and nobody writes code. DIY HTML/CSS/JS if a developer or technical founder builds reports and wants full control over layout and design. Hex if you have a data team that needs to explore data and build reports in the same tool. Match the tool to the person who fixes it when something looks wrong.
When is vibecoding a dashboard a bad idea?
When you don’t have enough data to justify a system. If you’re pre-revenue, running one ad platform, and checking numbers once a week, a spreadsheet is the right tool. Vibecoding a full dashboard stack becomes worth it when you have multiple data sources, multiple people consuming reports, and real money riding on the numbers being right. For most companies, that threshold sits around $10M ARR or $50K+/month in ad spend.
Vision Labs builds the full five-layer stack – data collection, warehouse, semantic layer, dashboards, and push/pull/agentic actions – in 30- and 60-day sprints. If your team is past the spreadsheet stage and needs a reporting system that holds up week after week, start here.
