A client dashboard that reports signals, not charts

The dashboard is an HTML product, not a data dump

The middle-of-funnel project in the course is data collection and tracking: UTMs, dashboards, and the connective skills between them 4:36:37. A UTM parameter is a tag appended to a URL so analytics can attribute a visit to a specific campaign or channel 4:36:50. The naive approach most people take is dumping every export into Claude Code and saying "visualize it" — which produces dashboards the speaker calls "absolutely ass to look at" 4:36:50. The fix is to stop treating the dashboard as a data-showcase tool and treat it as a web design project: a modern dashboard is essentially an HTML page 4:37:28. The reference build, a hypothetical agency called Northbeam Digital, tracks total ad spend, sessions, leads, blended CPL, and closed-won deals on one page 4:37:28.

Marketing data lives across Meta ads (spend and clicks), Google Analytics 4, Stripe (revenue), CRM, and email platforms, so a unified dashboard needs an ingestion pipeline — connectors that pull from every platform into one dashboard build 4:37:28. That pipeline runs on a daily, weekly, or monthly schedule to create a single source of truth where all of an organization's metrics live instead of disconnected per-platform reports 4:38:58. The speaker contrasts this with the old agency model, where a monthly "report day" produced polished reports used almost as sales mechanisms; auto-ingesting dynamic dashboards replace that pain 4:38:58. The build flow: read all API docs on the platforms, write the tests and queries, generate a dashboard page, and refresh it with a cloud routine on a schedule 4:40:29. The dashboard also gets an AI assistant so a client can query the data locally on the page 4:40:29.

Flat exports stand in for a production database

To avoid exposing proprietary or private real stats, the demo assembles hypothetical Facebook ads, website visits, sales data, salesperson logs, and email stats as fake-data CSVs — CSV being comma-separated values, effectively a Google-Sheet-style flat data file 4:42:01. Step one of the build is assembling all the marketing data files that will stand in for the databases 4:43:32.

The folder is a mix of dimension tables — channels, clients, sales reps — and fact tables — activities, ad creatives, agency financials, attribution paths, backlinks, budget pacing, call tracking, channel daily, client summaries, and monthly time series 4:43:32. The fact-activities file is a time log of reps' total calls and emails replied, usable for measuring rep productivity 4:44:15. The speaker filters it by activity date, clears all values except rep one, and sums total logged seconds: 2,940 — that is 49 minutes of activity, on the demo's "today" of August 6th 4:44:40 4:45:05. The GA4 file is 6.1 megabits with over 10,000 rows — client ID, channel, country, sessions, engaged sessions, bounce rate, pages per session — while the activity data runs to roughly 50,000 rows 4:45:05.

Let the strongest model find the structure the data forces

The speaker opens Claude Code, attaches the "marketing dashboard data" folder, and feeds it a brief description of all the data plus the folder itself 4:46:36. The prompt asks Claude to go through all the data, find ways to combine the databases logically, create a high-level structure, scaffold pages simply, and provide a few candidate structures 4:46:36.

Model intelligence decides how good this analysis gets. Most of the course ran on Sonnet 5, which the speaker calls a "last generation," "not very intelligent" model; for high-quality design he uses the best models available — Opus 5 or Fable 5 — to save time and energy 4:47:30. On a Pro plan he buys a one-time $20 block of usage credits, caps monthly usage at $20, and turns auto-reload off (not needed on Max plans) 4:47:40. Re-running the structure analysis on Fable 5 at high reasoning, he prompts: "I just swapped to Fable 5. I'd like you to revisit these — assume the analysis done before was by a dumber, less capable model" 4:48:07.

Fable 5 at high reasoning finds three properties of the data that force the structure: the anomaly distribution dictates the primary axis (client-scoped); coverage gaps — e-commerce data exists for only 4 of 12 clients and attribution for only 44% of leads — limit what can be shown; and grain mismatches dictate what cannot be client tabs 4:49:39. The recommended structure: overview, portfolio grid, alert feed, performance, traffic, pipeline, search presence, commercial, client-by-ID, channels, sales teams, and signals 4:49:39. Opus 5 at medium reasoning, by contrast, proposed a funnel-oriented structure (overview, acquisition, traffic and site) versus a client-centric one (agency overview, client detail, channel analysis), plus persona questions like "Are we hitting the number? Where's the money going? Is the site converting?" 4:50:30. The speaker points to "Fable Design 25," a showcase where Fable produced extremely high-quality, well-designed sites — the point being that design need not be constrained when the model is approached in the specific order: first get the data on the page, then "skin" it and select high-quality iterations 4:51:10 4:51:40.

The front page names anomalies, then shows the charts

Threshold detection is simple SQL logic: flag a metric as anomalous when it jumps or collapses faster than a chosen threshold over a period — for example, a ~5x difference 4:52:00. The Fable-5-built Northbeam Digital agency overview puts leads per day beside a client portfolio and active traffic-collapse signals 4:52:15.

Opening the Signals dashboard names the anomaly outright: a lead-flow collapse for CL-010 in the week of March 9, 2026 — 3 leads versus a typical weekly 57.8 — flagged via that ~5x threshold 4:52:41.

The same logic surfaces a conversion-rate jump for CL-005 — 2.23% against a 1.65% baseline — and a rep win-rate outlier: REP-05 Cole Baxter at 15.6% versus the 24.2% team average 4:53:15 4:53:20. The speaker claims the Fable 5 dashboard session cost about $13 — 94% of the usage limit — for a dashboard controlling ~20 clients and tens of thousands of rows, and the reason to spend it on Fable 5 is to get the scaffold and architecture, then spin it into variants 4:54:12. For an agency, this signal feed is the client-facing proof of monitoring — the reason the dashboard is a retainable product rather than a monthly export.

Five agents, one brief: search the solution space in parallel

The speaker switches from Fable 5 to Opus 5 at high reasoning ("extra high" is not available on Opus 5) and asks it to spawn five sub-agents all working on the project simultaneously in HTML, with key focus on interpretability, usability, and practical UX 4:55:43. The math favors parallelization: serial execution of 4 steps at 5 minutes each is 20 minutes; spawning parallel agents on steps 2–5 and merging into a final step 6 takes only ~10 4:57:14. A parallelizable process is one where agents work on the same data resource but their actions are independent and do not impact one another 4:57:40. Pros: much faster, and a broader search of the solution space; cons: more expensive, and because each process must be non-monitored, the risk of failure is higher — acceptable because only one of several variants needs to succeed 4:58:45.

The five subagents each produce a distinct variant — editorial, terminal, Swiss, dense-operations, and warm-print — and the brief requires every variant to find the same seven anomalies 4:59:58. A "bare bones," sleek reference scaffold keeps the structure honest 5:00:22; the background tasks panel shows the main agent "Northbeam" plus the five variant agents, each with transcripts and visible tools 5:01:05. The agents finish at staggered remaining times as the usage limit is approached 5:01:46. The dense-ops variant looks far better in light mode: overview, channels across clients, sales metrics (24.1% team win rate, 1,490 deals won, $39.8M revenue closed), and signals for wins with conversion-rate jumps and losses with traffic collapse; clicking a client exposes performance, traffic, pipeline, rep performance, and search 5:03:18. The terminal-style variant allows drill-down from LinkedIn ads through Google search with filtering and smooth-scroll numbering (1–6) for presentations; the portfolio-style variant shows active/churned/health/MRR on a 30-day timeline 5:04:48. The speaker claims a company making roughly $20–30 million a year took one of these dashboards and put it on office televisions as a slideshow so responsible reps constantly see the metrics 5:03:45.

Deploy it, gate it, and let the front page name anomalies

To put the dashboard on the internet, the speaker uses Netlify, which hosts static content for virtually free — a "free forever" plan that deploys far more sites a month than needed 5:05:50. The connection is a remote MCP server: settings → connectors → add custom connector, name it "netlify," paste only the connection info (not the npx sections), and authorize it 5:06:18. The instructor hit the usage limit mid-setup and bought another $15 of credits — unlikely for viewers, since the course was recorded in a single day 5:06:50. The speaker's voice-dictated deploy prompt works as a template: "My goal is I want to get this on the internet. My favorite was V3. I just connected the Netlify MCP. I would like you to add this and upload this to Netlify and then password protect it. Make the password 1 2 3 4 5" 5:07:18.

The finished app ships light and dark modes and gates per client via URL parameters and special passwords 5:07:30. It generates a health score per account — pipeline, funnel, drop-off points — and flags dead zones, like a CPL bump from an average of about $100 to literally $1,100 5:07:50. The critical example is Torque Auto Group (CL-010): health score 62, the same account flagged on the week of March 9th with 3 leads versus its usual 58 and CPC spikes 5:08:30. The speaker claims the whole build recreated agency-analytics-style marketing visualization software for about $15 in tokens — with 50,000 to possibly 100,000 rows of data, not just a Google sheet 5:07:50. The speaker also claims the entire dashboard was assembled in roughly 30 minutes even while rate-limited, and frames the economics against a monthly spend of $14,000–$15,000 on AI agents: blasting everything with Fable 5, spawning multiple sub-agents, and using fast mode's ~2x speed multiple delivers "extraordinarily economically valuable" work in half an hour 5:09:21.

Adopt this as a repeatable client deliverable: (1) assemble the client's flat exports as dimension and fact CSVs in one folder; (2) run structure analysis with your strongest model at high reasoning, letting anomaly distribution, coverage gaps, and grain mismatches set the page architecture; (3) spawn five parallel design variants against one brief that requires every variant to surface the same seven anomalies; (4) deploy the winning static HTML to Netlify through the MCP connector; (5) password-gate it per client via URL parameters; and (6) make the front page the signals feed — the page that names the anomaly (CL-010's lead collapse) before any chart loads. That is the product your monthly report was pretending to be.

Your score

One-sentence summary

Flat export silos become a deployed, password-gated HTML dashboard whose front page names anomalies like a lead-flow collapse — a structure forced by the data itself, designed by the strongest model at high reasoning, and shipped through Netlify's MCP connector.

Glossary

Every term marked with a dotted underline above. Skim it, or search the page for a word.

Ideas worth stealing into your own setup