07 / Marketing ops
Supergentic
Fieldhaus
The problem
The product names its own problem: the blank page. A journey, a segment or an email in Marketing Cloud each start from nothing.
Marketers know what they want in words. Supergentic starts from those words and hands back something ready to refine, export or build on.
How it was built
A single-page app on typed fixture data, so every screen could be clicked through and judged before any model was wired in.
- 01
One app, two front doors
React 19 on Vite 8 and TypeScript 5.9 in strict mode, styled with Tailwind 4. One BrowserRouter serves the landing page or the app depending on the hostname, and seven context providers wrap the routes.
The landing page: Turn Your Salesforce Team Into Superheroes. - 02
Blocks as the data model
BlocksContext owns the catalogue and writes it to localStorage on every change. A block has an id, type, cloud, status, creator, timestamps, connections, tags and free-form metadata; 40 block types are defined and 103 seed blocks ship with the demo.
Building Blocks as a grid, a list and a map. - 03
Generation as a phase machine
Builders run three phases, prompt, generating and preview, with a status message that changes each second. In the demo the result is picked from stored templates by the words in the prompt.
One sentence in, a branching journey out. The reply is scripted. - 04
Prompts built from fragments
Each Prompt Builder option carries a fragment of text, and choosing it appends that fragment, so the form and the sentence stay in step.
The Prompt Builder writes the sentence as you click. - 05
Segments to SQL
The export module turns segment criteria into SQL, so the preview beside the audience and the downloaded file come from the same function.
A segment with its criteria, audience and SQL. - 06
A Console of tool calls
Console replies come from one function of about 70 keyword checks. Each returns text, a tool-call line, data cards and follow-up actions, so the scripted answers have the shape a real agent's would.
A Console reply: a tool-call line, then data cards and follow-ups. Scripted. - 07
One list page, twenty uses
CloudListPage is a generic list, with a prompt, a generating state, a sortable table, a grid toggle, an export dialog and a drawer, reused by 20 pages across Service, Commerce, Data Cloud, Analytics and Agentforce.
One Labs switch turns on more clouds in the sidebar. - 08
Flags for what is unfinished
Eleven Labs flags, three beta and eight alpha, four on by default, live in sessionStorage and reset with the session. Nav groups and pages read them through one context.
Labs flags in beta and alpha tiers. - 09
Exports a developer can open
Downloads build a Blob and click a hidden link. The shared dialog writes JSON, CSV, XML and YAML; across the app there are eight file types, and the sales tools write SOQL, Salesforce ListView and Report metadata XML, an HTML dashboard preview and Markdown specs.
An export from a catalog list. Sample products. - 10
Sales tools that read the sentence
Pipeline Studio maps the words of a request to an object, a grouping, a metric, a chart type and a date filter, and the List View Manager turns its result into SOQL and metadata.
A list view built from a sentence, ready to leave as SOQL and metadata XML. - 11
The catalogue as a map
The map view lays the whole catalogue out with React Flow in four modes.
The catalogue as a map, by cloud. - 12
App Lab
App Lab reads eight kinds of source, calls Claude through a server-side proxy with twelve operations and prompt caching, generates a project with at most five calls in parallel, starts a Vite server for each generated project, and exports to ZIP, GitHub or Vercel. Each endpoint runs as dev middleware and as a serverless function. Without a key it stops with a plain notice.
App Lab says plainly when no model key is set. - 13
What is real and what is sample
Downloads, clipboard copies, search, filters, block editing, saved work, theme, density, font size and Labs flags are real. Records are typed fixtures kept in the browser, and the push and sync buttons only confirm with a notice.
- 14
Five bursts
86 commits between 30 March and 17 May 2026, on 12 days: the app and its exports; Superpowers, a QA pass and analytics with consent; the multi-cloud repositioning with themes and Labs tiers, 18 commits on 16 April alone; Disco meeting intake; and App Lab.
The solution
A clickable product across 43 screens: the landing page and 42 app routes, from Superpowers and Building Blocks to the Journey Wizard, Console, Data Hub, Sales Tools and App Lab.
Downloads, clipboard copies, search, filters, block editing, saved work, themes and Labs flags all work. Generation and Console replies are scripted from local data; App Lab calls Claude through a server-side proxy when a model key is set.
Decisions along the way
Script the AI first
Replies stream from local data, so every screen could be judged before a model was involved. The proxy to a real model was scaffolded only in the last week.
Both platforms from week one
Support for the older Marketing Cloud Engagement and the newer Marketing Cloud Next went in together on the fourth day, so no block is tied to one of them.
Tiers for what is unfinished
Newer features sit behind Labs flags in beta and alpha tiers. The console was moved down to alpha and off by default until it earned its place.
Fewer views, working buttons
Board and map views of the block library were removed, the list view was fixed, and the drawer buttons were made to do what they say.
What I delivered
- A Superpowers gallery of 140 automations and a six-phase wizard
- Building Blocks: 103 seed blocks in grid, list and map views
- Journey, Flow, Segment, Content and Campaign builders on one prompt pattern
- A Console with tool-call lines, data cards and shortcut menus
- Data Hub and Customer 360 on sample data
- Sales Tools that turn a request into SOQL and Salesforce metadata
- Exports in eight file types
- App Lab, a project generator wired to Claude through a server-side proxy
Value added
For marketers
Work starts from a sentence and leaves as JSON, CSV, XML, YAML, SQL or Salesforce metadata.
For the build
A fixture-first app with one data model, one list page and one prompt pattern, so a real model can be dropped in behind the same screens.
By the numbers
Commits per week, spring 2026
86 commits between 30 March and 17 May, in three bursts.
View as table
| Period | Value |
|---|---|
| 30 Mar | 22 |
| 6 Apr | 16 |
| 13 Apr | 32 |
| 20 Apr | 0 |
| 27 Apr | 2 |
| 4 May | 0 |
| 11 May | 14 |
Source: git history of the Supergentic repo
The result
Supergentic is a working demo. Its screens, builders, blocks and exports run on local sample data, and generation and Console replies are scripted. App Lab is wired to live Claude calls through a server-side proxy and stops with a plain notice when no model key is set.
Results
- 43routed screens
- 140superpowers in the gallery
- 86commits, 30 March to 17 May 2026
What I took from it
Scripting the replies first meant all 43 screens could be clicked through before a model was involved, and App Lab is where the same product meets a real one. The honest next step is the one App Lab started: put a model behind the builders and see which drafts survive.
Background
Salesforce sells two generations of the product Supergentic targets. Marketing Cloud Engagement is the established platform. Marketing Cloud Next was announced on 11 June 2025 as an agentic marketing platform built on the core Salesforce platform. A tool that starts from plain language has to produce work for both, which is why Supergentic supported both from its first week.