AI · CONTENT OPS · CLIENT UNDER NDA
An AI workflow that cuts a landing page from 16 hours to 15 minutes
A subscription-based AI tool that turns raw text into pages that already fit the CMS schema, replacing a time-consuming manual task with an AI workflow.
- −96%
- manual writing time
- −70%
- cost vs. hourly content writers
- 100%
- content and images AI-assisted
- Role
- Sole creator and product owner
- Duration
- 5 MONTHS
- Year
- 2026
- Client
- Under NDA
The problem
In the first quarter of 2026, the commercially available AI tools could not map unstructured text onto the strict set of dynamic components of an Astro and Strapi v5 site. Building a landing page by hand took 16 working hours, two working days. Three things caused it:
- 01
The content team wrote raw text in Google Docs or Word without knowing the limits or the schemas of the frontend.
- 02
Moving that text into dozens of nested components (Dynamic Zones) in Strapi meant choosing blocks, labels and colors by hand and linking every asset in the Media Library.
- 03
Nothing checked a draft against the schema. General AI tools did not know this project's schema, so they wrote content that did not fit the Astro components.
What I built
I built a dedicated orchestration system in TypeScript. It has five parts:
- 01
Teaching the tool the schema
The system parses the Strapi component structure in real time: field names, types, relations, validations, descriptions and labels. It also reads pages that are already published and uses them as reference examples (few-shot in-context learning), so the output follows patterns that already work on the site.
- 02
Getting text and assets in
The author pastes raw content into an editor that takes rich text or Markdown. Images come from four places: AI generation based on the text, web search, the Strapi Media Library, and a direct upload from disk.
- 03
An assistant that works in the document
An assistant reads the document live, checks facts with web search, and tightens the copy. A skill registry lets me add formatting and business rules on the fly, written by hand in JSON Schema or created by the assistant itself. When the author approves the text, the agent compiles the document into JSON that matches the page components (sections, blocks, color variants) and pushes it to the Strapi API with the status Draft, through an authorized call. A person reviews it there and publishes.
- 04
Validation before the CMS
Every payload is validated against the JSON Schema before it is sent, so type errors are caught before they reach Strapi.
- 05
Learning from what works
The app has an evaluation and telemetry module. When a generated page is flagged as good, its structure joins the reference set for later requests, and later pages choose sections, design tokens and block hierarchy more accurately.
- 06
Ready to be sold as a subscription
The app is prepared to be sold on a subscription basis. Users create projects and share them with other content creators. Everyone in an organization gets a full overview of the content it has generated, along with statistics per user.
How it is built
The whole tool is one Next.js app hosted on Vercel, with a WYSIWYG and Markdown editor, streamed chat and draft previews. The API, the schema validation pipeline and the batch worker are part of the same app. Firebase provides the rest: Firestore holds the cached Strapi schemas, the business rules, the prompt and skill repository, and telemetry, and Firebase Authentication handles sign-in. Sign-in supports multi-factor authentication. A separate management panel gives the support team its own tools. The model layer uses the Vercel AI SDK, which puts different model providers behind one interface, so a model can be swapped without changing the client code.
My role
I was the sole creator and the product owner. The project took five months.
Outcome
A landing page now takes about 15 minutes: 10 minutes of automatic synthesis and 5 of human review, down from 16 working hours. This was measured on real pages. Manual writing time dropped by 96% and content cost fell by 70% compared with hourly content writers. All content and images are now AI-assisted. Every payload is checked against the JSON Schema before it reaches Strapi, and swapping the model needs no change in the client code.
What comes next
Two things are planned. The first is a small assistant widget inside the Strapi admin panel, so editors can change blocks one by one with the state synced both ways. The second is a dedicated background worker for batch processing, for migrations of many pages and heavier layout changes.
TECH STACK
Built with
LANGUAGE
- TypeScript
STACK
- Next.js
- Firebase
- Vercel AI SDK
INFRASTRUCTURE
- Firebase
- Vercel
NEXT STORY
PLATFORM · CRM
Enterprise real estate →
05 / CONTACT
ACCEPTING NEW STRATEGIC ROLES FOR 2027
I thrive where complex engineering meets commercial reality. Open to Head of Web & Product roles.
KRAKÓW, PL
BUILT WITH BEARMINION