Solo founder engineer · Own product · 2026- · private beta · Next.js 16, React 19, Postgres, Drizzle

Giáo Án AI

A lesson plan generator for Vietnamese primary school teachers, designed to replace the two to four hours of nightly paperwork teachers describe with something a teacher can finish before dinner. I built the code, the pricing, and the funnel; when it opens, I am also the support inbox.

The Giáo Án AI template library, a catalogue of lesson templates filtered by grade and subject.
The template library: 1,489 lesson templates across grades 1 to 5.

  • 1,489

    lesson templates

    counted from the template files in the product repository, August 2026

  • 13

    subjects

    Kết Nối Tri Thức curriculum, grades 1 to 5

  • 8

    subjects, TT27

    TT27 evaluation coverage in Giáo Án AI

docx and xlsx exportCV 2345 and Kết Nối Tri Thức compliant

Problem

Vietnamese primary school teachers write lesson plans to a national curriculum standard, and the paperwork around them is heavy: plans compliant with CV 2345 and Kết Nối Tri Thức, then TT27 evaluations on top of them. Most of it happens at night, after the actual teaching.

The work is highly structured, which is exactly what makes it a good target. The teacher's judgment is the valuable part. The formatting is not.

Constraints

Compliance is not negotiable. A plan that does not match the required structure is worthless to a teacher, no matter how well written it reads, so the output format is a hard requirement rather than a preference.

Teachers need the result in the tools their schools already use, which means docx and xlsx export rather than a web view.

I run this alone. Anything I ship, I will also support, so the failure modes have to be ones I can explain to a teacher over chat.

What I did

  1. The lesson creation form with grade, subject, and lesson selectors.
    Generation starts from a compliant template, chosen by grade, subject, and lesson.

    Template library

    I built a template library covering the curriculum, 1,489 lesson templates across grades 1 to 5 and 13 subjects, so generation starts from a compliant structure instead of trying to reach one by instruction alone.

  2. Server-side normalization

    Around the AI providers (Anthropic, Gemini) I put a server side layer that shapes responses and normalizes errors, so the product fails in a small number of predictable ways rather than passing raw model behavior to a teacher.

  3. docx and xlsx output

    Output goes out as docx and xlsx, and TT27 evaluations for 8 subjects export to Excel.

Decisions

DecisionWhyWhat I rejectedWhat it cost
Start generation from a compliant templateThe plan lands in a shape the teacher already recognizes, so reviewing it is fast and editing it is local.Reaching compliance by instruction alone.
Normalize provider errors on the serverThe product fails in a small number of predictable ways I can explain over chat.Passing raw model behavior through to a teacher.
Export docx and xlsxTeachers need the result in the tools their schools already use.A web view.

Details that mattered

The interesting decisions are not the model ones. They are about which parts of a teacher's work should stay manual, and how much structure to impose before generation rather than after it.

Starting from a compliant template makes the output boring in the way it needs to be. The generated plan lands in a shape the teacher already recognizes, so reviewing it is fast and editing it is local.

The stack is Next.js 16, React 19, Postgres with Drizzle, and Clerk for auth. I also own positioning, pricing, and the funnel.