Why I did this: to test how far a single person can go from RFI → product using vibe coding + AI helpers.

Edit — 27 Aug 2025

I’ve also expanded this experiment into a more formal vendor-style brief. You can read that follow-up here: Project Brief: TEC Tahatū CV & Cover Letter Builder (Vendor Simulation)

TL;DR

  • RFIs can look simple. The work is not.
  • Vibe coding gets you ~80% (exploration, scaffolding, first drafts).
  • The final 20% (debugging, infra, security, accessibility, export fidelity, compliance) still takes months, not weeks — but it’s months, not the year+ you might expect.
  • Tooling friction is real: Copilot vs ChatGPT context split slowed me down.
  • Would I build a production system solo? Yes — if the product has a broad market, not a single-customer (govt-only) dependency.

A quick note on PM & architecture roles

While I wore the hats of product manager and architect here, I want to acknowledge these are deep, strategic professions. They can make or break products. My sketches came quickly because I’ve spent time around these functions, but it’s not trivial. Respect to those crafts.

Why this exercise

I spotted an RFI from NZ’s Tertiary Education Commission (TEC) about a CV & cover letter builder for their Tahatū careers site and thought: that looks straightforward.

Here are the tender details for context:

  • RFx ID: 32236908
  • Title: CV and Covering Letter Builder
  • Tender Type: Request for Information (market research)
  • Purchaser: Tertiary Education Commission
  • Closing Date: 29/08/2025

I wanted to stress-test vibe coding on this real brief:

  • A few hours thinking,
  • A few hours exploring features/roadmap,
  • A few more vibe-coding a small feature.

The goal wasn’t to finish a full app — it was to see what a single person could realistically do quickly.

If you’re curious about tenders like this, NZ procurement is centralised at GETS (Government Electronic Tenders Service) where you can browse or subscribe.

My process (step-by-step)

  1. Skim the RFI → translate to product goals.
    Outcomes first, not features: what must a learner do in <5 minutes?
  2. List features → slice a roadmap.
    Minimal MVP, then stack value (parsing, tailoring, ATS, analytics).
  3. Architecture sketch.
    Next.js + Vercel, Supabase Postgres/Storage, NextAuth, Prisma, LLM via OpenAI/Azure, Playwright PDF, docx-templater.
  4. Vibe code a vertical slice.
    Paste profile text → LLM → structured JSON → preview → PDF.
  5. Reality check.
    Accessibility, exports, data governance, i18n, and integration always take longer than you think.

My background & reliance on AI

I’m not a full-time developer. This was an experiment. I relied heavily on ChatGPT and Copilot.
Some choices were new territory (Prisma ORM, Playwright for PDFs, docx-templater) and I trusted AI suggestions.
But the big calls (Vercel vs AWS, Postgres, Supabase) I knew enough to choose.

Lesson: know your foundations. For production, research each component. Lean on AI for secondary tools, but validate core architecture.

Lifestyle context: I also split my work between mobile and desktop. ChatGPT is my brainstorming buddy on the train and in spare moments (mobile app), while VS Code + Copilot is my coding setup on the laptop. This causes some friction because context doesn’t carry smoothly between the two, but it’s how I fit the work into my life and make use of idle time.

Feature set I landed on

  • Auth (email or OAuth)
  • Profile: paste natural language or upload CV → parse → structured JSON
  • Master CV: pick template, live preview, edit, export PDF/DOCX
  • Cover letter: generate from job ad + profile, editable
  • Applications: save tailored CV/CL per job, track outcomes + feedback
  • ATS check: keyword coverage + suggestions
  • Analytics: funnel + template popularity
  • Accessibility & inclusion: WCAG 2.2 AA, plain language, te reo Māori + English, co-design input

The roadmap that felt honest

  • Start: thin MVP — auth, profile paste, one CV template, PDF export.
  • Next: upload parsing, multiple templates, cover letter basics.
  • Then: job ad tailoring, DOCX export, tracker.
  • After: ATS scoring, analytics, low-bandwidth support.
  • Finally: employer integrations, template uploads.

Even with vibe coding, shipping polished features that meet gov-grade accessibility, privacy, and export fidelity needs months, not weeks. Still better than a year.

What surprised me

  • PM hat felt easy with AI: you can draft features + architecture in an afternoon.
  • Engineering is still the grind: edge cases, auth, exports, accessibility.
  • Debug/deploy tax: CI/CD, secrets, PDF quirks turn days into weeks.

Key tools & friction

  • VS Code + GitHub Copilot for scaffolding. Smooth.
  • ChatGPT premium for brainstorming.
    👉 Pain point: context doesn’t flow between Copilot and ChatGPT.

Tip: Keep a requirements.md in-repo. Feed Copilot, not tabs.

What I’d do differently

  • Unify context in-repo. Keep requirements, architecture, prompts next to code.
  • Print-first design. Build the print route early; test PDF often.
  • Accessibility baked in. Start with WCAG checks.
  • Plan DOCX earlier. Know placeholders and loops up front.
  • Think bandwidth. PWA or lightweight exports matter.

Would I ship a production system solo?

Yes, but with caveats:

  • Choose a broad market product, not just one govt buyer.
  • Be full-stack curious (frontend, backend, auth, exports, privacy).
  • Budget months, not weeks.

Advice:

  • Translate RFIs to outcomes first.
  • Ship vertical slices.
  • Validate schemas early.
  • Design for print + a11y first.
  • Treat AI as advisor.
  • Show you understand governance.

Covering tenders yourself? Start with GETS and subscribe. It’s free and gives you visibility.

Closing

Vibe coding is powerful. It gets you far, fast. But the last mile needs craft, discipline, and respect for the work. AI can help you sprint — but it won’t carry you over the finish line.

Written for KiwiGPT.co.nz — Generated, Published and Tinkered with AI by a Kiwi