dev-diary

Aura week 1: I killed my voice-first ECE app on day 14

Two weeks. 60 commits. One 4-hour bubble bug. Then I killed the project. What Claude Code got right, and what I should have checked first.

Stone

Stone · Registered Nurse

AHPRA-registered · 5+ years primary care, Sydney · 2026-05-26

Aura week 1: I killed my voice-first ECE app on day 14

Last Wednesday at 8pm my partner said something that quietly ruined the next two weeks of my evenings.

She runs the pre-school room at a Sydney centre. Five-plus years as an ECT. Mid-conversation about her day, she dropped this line: “the kids see the back of a clipboard more than the educator’s face.” Then she walked me through her usual night-time routine — 30 to 40 minutes writing 12 individual learning observations, one per child, every shift.

I went to my laptop. Two hours later I had a Claude Code session open and the first commit of an app I was calling Aura.

Two weeks later I killed it. This is the post-mortem.

What Aura was meant to be

Voice-first iPad PWA for Australian early-childhood educators. One voice recording — say, two minutes about today’s room — fans out into 12 personalised per-child diary entries via an LLM. All AI running on a local server in the centre, not in some US cloud. Privacy was the moat.

That was the pitch. Local-first. Voice in, structured per-child out.

Day 1: the Whisper miss

Monday, 9pm, kitchen table. First commit at 11:40pm. Three hours, mostly thinking.

I gave Claude Code a vague prompt: build me a voice recorder UI plus a Whisper transcription endpoint. It did, and the code looked fine — until I noticed the fetch was pointed at api.openai.com. That’s the wrong Whisper. I had a Whisper Docker container planned for localhost:8000 on a NUC running at the centre, not OpenAI’s hosted API.

The fix wasn’t the code, it was the prompt. Round two I gave Claude Code a 7-line network topology in plain English (iPad → local NUC → Ollama and Whisper containers, no internet egress) and the second pass came back correct.

Lesson I’m keeping: network topology in the prompt matters more than feature description. Claude Code is fine at building the thing. It’s not psychic about where the thing connects.

The day-6 moment

Saturday night, 11pm. I was tired and decided to do the thing I had been putting off: collapse six separate HTML pages (Parts 1-4 plus dashboard plus settings) into a single-file SPA with a hash router.

I wrote a ~500-line prompt. Network topology, yes — lesson stuck. I specified hash router, IIFE-per-Part for state isolation, three new helper modules to extract, and a verification script that should pass at the end. I expected to wake up Sunday and spend most of the day debugging.

What I got, thirty minutes later: a 540KB single-file SPA. Three new modules extracted cleanly. A merge_build.py and verify_aura.py written without me asking — Claude Code inferred I’d want them. Smoke test green. The hash router worked first try.

I spent the next hour reading the diff because I couldn’t believe it. That was the wow moment.

The reason isn’t that Claude Code is magic. It’s that I had built up enough context across the project over six days that one fat prompt could carry the whole refactor. Big prompts work when you’ve earned the context. Day 1 me couldn’t have written that prompt.

The 4-hour bubble bug

The other shoe dropped on day 11. Part 2 of the app had 5 inspiration bubbles meant to arrange in a ring around a central mic button. They wouldn’t align. I’d nudge one CSS property, two new misalignments would surface. I’d fix those, the first one would come back. Mole-whacking for hours.

I made the mistake every indie dev with a Claude Code MCP browser tool makes: I went into the browser and started injecting CSS overrides live. Watching the layout shift in real-time felt productive. By midnight I had five layers of !important piled up and zero idea what was actually authoritative.

Root cause, when I finally stopped: there were two inspiration-bubble systems in the codebase. The new ring layout I had been building, and an older openInspiration() function from an earlier sketch that still ran on click and wrote inline style="visibility:...; opacity:...; transform:..." on each bubble. Inline styles win over stylesheet rules. My new CSS was being painted over by ghost code I hadn’t deleted.

The fix took ten minutes once I stopped browser-injecting and asked Claude Code to dump the relevant 200 lines of source. It pointed straight at the old function. One commit: -201 / +10. Old system gone. New system worked.

Four hours of “debugging.” Ten minutes of actual debugging. Math doesn’t work.

Numbers from the two weeks

Things I tracked, roughly:

The bubble-rewrite count is the one that stings. Eight rewrites of the same component is not engineering. It’s flailing.

Controversial opinion: browser MCP debugging is a trap

Claude Code’s browser tools let you inject CSS or JavaScript into a live page and watch the effect. It feels like the fastest possible feedback loop. It is — for the first five minutes. After that, you’re not debugging the code anymore. You’re debugging a temporary overlay that will vanish when you refresh.

My rule, formed at hour 4 of the bubble bug: any browser-inject session over 10 minutes is a sign to stop and dump source. Ask Claude Code to show you the actual files. Read them. Find the conflict. Edit the source. Push. Refresh.

I expect pushback on this one. Browser-inject feels fast. The dopamine is real. But the reason it feels productive is precisely because there is no friction — and friction is what tells you when you’ve gone off-piste. Without friction, you can spend four hours convinced you’re nearly there. You’re not.

Why I killed it on day 14

This is the part that hurts to write. I did my competitor research on day 14. Should have been day 0.

Three things I found in one afternoon of searching:

  1. Mana announced a “Live” voice-first feature for autumn 2026. That’s Aura’s USP. Already shipping.
  2. Early Edu AI is at $8.99/month for individual educators, with a stated 1,000+ subscribers. The single-educator long-tail I assumed was unserved is not.
  3. Mana / LoveHeart has SOC2 and a public trust center. They’re already on the Australian Childcare Alliance’s official AI-partner list.

I had also quietly decided — for cost and complexity reasons — to drop the local-first deployment. Cloud SaaS. Which meant I had no remaining moat against any of the above.

Cloud SaaS Aura, by a solo indie, against an ACA-backed compliance-credentialed incumbent with the voice feature also shipping. That’s not a fight I can win on technology, and I have no distribution advantage.

Killing the project on day 14 cost me 60 hours. Continuing would have cost me 600. Sunk-cost reasoning is the most expensive bias in indie dev — more expensive than any cloud bill.

Three things I’m taking with me

  1. Competitor research happens on day 0, every single time. Even (especially) when the problem feels obvious.
  2. Network topology belongs in the prompt before features do.
  3. The browser MCP inject tool is a 10-minute thing, not an hours-long workflow.

The blog you’re reading is what I’m building instead. Aura’s repo is archived. The next post in this dev-diary will be about rebuilding stonemegan.dev itself with Claude Code — failures included on purpose.

If you’re an indie dev mid-build right now and haven’t checked who else is shipping the same thing this quarter: close this tab and go do that first. Day 14 is too late.

TL;DR