Built three working apps in five months with no coding background

ACHIEVEMENTS

I am not a developer. I am a holistic health practitioner and somatic therapist in rural Central Queensland, with a business degree and a background in mining operations and training. In March 2026 I could not build software. By August I had three working builds, a trade mark application filed, and developer accounts open on both the Apple iOS and Google app stores.

The three apps come from three parts of my working life, and each solves a problem I have watched go unsolved. Both web apps are live at the links at the end of this entry, and the iOS app is shown running on my own phone in the attached recording.

APPLICATION NAMES

- SONIC SYMPHONY

- MIND BODY SOUL COMPASS

- GROUND RULES

~ Sonic Symphony is an iOS app built in Flutter, on Firebase, running to 6,052 lines of Dart. It plays a library of 23 tracks generated with Suno and directed by me, category by category, against a written brief. It opens with a needs question rather than a content wall, so a person picks by what they need now, not by what a playlist thinks they want. It has a searchable library sorted by state, a player with a sleep timer, real accounts, a members-only tier enforced in the database rather than in the interface, child-safety flags carried on the content itself in the database rather than checked in the interface, and an account deletion flow with plain-language data wording. I built account deletion before launch rather than after, because it is the right default and because both stores require it.

My own recordings are deliberately not in it. I have played crystal singing bowls, handpan, medicine drum, monochord and tuning forks for years, and that material stays out of the product until the app is launched and the rule protecting paid audio has been tested against a second account.

I wired the Sonic Symphony subscription two days ago, RevenueCat into Firebase, with a webhook deployed. It hasn't been exercised by a real purchase yet, because the store products cannot be configured until the Apple process completes next week.

A trade mark application published, the day after the first code was written. It is an application under examination, not a registration. The Apple Developer organisation enrolment ran from 18 July to 29 July, when the programme agreement was accepted. The Google Play developer account was verified on 6 August. SHE IS MAGICK PTY LTD has been registered since January 2026, and the business name for the app was registered on 8 August.

~ Mind Body Soul Compass is a web app built via Lovable with TanStack Start and Supabase, and it is the largest of the three. A person logs what is stacking up today across fifteen kinds of load, including heat, driving, noise and screen load, and gets one chakra, one crystal, one sound, one journal prompt, one body reset and one practical next step. It has about twenty-two routes, nineteen database tables, and thirty-nine row-level security policies. It includes a Shift Worker Mode with resets built for crib (lunch) breaks, night shift and the drive home, and a Kids Calm Corner with a written list of what must never appear in it.

The opening instruction I wrote for it sets six prohibitions before it names a single feature: not generic, not fluffy, not tied to one town or workplace, no medical claims, do not skip my source rules, do not invent a new brand system. Two lines from that same instruction are the clearest statement of what I was trying to make. "Useful before mystical." And "Do not diagnose." It ends by telling the AI what to report back to me when it was done: what was built, what still needs manual setup, what tables were created, what keys would be needed later, and what to test next. I wrote the acceptance criteria before the work started.

~ Ground Rules is a five-route interactive demo for mining sites, built for phones. Its headline is "Courtesy is a safety system". It reframes workplace respect as risk control rather than a values poster, grounded in Queensland's Code of Practice on managing psychosocial hazards and cited incivility research. It runs a branching pre-start radio scenario with nine graded choices and nine coaching responses, an eight-statement site self-check, and an anonymous team pulse where supervisors see trends and never names.

Ground Rules is a demo and says so on every screen. Its forms deliberately do not transmit. I would rather show a mining manager something honest than something that pretends.

TOOLS

Lovable for the two web apps. Supabase for Postgres, auth, storage and row-level security. (MBS & GR)

Claude and Claude Cowork for research, specification and build packs. Flutter and Dart, built through Xcode. Firebase for auth, Firestore, storage and cloud functions. RevenueCat for subscriptions. Airtable as a structured brain for content. Suno generated the entire audio library, directed track by track. ChatGPT for early exploration and image generation. Claude Code started two days ago, for the membership and subscription work on the Sonic Symphony codebase.

Tooling, registrations and the trade mark came to under AUD$4,000 across the whole period. The course itself was AUD$9,898 as well as Mindvalley Membership AUD$574.76. The entire cloud infrastructure bill for a working iOS app came to four cents (o.o4c).

METHODOLOGY

Before any of the building started, AI Mastery taught me something that shaped every tool decision after it: owning my own subscriptions mattered for security, not just convenience. I went from free ChatGPT to paid ChatGPT Pro for that reason, then moved again to a Claude subscription, which felt more secure. I went through the settings on every tool I used, checking what I was actually exposed to.

I learned to write a specification before I wrote a prompt.

The clearest example is Ground Rules. The entire app is three prompts, four hours and twenty-two minutes, one morning.

I am a mining operator myself, and as a woman I have struggled with the toxic cultures on site. I have had several conversations with our SSE and GM about ways to help implement real change. My boss's current campaign is "Attitude is everything." Ground Rules is my attempt to build on that, using AI and human behaviour together to make a real difference for the people working on-site.

By 30 June I had written a complete Courtesy Culture workshop package for Queensland coal crews: six sections, a facilitator run sheet, and a pilot proposal addressed to the SSE and GM, sourced against the Coal Mining Safety and Health Act and RSHQ guidance. Ground Rules is that thinking built as software two days later.

The very developers beginning - My first prompt was seventeen characters long: "blank starter app". The AI declined to build and asked what I wanted, which was fair. What happened over the next two and a half hours I can no longer fully reconstruct in detail, but I remember the shape of it: I was working with ChatGPT to build a version, look at what it produced, and refine what I actually wanted. Somewhere in that back and forth I was also learning to specify things in technical language such as breakpoints and colour tokens, terms that had never been part of my working vocabulary. By mid-morning a full five-route version of the app existed, and I had also been sharpening a tighter written specification alongside it, using what the build had shown me. I deleted the working version, and five seconds later pasted in that finished specification: 13,953 characters covering routes, colour tokens, breakpoints, state changes, exact copy, scoring bands and reduced-motion behaviour. It built in one pass with no errors. My content survived word for word.

The delete was not small. The revert removed 1,149 lines across twelve files, of which a thousand were application code, including all five working routes. I was not starting over. I threw away the code, not the thinking. The specification already held what I had learned from building the first version. The five-second gap matters because it shows nothing was salvaged after the deletion — the thinking had already moved somewhere else.

Two hours later I sent the third and final prompt, which changed only the colour system and three lines of text, ending with "do not change anything else, no layout, route or logic changes". Learning to describe a change rather than an app was the real step forward.

Then I learned that the method has to suit the job.

Ground Rules took three prompts. Mind Body Soul Compass took fifty-two, in the same fortnight, and each one was short. That is not inconsistency. A pitch site is knowable in advance, so it is worth specifying completely and building once. A database application with accounts, roles and paid content is not knowable in advance. It gets negotiated with the thing as it takes shape, one small move at a time, and every move has to be fenced so that nothing else changes while you are working. Knowing which of those two a job is, before starting it, saved me more time than any prompt technique did.

I also learned to set the boundaries first, and to treat them as architecture rather than as policy.

Before any code existed for Mind Body Soul Compass I wrote the rules: no medical claims, no diagnosis, child-safe content only in the kids section with an explicit list of what must never appear there, consent required before saving birth data. Every one of those became enforced behaviour in the running app rather than a line in a document.

The one I am most pleased with is smaller and more personal. I live in a small town, and I set a standing requirement that the town must not be discoverable through anything I ship.

I can date it precisely, because the record kept it. At 10:19 on the morning of 2 July, four hours into building my first app, I sent a prompt making four copy changes. Three of the four removed my town: from a positioning paragraph, from a footer line that named both my business identities alongside it, and from the site footer, which became "Central Queensland" instead. That was on Ground Rules, a demo with no accounts and no payments, months before anything of mine went near a shopfront.

The same requirement then had to survive a real one, on Sonic Symphony, the app that will eventually take payments. Both app stores publish a physical address for any developer taking money, and Apple pulls that address automatically from the business credit record, so the requirement reached back into the D-U-N-S registration and the payments profile, which now carry a Brisbane service address. It forced a change of app identifier, because on Android the package name appears in the public store URL. It caught a "made in" line in both store descriptions before either shipped, and a backup file still carrying the town name, which was moved out of the source tree.

One personal boundary, set on day one and held, changed the copy on one product, then the business record, the app identifier, the store listings and the file structure of another.

What went wrong, and what I did about it:

SONIC SYMPHONY

Claude told me, with confidence, that Apple's price ladder moves in whole dollars above fifty, so the yearly price I wanted could not exist. I did not accept it, and checked with ChatGPT for accuracy. Apple's published sentence stops at "etc." and never states the increments above fifty dollars, which means the claim was a third-party summary presented as fact. The price exists. It is now in the app, and the project carries a standing rule not to assert what that ladder contains from a web source. When I was then asked to approve a fallback price in advance, in case the one I wanted was unavailable, I refused, and asked to be shown the real list instead. A price is a promise to a customer. I will not have one substituted while I am not looking.

A Firebase tool reported "Found 0 Firebase projects" and offered to create a new one. The real cause was an expired login. Accepting would have cut the app off from all 23 recordings and every account in it. Reading the message rather than clicking through it is the whole skill. Thankfully we didn't lose all our work, from a simple pause and read.

An app that should have been under three megabytes was downloading at fifty-nine. I asked for an assessment of why, and the cause turned out to be one configuration line bundling every source image into the build, including drafts I had rejected.

A webhook failed four times in a row. Twice, the fix was a code change, and twice it made no difference, because the problem was never in the code. I recognised the pattern: it was circling, and no amount of further guessing was going to solve it. I stopped it and stepped in myself. The actual fault was a permission setting on the server, sitting entirely outside the code, and it needed a login only I held. The moment I switched it on, the webhook went live.

Two typefaces passed on my desktop and failed on mobile. Italiana and Bodoni Moda both looked right in mockups and were nearly unreadable on my own phone. No AI was going to tell me that.

The first batch of generated music all sounded the same. I said so. The cause was not the model, it was the prompt skeleton the SUNO AI had been reusing, and ten prompts were rebuilt on the things that actually vary a piece.

MIND BODY SOUL COMPASS

On one day in July I ran security scans on MBS Compass and worked through ten remediation items covering nine distinct findings, in three batches, each time naming the exact identifiers and adding the same instruction: fix only these, do not fix or ignore anything else. Left unfenced, an AI will improve things you did not ask it to improve, and a security surface is the last place you want that.

The middle of that day is the part I would show anyone. A fix I had asked for locked me out of my own admin panel. I pasted the exact database error and the exact screen I was stuck on, rather than describing them, and we found the chicken-and-egg problem and added a first-admin rule. Two hours later that fix appeared in my own next batch of findings as a privilege escalation risk, so we took it out and moved admin promotion to the server side. Three passes on one problem before it was right, and the middle pass was my own fix being caught and reversed.

When we needed astronomical calculations, the obvious library was Swiss Ephemeris. I blocked it, because its licence would have been a problem in a paid app, and specified an MIT-licensed alternative instead. I specified the precision it had to meet and the margin it had to sit inside. I also asked for a validation screen so I could compare output against known charts before trusting the engine. That instinct was warranted. The AI populated that screen with three celebrity charts carrying confident expected results, then removed them itself, on the grounds that published data for those charts varies by source. An unverifiable claim was caught before it shipped.

LEARNING'S

That the bottleneck was never the code. It was knowing what should exist, and being willing to say no.

The AI wrote the software. I brought the mining vocabulary, the somatic practice, the sound direction, the safety judgement and the standards. On 11 June, before any AI touched the build, I wrote the sound app's specification myself: fifteen screens, fourteen categories, the quiz, the track metadata model, a file naming convention, and two word lists — what the app may say, and what it may never say. I wrote the pre-start radio scenario because I have stood in those pre-starts. I wrote the middle answer in that scenario to be genuinely defensible rather than an obvious wrong choice, because that is how adults actually learn. I set the child-safety rules because I know what a child should not be handed. I caught the licence problem because I read the terms. I hold a written wording list for the sound app that allows support, relaxation, focus, sleep routine, emotional regulation and wellbeing, and forbids cure, treat, diagnose, guaranteed, clinically proven, and any named condition.

The judgement came before the building. On 27 June, five days before the first app existed, Claude searched its own documentation twice and told me flatly that Claude modes did not exist. I had been using six of them. I uploaded the skill file that contained them and it retracted in full — they were real, but they came from a skill I had installed, which is why its search had missed them. I then had a seventh mode written, and when the platform warned that installing it would replace the existing skill, I made the AI confirm the original six were preserved before I went ahead. I used the new mode the same day.

None of that is specific to these three apps. The specification before the prompt, the source rules, the fenced fixes, the wording lists, the check before trust: that is the practice, and the apps are the first three things it produced.

None of the three has launched. Nothing has been submitted to either store, there is no TestFlight build, and there are no users and no testers. Mind Body Soul Compass has one account, mine. Payments are not connected on the web app, and the astronomy engine I am proud of is not yet wired into a user-facing screen.

On Sonic Symphony specifically: the subscription is deployed but not proven. The webhook answers correctly, and the code that writes a membership on a real purchase has never run, because no purchase can be made yet. The rule protecting paid audio compiles, deploys and fails safe, and has never been tested against a second account. The tracks in the app are one to four minute pieces, not the full-length sessions a membership will eventually need, and seven of the fourteen categories hold no tracks at all. I know the difference between deployed and proven, and I am not going to describe one as the other.

The audio is generated rather than recorded, and that is a decision about protection. The generated library let me build and test the whole app without putting my own recordings into an unlaunched product with an untested paid-content rule. Bringing them in is a later step, and it comes with its own protections.

One last thing, and it is a fact rather than a feeling. Ground Rules was built on 2 July. Twenty-three tracks went into Sonic Symphony on 2 July. The trade mark was filed on 3 July. Mind Body Soul Compass was started on 3 July, and its database, its security model and its astronomy engine were all built by the morning of 5 July, which was the day I flew out. I was away until 17 July with no reliable service, and there is no build artefact of any kind, in any of the three projects, across those twelve days. The store and infrastructure work picked up the day after I landed back in the country.

Two applications, twenty-three tracks, a trade mark and a hardened database inside seventy-two hours, by someone who could not build software in March.

I would rather be judged on three honest, working things than on one polished claim.

RELEVANT URLs

- Ground Rules, live: https://ground-rules.lovable.app

- Mind Body Soul Compass, live: https://mbscompass-com.lovable.app

- Sonic Symphony: iOS build, screen recording attached

- Trade mark: published.

COMPLETELY AI GENERATED REPORT* (with my approval and scrutiny)

Thank you for your time and consideration, have a splendid day :)