The vault

Things I killed

A public record of business ideas I stopped working on — what each one was, why it looked good, what actually killed it, and the skill I kept. Most idea lists are graveyards without headstones. This one has the reasoning attached.

16
Deletions: 100%
Kept the skill

I was telling myself I was short on ideas. I was long on ideas and short on deletions. Plausible ideas are the dangerous ones — they never fail loudly enough to bury, so they sit there quietly eating the attention the real thing needed.

№ 01Killed2023 — research Mar, live Oct
Wise Lightning — e-commerce brand
My first e-commerce brand, 2023. A wireless-charging lamp with custom branding, ~300 units, sold on Amazon via the traditional FBA route.
Why it looked good
A real physical product with custom branding, not a dropship reskin. I took the slow, legitimate route — samples, freight, proper listings — and the freight and launch actually went fine.
What killed it
Margin, structurally. I ran roughly a 3× multiplier from cost to sale price when the floor is 5×. At 3× the ad spend, returns and marketplace fees eat the product alive before you make a mistake — and I made mistakes: stock bought upfront before any proof of demand, a listing pulled a month post-launch over a verification issue, and a ~10% return rate from units failing inside warranty.
What I kept
The margin rule, which now governs anything physical I look at. Plus a content-first validation method: prove attention before buying inventory, never the other way round.
Testing the lamps
Building the listing
The power tower — the one I passed on, and I'm glad I did
Rule5× cost-to-price or don't start. A 3× margin is not a thin business, it's a slow loss.
№ 02Killedkilled 2026
SOTE — faceless history YouTube channel
A faceless, documentary-style YouTube channel narrating Asian history. Long-form, production-heavy, assembled with generative tools.
Why it looked good
Evergreen library content, no personal exposure, and other channels were visibly succeeding with the exact format. Every input had a tool that could produce it.
What killed it
Cost per datapoint. On YouTube the only validation loop is the algorithm — you cannot test a channel with friends or DMs, only by shipping. With episodes running up to an hour and three quarters, every single datapoint cost weeks of production. That's structurally untestable for one person. I'd also mistaken replication for validation: their success validated *their* channel, not the format for a new entrant.
What I kept
Video-form fundamentals — pacing, hooks, script structure — which went straight into the weekly vlog. And a rule about how to test anything algorithm-judged.
One of the thumbnails. The packaging was never the problem.
RuleWhen an algorithm is the only judge, minimise cost-per-datapoint. Earn the right to produce long-form after it confirms the audience.
№ 03Ruled outretired 2026
Anything faceless
A standing rule rather than a single project: building audience assets with no face attached.
Why it looked good
Privacy, and the belief that the work should stand on its own without a personality attached to it.
What killed it
Faceless removes the one thing that is genuinely hard to copy — you. In a market where anyone can generate competent content instantly, the face is the moat, not the constraint.
What I kept
The decision to put myself on camera at all, which is what made the weekly vlog possible.
RuleNever faceless again.
№ 04Ruled outresearched 2025
An AI influencer
A synthetic avatar run as a niche media brand. Generated face, generated voice, daily short-form output, monetised through brand deals and affiliate links.
Why it looked good
Every input had a tool that could produce it, and the tools were cheap. An avatar, a consistent voice and a daily posting cadence could all be bought for a few hundred dollars. It looked like a media business with the human cost removed.
What killed it
The costing exercise answered it. You are not paying for the avatar. You are paying for distribution, exactly like a human influencer pays for ads, collaborations and PR. A serious six-month run came to somewhere between six and eighteen thousand dollars in promotion alone. The synthetic face removes the talent cost and leaves the only expensive part untouched.
What I kept
The habit of costing the whole funnel rather than the artifact. When a business model looks cheap, check whether the cheap part is the part that was ever hard.
RuleIf a technology makes the production cheap, the constraint just moves to distribution. It does not disappear.
№ 05Ruled outresearched Dec 2024
A custom GPT as a product
A website selling a fine-tuned GPT for one industry. Subscriptions, pay-per-use, and custom builds for businesses on top.
Why it looked good
It was genuinely quick to build and the pricing table worked on paper. Pick a vertical, load it with domain context, charge a monthly fee. In late 2024 that still felt like a product rather than a feature.
What killed it
It is a thin layer over somebody else's API, which means two things at once. The API cost sits directly in your margin, and the platform underneath can ship your entire product as a checkbox. There is no defensible position between the model provider and the customer.
What I kept
The wrapper test, which now runs on every software idea. If the underlying provider could ship this as a feature next quarter, it is not a business.
RuleNever build the layer directly above a platform that is still expanding. That is the first place it grows into.
№ 06Ruled outresearched 2022
Vending machines
Buying and siting vending machines as a semi-passive cash business. Stock them, service them, collect.
Why it looked good
It is real, it is simple, and the cash is immediate. No customer acquisition, no marketing, no pitching. For someone with capital and no track record, a machine that takes coins is an appealingly literal business.
What killed it
The return on capital and time did not justify it. You buy the machine, you buy the stock, and you keep buying your own labour to restock and service it. The good locations are already taken and the rest do not clear the maths. It is a job with a vehicle attached, priced as an investment.
What I kept
Judging any opportunity on return on invested capital including my own hours, not on whether the revenue is passive-sounding.
RuleSemi-passive usually means the labour is just less visible on the spreadsheet.
№ 07Ruled outresearched 2022
Amazon retail arbitrage
Buying discounted retail stock and reselling it on Amazon for the spread.
Why it looked good
The mechanics are obvious and the feedback is fast. Find a price gap, buy, list, sell. No product development and no brand to build.
What killed it
There is nothing to own. The margin is the gap. The gap closes the moment anyone else finds it. And you have no claim on the listing, the customer or the supply. Every hour spent scanning shelves buys one transaction and builds no position.
What I kept
This is what pushed me toward a private label instead, which became Wise Lightning. Owning the listing was the whole difference, and that lesson cost nothing here.
RuleIf the entire business is a price gap, you are renting someone else's inefficiency until they notice.
№ 08Ruled outresearched 2024
A protein supplement company in Thailand
Manufacturing and selling protein powder into the Thai market as a local consumer brand.
Why it looked good
The fitness market here is growing fast and visibly under-served compared to Australia. Living in the market felt like an information advantage, and the product category is one I actually use.
What killed it
Complexity, and not enough data to justify it. A consumable ingested product means local regulation, import rules, manufacturing partners and testing. All of it in a second language and a legal system I do not know. And the demand signal was assumed from what I could see around me, not measured. High operational complexity on top of an unvalidated assumption is the worst combination there is.
What I kept
Being in a market is not the same as understanding it. Living somewhere generates hypotheses, not evidence.
RuleDo not take on regulatory complexity for a demand signal you have not actually measured.
№ 09Ruled outresearched 2024
Selling website themes
Building and selling premium website templates as a productised, buy-once asset.
Why it looked good
Build once, sell repeatedly, with no client work attached. It sits close to skills I already have, and a good theme keeps earning long after it ships.
What killed it
It needs a team. A competitive theme is design, front-end, documentation, demo content and then ongoing support for every buyer who cannot get it working. Support is the part nobody prices in, and it never stops. As a single operator I would have been shipping one theme while answering tickets for the last one.
What I kept
The engine version of the same idea, built for one narrow use rather than a marketplace. Templated sites for local service businesses, where I own the customer and there is no public support queue.
RuleA one-to-many product with a support obligation is a team business wearing a solo costume.
№ 10Ruled outruled out 2022
Anything NFT
The 2021-era opportunity set. Minting a collection, flipping profile pictures, or building something adjacent to it.
Why it looked good
It did not, particularly, and that is why it stayed a one-word entry. The returns being posted at the time were real and very loud, which is its own kind of argument.
What killed it
It was speculation, not a business. The value depended entirely on the next buyer paying more, with nothing underneath producing anything. That is a position, and I was trying to build an income.
What I kept
The clearest line I have between investing and operating. Both are legitimate. Confusing one for the other while you have no runway is not.
RuleIf the return depends only on the next buyer, it is a trade. Do not file it under business.
№ 11Ruled outresearched pre-2023
Buying into a franchise
Owning a franchise outlet as a first business — a proven system, a known brand, and someone else's playbook to follow.
Why it looked good
It removes the hardest part of starting: you don't have to invent the offer, the brand or the operations. For someone with capital and no track record, buying a system that already works is a rational-looking shortcut.
What killed it
The economics reward whoever owns the system, not whoever buys into it. The franchisor collects fees on revenue regardless of your margin, sets your suppliers, and can change the terms. You take the operational risk and the capital risk while the party with actual leverage takes a cut of the top line. Buying in is the expensive route to the wrong side of that trade — the position worth holding is franchisor, not franchisee.
What I kept
The lens itself: before entering any arrangement, work out who holds the leverage and whether the economics reward them or you. It's the same question I now ask about platforms, marketplaces and retainer clients.
RuleIf the system's owner profits whether or not you do, you're the product line.
№ 12Ruled outresearched pre-2023
A pressure washing business
A blue-collar service business with low startup cost, immediate cash flow and visible demand. I got reasonably far into researching it.
Why it looked good
Genuinely sound as a business. Low barrier, quick to revenue, easy to demonstrate value, and a market that never disappears. Plenty of people have built something real out of exactly this, and I'd still say it's a better first business than most online ideas.
What killed it
It anchors you to a single location. The work only exists where you physically are, which is the precise opposite of what I wanted. I'd have been buying a job with a van attached, in one city, at the exact moment I was trying to get geographic freedom.
What I kept
The filter that has governed everything since: does this survive me being in another country? If the answer is no, it doesn't matter how good the economics look on paper.
RuleA good business you can't leave is still a place you have to be.
№ 13Killedresearched and killed 2026
Faceless KDP publishing
Publishing books at volume through Amazon KDP, including a children's-book line, produced largely with AI assistance.
Why it looked good
Low marginal cost per title, a real marketplace with buyer intent, and an obvious pairing with advertising skills I already had.
What killed it
Same treadmill shape as the agency, plus a values problem: the volume approach only works by flooding a category, and I did not want to be the person doing that.
What I kept
The publishing research fed a decision to write one authority book properly, audience-first, on a multi-year horizon — instead of many disposable ones.
RuleIf the strategy only works at flood volume, it is a channel exploit, not a business.
№ 14Ruled outresearched 2026
Affiliate content sites
SEO-driven affiliate sites monetising review and comparison traffic.
Why it looked good
The classic passive-income shape, and I have the SEO and content-production capability to execute it.
What killed it
The model is mostly dead and what survives is brutally hard — AI summaries absorb exactly the informational queries the model depends on. I set a validate-at-15-pages gate and the research said don't start.
What I kept
A framework for validating any content play before writing at volume, plus a much sharper read on which content types survive AI summarisation.
RuleValidate the channel before producing for it, not after.
№ 15Killed2026
Cold email to local trades
Scaled cold email and SMS outreach selling automation systems to local service businesses.
Why it looked good
Deliverability and sequencing were solved problems for me — I'd hit 60–80% open rates elsewhere. It looked like a volume equation.
What killed it
Roughly 5,000 emails and 100 SMS produced zero replies. Not a poor rate — zero. The diagnosis wasn't the copy or the list: the audience had no felt pain about the problem I was solving.
What I kept
The infrastructure, and a hard lesson about channel-to-audience fit. Trades get reached by phone; email works for higher-ticket buyers who already know they have the problem.
RuleChannel follows audience. Never the reverse.
№ 16Ruled outtested 2026
Physical noticeboard advertising
Reaching local trades through physical noticeboards in supply stores and depots.
Why it looked good
Cheap, hyper-local, and it put the message where the audience physically is — with almost no competition for attention.
What killed it
No measurable path from placement to enquiry. Untrackable channels can't be optimised, which means you can never tell a bad message from a bad channel.
What I kept
A bias toward channels with attribution, even when the untracked one is cheaper.
RuleIf you can't measure it, you can't iterate it — and you'll blame the wrong variable.