candidates, from a weekend script to a multi-year hardware programme, ranked against the hours you actually have, the runway you can actually survive, and the two things every idea list ignores — whether you can reach anybody, and whether they would just build it themselves.
The address bar keeps your settings — copy the URL to send this configured.
Ranked by your weights and shown with the whole arithmetic. . The ranking is a sorting aid, not a verdict — and the draft bet at the bottom of each card is the actual output: something small enough to be wrong about in a fortnight.
Every list of things to build is written about the ideas. Here is a market, here is a gap, here is a total addressable market with a comma in it. The idea is scored and the person holding it is not, which is precisely backwards, because the idea is almost never the binding constraint.
The constraint is reach. candidates are on this page and the difference between the ones you could start on Monday and the ones you could not has almost nothing to do with their merit. It has to do with hours a week, months of patience, and the fact that you can already get in front of people through and not through anything else. Change those three and the page changes its mind.
Two things fix where an idea sits. How much building it needs before anybody could pay for it, along the bottom; and how large the thing that writes the cheque is, up the side. Neither moves. What moves is the line — everything inside it can produce a paying customer before your runway runs out, and everything outside it is a different decision about your life rather than a worse idea.
Time to the first dollar is two things stacked, and people plan for one of them. There is the building, which is in your control and which you will estimate optimistically. And there is the waiting — the months between shipping a thing and somebody you have never met deciding to pay for it, which is not in your control at all and which is set almost entirely by the channel.
The spread in what these cost to build is wide — about thirty times, from a few weeks to a few years. The spread in what it costs to put ten customers in front of them is far wider than that, and it is not set by the thing at all. It is set by who buys it and how they are reached, which are fixed long before any code is written.
Which is why the useful question is not "what could I build" but "what could I build that I can also reach". Those are different lists, and only one of them is on this page. It is also why every bet worth opening names a channel before it names a feature: the channel is the part you cannot change later without changing the product underneath it.
Suppose it works. turns over in its third year of trading. That is the number that goes in the tweet. What reaches your account is , because running it costs money, because self-employment tax is both halves of FICA with no employer to pay the other one, and because federal and state tax then apply to what is left. The total take is .
And underneath that sits the line that appears in no spreadsheet at all: the hours you spent getting there, worth at what your time earns elsewhere. It is never invoiced and it is always spent, and on most of these candidates it is larger than every cash cost combined.
There is one filter that matters more than market size, and until recently it had one half. The first half is whether a competent assistant can deliver the value without the person ever arriving at your thing. If it can, you are not building a product, you are building a citation. So each candidate is scored on four things an answer cannot do on its own: compute over data that is live or private, hold state across a decision longer than a conversation, take an action with consequences, and accumulate data nobody else has.
The second half is newer and it is the one that has quietly killed more small software in the last two years. The buyer arrives, looks at the price, and builds their own by Thursday. Not a worse version — a version that fits their workflow exactly, that they own, and that cost them an afternoon. Four tests again, and this time a high score is a warning: the buyer writes code, the data is already theirs, it stays built once built, and nobody has to be accountable when it is wrong.
These are not the same failure and they do not correlate. survives both — and out of twelve. survives neither, at and .
It is worth being specific about what changed, because "AI will eat SaaS" is not an argument. What changed is that the cost of a bespoke internal tool fell through the floor for anyone who can describe what they want. A product manager with an afternoon and a coding assistant now produces the thing that used to be a $39-a-month subscription — and it is better for them, because it does exactly their thing and nothing else.
Which means the question is no longer "is this valuable". It is "is this valuable to somebody who cannot make it". Four things keep a buyer from making it, and only four. They have no engineer — a pharmacist, a funeral director, a club treasurer. The data is not theirs — they could write the software in a morning and would still have nothing to run it on. It never stays built — the rules move, the site changes, the schema drifts, and now they own that forever. Somebody has to be accountable — a self-assessment is not an assessment, and a self-built inspection record is not a record.
The uncomfortable part is that exposure runs almost exactly opposite to appeal. The candidates a technical person most wants to build — tools for developers, tools for themselves, tools whose whole charm is that you would use them — sit in the corner where the buyer is most able to build their own. The ones that are safe are boring on purpose.
Nothing on this page is both. The candidates that pay within the quarter have ceilings you could hit and then sit at for a decade; the ones with ceilings worth the word take years to produce a dollar. The dashed line joins the ones where nothing else is both sooner and bigger — everywhere else on the chart, something dominates you.
The channel is not a marketing decision made after the build. It sets the sales cycle, the cost per customer and the ceiling, and it is fixed by who the buyer is long before you write any code. This is the column to read first.
| Channel | What you need for it | Wait | Per customer | Uses it | Who it reaches | Yours? |
|---|
| Candidate | Family | Build hrs | First $ | Cash in | Mature net | Payback | Odds | Answer | Build | Moat |
|---|
The tax is computed, not assumed. Year-three net runs through the same self-employment calculator as worklets.ai/self-employment-tax-calculator: 92.35% of net income as the SE base, 12.4% Social Security to the 2026 wage cap plus 2.9% Medicare and the 0.9% surcharge, the deductible half, then 2026 federal brackets and the real state schedule for whichever state you picked. That is why the numbers move properly with revenue instead of sitting at a flat effective rate.
Adoption is a bounded curve, not a hockey stick. After the first
customer, the count approaches the stated ceiling as 1 − e^(−t/τ),
where τ is between six months for an audience and eighteen for procurement. It
deliberately cannot produce a surprise, because a model that can produce a
surprise will, and you will believe it.
Reach is the load-bearing assumption. Time to the first customer is the channel's lead time plus the buyer's sales cycle. Having the channel an idea needs multiplies the wait by 0.45 and the cost per customer by 0.4; having no channel at all multiplies them by 1.35 and 1.3. Those four numbers move this page more than anything else on it, and they are judgement rather than measurement. They are stated here so you can disagree with them precisely.
The page draws a shortlist; the catalogue is bigger. There are candidates behind this and the figures draw of them by default, because a row chart with seventy rows is a chart nobody finishes. Reachable candidates are taken first regardless of score — "you could actually start this" beats "this scores slightly higher and you cannot" — and the rest of the slots are filled by fit. Nothing is hidden: the scatters plot every candidate, the table has a button for the full list, and the control above sets the number.
The self-build score is exposure, not quality. Four tests, three points each: whether the buyer writes code, whether the data and access are already theirs, whether it stays built once built, and whether anybody has to be accountable when it is wrong. Twelve means they will make their own; zero means they could not if they wanted to. The chart plots twelve minus that, so both tests point the same way. These are judgements about buyers, and the one worth arguing with is "the buyer writes code" — it is a claim about a whole category of person, and it is getting truer every year.
What is not modelled. Competition arriving, price changing, churn, hiring, raising money, and the fact that you will enjoy some of these and resent others by month four — which in practice decides more of these outcomes than any column here. Nothing on this page knows what you would actually be good at.
The odds column is not a forecast. It is a prior on reaching $10,000 a month at all, used only to produce the discounted figure on each card. Roughly 70% of small software products never clear $1,000 a month and only a few percent pass $50,000, so an optimistic-looking 30% here is already a claim that the candidate is well above the base rate.
Related tools on worklets.ai: self-employment tax, 1099 against W-2, how long the money lasts, and the companion piece on where to live, which works the same way on a different decision.