After three shutdowns, idea selection is no longer a creative act. It is a structured one. The cost of starting the wrong thing is now measured in years, not in weeks of energy. The rubric below is what I now run before starting anything. It is not glamorous. It is a series of filters, ordered by what they prevent, and most ideas die inside the first three.

The framework underneath this rubric is the same one in three startups, three shutdowns: 10x value, scaling potential, sustainable margins. The seven filters here are an operational version of those three checks plus four scaffolding filters that protect the founder from themselves.

The selection problem after three failures

The standard founder selection problem is "what do I find interesting and what does the market want". After three shutdowns, the question shifts to "what do I find interesting, what does the market want, and what failure modes am I most likely to repeat". The third question is the one that turns selection into a structured exercise.

The pattern across my three shutdowns: 10x value missed (MindWave), category timing missed (Super AI), unit economics missed (Vibe AI). Each was a different filter; the absence of any single filter was sufficient to fail. The rubric below ensures every filter gets checked, in order, before any code gets written.

The seven filters in order

Filter 1: 10x value test

Is the proposed product 10x better than the closest alternative on a dimension the user cares about, or is it incremental? Peter Thiel's framing in Zero to One is the canonical reference (Volt Equity summary). MindWave failed this filter; the alternatives existed and the new product was not measurably better. If you cannot articulate the 10x dimension in one sentence, the idea fails this filter and stops here.

Filter 2: scaling potential

Can the value scale linearly or super-linearly with users? Are there fixed-supply gates that would create handcrafting bottlenecks at scale? MindWave had this problem too: the warmth that made the product good was hand-crafted by the team, and the warmth was not a scalable supply. The filter rejects ideas where the value is dependent on a non-scalable resource.

Filter 3: sustainable margins

Does the unit economics work at one user, ten thousand users, and a million users? This is the filter Vibe AI failed. The cost-per-active-user did not balance against price at any scale because the price model was structurally mismatched to the cost curve. The filter requires a unit-economics sketch on a napkin before the idea moves to filter four. Detail is at economics of bootstrapped AI agents.

Filter 4: defensibility

What stops a well-resourced incumbent or a well-funded competitor from doing the same thing in eighteen months? Defensibility can come from data, distribution, integration depth, switching cost, or speed of the founder. Super AI had no defensibility because the underlying model was the moat and I did not own it. The filter rejects ideas with no path to a defensible position.

Filter 5: time-window read

Is the idea early, on time, or late for its category? Early ideas waste runway educating the market. Late ideas land in saturated competition. Super AI was late by the time the all-in-one assistant category started commoditising. The filter requires a clear read on where in the category cycle the idea sits, with a calendar reason if it is not on time.

Filter 6: founder fit

Am I the right person to build this for the next five years? Do I have the technical depth, the audience, the conviction to keep going through the slow eighteen months that come between launch and traction? Founder fit is the filter that founders skip. I skipped it for MindWave, partly. The filter requires honesty about whether the idea fits the founder's strengths and tolerance for the specific failure modes the idea will produce.

Filter 7: refusable scope

Can the scope be cut down to one capability that ships in week one and stays single-capability for the first six weeks? Ideas that require six capabilities to be valuable fail this filter. Super AI failed it. The all-in-one positioning was incompatible with refusable scope. The filter forces a single-capability v1.

Selection funnel: 100 candidate ideas in, 1 build out 100 candidate ideas 35 pass 10x value 20 pass scaling potential 10 pass sustainable margins 6 pass defensibility 4 pass time-window 2 pass founder fit 1 builds Source: Aryan Agarwal, founder rubric, 2026.
The funnel is intentionally narrow. The cost of starting the wrong thing is years.

The three-week evaluation

The full evaluation runs three weeks, never more. The clock matters because deliberation can become its own form of avoidance.

If the idea is still standing after three weeks, it gets a six-week ship-or-kill build cycle. Six weeks of building, end-to-end, single capability, with a unit-economics check at the end. The cycle is detailed in bootstrapping an AI agent platform in 2026.

The dead-idea log

Failed ideas do not get discarded. They get logged with the failed filter labelled. The log has three uses. First, it surfaces patterns; if I notice I keep killing ideas at filter five, I might be reading the time-window wrong. Second, it lets me revisit ideas when the failed filter changes; an idea that died on time-window in 2024 might pass it in 2027. Third, it is a memory system; when a similar idea shows up later, I can compare against the last version honestly.

The log is a markdown file with seven columns, one per filter. Each row is an idea, with the filter that killed it noted in red. Simple. Inspected monthly.

How Gravity passed every filter

Gravity, as the bet four idea, was the only one of dozens of candidates I evaluated in early 2026 that passed all seven filters cleanly.

Passing all seven on day one is necessary, not sufficient. The build still has to ship. But starting from a clean filter pass is a different game than starting from a half-failed one.

Frequently asked questions

How do you decide what to build after a shutdown?

A seven-filter rubric run in order. Most ideas die in the first three filters: 10x value, scaling potential, sustainable margins. The remaining four filters add nuance: defensibility, time-window, founder fit, and refusable scope. An idea that fails any filter does not start. The rubric protects against the missing checks I had in three previous shutdowns.

What is the most important filter?

Sustainable margins. The other filters describe whether the idea is interesting or possible. Sustainable margins describe whether the idea can survive long enough to matter. The lesson from Vibe AI is that loved products can fail this filter without warning. Every idea now gets a unit-economics sketch before any code is written.

How long do you spend evaluating an idea?

Three weeks of deliberate evaluation, never more. The first week is research and the 10x value test. The second week is the unit-economics sketch and the time-window read. The third week is the founder-fit and refusable-scope check. If the idea is still standing after three weeks, it gets a six-week ship-or-kill build cycle.

What do you do with ideas that fail a filter?

Log them with the failed filter labelled. Some ideas come back when the filter that killed them changes. The model-routing thesis behind Super AI failed the time-window filter in late 2024; the same thesis at a different layer might pass that filter in 2027. Logging the failure mode is what makes the idea reusable.

How does this rubric apply to Gravity?

Gravity passed every filter on day one. 10x value (autonomous outcome versus workflow-based alternatives), scaling potential (variable cost curve, no fixed-supply gating), sustainable margins (capability-based pricing aligned to cost-of-inference), defensibility, time-window aligned to the agent platform shift, founder fit, and refusable scope built into the weekly cadence.

Three takeaways before you close this tab

Sources