Picking a SaaS idea in the age of LLMs
I'm trying to ship 10 SaaS projects this year and honestly the hardest part isn't building. It's choosing what to build.
Coding agents made code cheap, but validation is still painful. So I wrote down the filter I'm using before I commit to anything.
Roughly: can I get to MVP in 2 weeks. Is there an obvious place to promote it (a subreddit, an ad audience, a platform like RapidAPI). Will I get a real signal within 6 months. And is it different enough from my other projects that I'll actually learn something new.
I wrote up the whole thing (criteria, the workflow, a couple of diagrams) over here if you wanna see it: https://codesolo.substack.com/p/how-to-pick-a-saas-idea-in-the-age
Curious what your dealbreaker is when you look at a new idea. Mine's time to signal. If I can't tell within 6 months whether it's working, I don't start.
Time to signal is the right dealbreaker. But the question I now ask before that is, who is already paying to solve this badly?
If someone is using a spreadsheet, a Zapier hack, or a VA to do something that should be software, that is a better signal than any market research. The pain is proven. The budget already exists. You are just building a better version of what they are already doing.
The 10 projects approach is interesting too. We build mobile apps at our agency and the clients who shipped something narrow and specific almost always outperform the ones chasing a broad market. The niche ones find their audience faster and get real feedback before they run out of runway.
Curious, are you validating the distribution channel before building or after MVP?
The filter is mostly right, but “can I build it in 2 weeks?” is probably the least important part now.
That used to be a real constraint.
Now it mostly just filters for what is easy to ship, not what is expensive to ignore.
The stronger screen is usually:
does the problem already cost someone enough that they are solving it badly right now?
That tends to predict distribution, urgency, and willingness to pay better than build speed.
Fast MVPs are useful.
But “easy to build” has become a weak proxy for “worth building.”
That’s the part LLMs changed most.
The dealbreaker for me is a free alternative that already does it well enough. Not a paid competitor — a free one. If someone can Google the problem and find a solid free tool in the first three results, I stop the evaluation there regardless of how exciting the idea feels. Excitement doesn't survive free.
I built a 5-dimension scoring system for new ideas — speed to market, revenue potential, differentiation, audience fit, build cost, scored 1–3 each. Anything with a well-established free alternative is an automatic discard before scoring even starts. Learned that after building products that felt differentiated until someone pointed out the free version.
Your "obvious place to promote" filter is underrated and I'd put it higher in the stack. I built four digital products before realising I had no channel where my buyers actually hung out. The product quality didn't matter — I was promoting in the wrong place entirely. Channel clarity before build would have saved me a month.
"Does the problem already cost someone enough that they are solving it badly right now?" is the filter that matters.
The one I'd add: is the workaround invisible to the person doing it?
I built a workflow system for non-technical vibe coders — session protocol that keeps Claude and Cursor in context across sessions. The workaround I was replacing was 30-40 minutes of re-explaining context every morning. People were doing it daily and had completely normalized it. Didn't think of it as a problem worth solving. Just thought "this is how AI works."
That invisibility is actually a stronger signal than visible pain. Because when you name it — "you're losing 40 minutes every morning rebuilding context" — the reaction is immediate recognition, not education.
Time to signal is real. But invisible normalized pain converts faster than visible pain that people have already learned to live with.
The honest filter I use now is whether the LLM can do the boring 80% but cannot survive the last 20%. If a foundation model alone solves the problem end-to-end, your moat is whichever lab is shipping that week. I tried building two SaaS in the last year that did not pass that test, the model improved before I shipped and the work evaporated. The ones that survive are usually domain workflows where the LLM is one node in a pipeline that owns data, history, and a feedback loop. Pick the workflow first, the model second.
tbh the LLM-as-idea-generator stopped working for me pretty fast. asked it "what SaaS to build" across like 5 different projects and got the same productivity/AI-wrapper soup every time. flipped it the other way around — pulling raw rants from reddit/HN/PH and using LLM only as synthesis on top of that signal. way better. literally turned this workflow into a side product (discury) bcs i needed it for my own pipeline of SaaS launches. the bottleneck was never the model, it's always feedstock. what are you using as input?
My dealbreaker is different from yours and I keep going back and forth on whether it's right — I care less about time-to-signal and more about whether the first 10 users will actually talk to me. If I can't picture who they are and where I'd DM them tonight, I don't start. Six months feels generous in the LLM era because cloning cycles are like 3 weeks now.
The part of your filter I'd push back on a little: "obvious place to promote it" can be a trap. A clear distribution channel often means the audience is already saturated with AI wrappers fighting for the same subreddit. I've watched a few launches die that way — perfect ICP on paper, but the channel was burnt.
I'm Shirley, building ZooClaw (AI agents for solo founders) and wrestling with the same idea-selection problem in real time. What's been your hit rate so far on the projects you've shipped?
hit rate honestly depends how you count it. shipped a bunch of things over the years, 1 was a real win (hosting company i ran 18 yrs and exited in 2020), couple are still alive and growing slowly, few died quietly. but the dead ones taught me way more about ICP than any amount of upfront research bcs you only really see the channel saturation problem you mentioned once youre in the trenches. totally with you on the "can i DM the first 10 tonight" filter btw — thats the one i use now too. discury is actually pre-revenue rn, im in the "talking to SaaS founders before writing more code" phase and the convos are more useful than any pricing test would be at this stage. half the value is finding out why ppl wouldnt pay. zooclaw sounds like its in similar territory (solo founders is a noisy space tho) — hows your DM-the-first-10 list looking?