Mihai Sandu

Micro-SaaS ideas — validated, priced, no hype. A new one every week 👇

[Mihai Sandu @mihais77]

I built a working Zapier→Make converter in 4 minutes for about 60 cents. It passed every test and imported clean. It was also silently corrupting data on every real workflow. Here's the bug that almost shipped:

[Mihai Sandu @mihais77]

@arvidkahl in B2B it's a nice-to-have for the buyer and a quiet deal-breaker for the daily user. rarely the same person, which is why so much enterprise stuff stays light-only.

[Mihai Sandu @mihais77]

@arvidkahl we didn't lose the speed. we lost permission to be slow.

[Mihai Sandu @mihais77]

@arvidkahl yeah, I have built in the last week my own "control center" dashboard that centralizez data from all of my active places. A.M.A.Z.I.N.G that I could do that in a few hours with claude code

[Mihai Sandu @mihais77]

@damonchen yeah. MRR's a proxy, revenue's the thing. annual prepays and one-off setup fees barely move the MRR line but they're the cash that keeps the lights on.

[Mihai Sandu @mihais77]

@hnshah wrote the spec for my agent three times and it still missed the point. will try to see

[Mihai Sandu @mihais77]

@joshelman @a16z Congrats Josh. genuinely curious where you draw the 'consumer company' line now — half the breakout apps this year are one dev, an AI wrapper, and a tiktok account. new consumer, or the thing it eats?

[Mihai Sandu @mihais77]

@yongfook yeah, and watch where the chat went: into DMs. now half the decisions live in 1:1s nobody can search six months later. the channel was the searchable record.

[Mihai Sandu @mihais77]

@gregisenberg point 2 is the whole ballgame. mine forgets what we did yesterday unless i wire the memory layer up by hand. the 'two-year agent' gets real the day that stops being my job and it doesn't eat too many tokens.

[Mihai Sandu @mihais77]

Link in the replies 👇 I tear down a micro-SaaS idea like this every week — the build, the bug, and whether it's actually worth shipping:
https://t.co/bEvUmiQWVE

[Mihai Sandu @mihais77]

I built a working Zapier→Make converter in 4 minutes for about 60 cents. It passed every test and imported clean. It was also silently corrupting data on every real workflow. Here's the bug that almost shipped:

[Mihai Sandu @mihais77]

Final: 12 tests green, ~4 minutes, ~$0.60. The lesson isn't "AI builds fast." It's that the edge-case test is the only thing standing between "looks shipped" and "actually works." Validated price: $19/workflow. An agency moving 10 Zaps clears that instantly.

[Mihai Sandu @mihais77]

That's the dangerous kind of bug: not a crash — a silent pass. Every labeled form/Sheets field would resolve to nothing, and the migration would still look successful. The fix was one regex group: [A-Za-z0-9_.]+ → [^{}]+?. One character of bad assumption

[Mihai Sandu @mihais77]

Python stdlib, zero dependencies. Two passes: assign new sequential IDs and build an old→new map, then rewrite every reference token. Wrote the happy-path test first — 7 assertions. Implemented to green in about a minute. Felt done.

[Mihai Sandu @mihais77]

What it does: take a Zapier export, renumber every step to Make's format, and rewrite each {{step.field}} reference to the new IDs. Miss one reference and that step quietly points at a field that no longer exists. Deterministic JSON in, JSON out.

[Mihai Sandu @mihais77]

@damonchen The version that saved me: crawl the whole docs site to markdown once, drop it in a repo, let the agent grep all of it. Far fewer hallucinated API params than pasting pages one by one — it sees how things connect, not just one page.

[Mihai Sandu @mihais77]

@hnshah The dangerous part: 'shippable-looking' clears code review precisely because it's clean. The misses are never syntax — they're a missing edge case or a wrong default that only bites in prod.