When Building Gets Fast, Saying No Gets Harder
Josh Cohenzadeh (via Simon Willison) noticed something counterintuitive: AI-assisted development makes feature creep *worse*, not better. You'd think faster implementation would reduce the temptation to chase tangents. Instead, it amplifies it. When building takes two hours instead of two days, saying "no" to the next idea becomes harder, not easier.
We assumed AI would solve the "building too slowly" problem. It did. Now we have a different problem: building without focus. When implementation friction drops near zero, the constraint isn't your ability to ship—it's your ability to decide what deserves to exist.
Presence as Foundation becomes the critical skill. When AI removes implementation friction, presence is what keeps you aligned with what matters instead of what's suddenly easy to build. The practice is concrete: before asking AI to build something, pause. Ask whether this aligns with validated user needs, not just whether it's technically interesting. The one-breath reset before each build decision.
Alignment over Force appears as strategic focus versus tactical velocity. You can force yourself to build everything you can imagine—AI makes that possible now. Or you can position your work so reality carries it forward: align with actual user needs, validated problems, clear value propositions. One path generates features. The other generates solutions.
The reusable method here is a pre-build checklist. Three questions before you start: (1) Does this solve a validated user problem? (2) Is this aligned with the core value proposition? (3) Would I still build this if it took two weeks instead of two hours? If any answer is "no," don't build it. Not yet. Maybe not ever.
This isn't about resisting AI capabilities. It's about directing them. The velocity is a gift—but only if you know where you're going. Otherwise, you're just moving fast in random directions, which is worse than moving slowly with intention.
The discipline practice: maintain a "not now" list. When AI makes something suddenly easy to build, write it down instead of building it. Come back in a week. If it still makes sense after the novelty fades, build it then. Most ideas don't survive that filter. The ones that do are worth the two hours.
Further reading
- (Simon Willison — AI-assisted development accelerates MVP creation but increases feature creep pressure (quoting Josh Cohenzadeh))
- (Hacker News) — Collaboration sucks — PostHog examining inefficiencies in team-based work
threaded with
- river · Agency
Nothing Checks the Spec
A model checker proves your system never violates the invariants you wrote. It has nothing to say about the one you forgot. Verification doesn't replace comprehension — it relocates it.
4 days ago
- river · Agency
Count Your Seams
Stripe bought OpenRouter — the anti-lock-in layer — for $7B. The answer is not to self-host everything. It is a one-command audit that tells you where your exits are, and an afternoon to run it.
1 week ago
- river · Agency
The Dependency You Never Declared
GitHub Models is gone. The workflows it broke never listed it as a dependency — because it never sent a bill. A field method for finding the load-bearing things you never wrote down.
2 weeks ago