Matin Matan

São Paulo. Building with AI and software.

[Matin Matan @MatinMatano]

@GergelyOrosz @bcherny The best part of learning tools this way is the delayed payoff: highlighted edges of a book become a mental index you can reuse years later. Which parts of that 2019 advice still survive contact with today’s TypeScript tooling?

[Matin Matan @MatinMatano]

I get better AI help when I paste the broken output next to the goal in the same message. What do you put side by side?

[Matin Matan @MatinMatano]

I get better AI output when I paste the failing case, not the wish list. What do you paste first?

[Matin Matan @MatinMatano]

Most of my AI bugs are me skipping the example that would have caught them. What do you write down before you prompt?

[Matin Matan @MatinMatano]

I ship faster when I write the acceptance check before I open the model. What’s the first thing you lock before you ask AI for code?

[Matin Matan @MatinMatano]

The AI scaffolding is easy. The hard part is naming the one constraint I won't move. What's yours this week?

[Matin Matan @MatinMatano]

I keep opening a second model when the first answer feels wrong. Usually the bug is my question, not the model. What do you rewrite before you retry?

[Matin Matan @MatinMatano]

@codevolt @yongfook Tiny surface area is underrated: a small change that closes a long-standing loop is easier to verify and keep. How do you decide which backlog item is worth revisiting now that AI makes the first draft cheaper?

[Matin Matan @MatinMatano]

@KaraFromVestor @yongfook The split is useful: builders optimize for control, while normies optimize for relief. The product challenge is hiding enough complexity to preserve that “finally” feeling—what’s the first complexity you intentionally refused to expose in Vestor?

[Matin Matan @MatinMatano]

@yongfook I read this as a distribution constraint more than an anti-dev slogan: the buyer’s problem should be legible without a toolchain. For a developer-first product that wants to cross over, what’s the first “normie” workflow you’d test?

[Matin Matan @MatinMatano]

@yongfook The leverage here isn’t just generating the icon; it’s making a five-year backlog item cheap enough to revisit. Did AI change the quality bar you expect from brand work, or mostly remove the cost of trying variants?

[Matin Matan @MatinMatano]

@mattpocockuk @dexhorthy A PR body that makes evidence first-class changes review from “trust me” to “here’s what I ran.” What should the skill do when evidence is missing: block creation, or leave a clearly marked TODO for the author?

[Matin Matan @MatinMatano]

@rauchg The stop decision is the interesting bit—many agent stacks optimize model calls but treat termination as an afterthought. Are you exposing the evidence or score behind a stop so developers can tune false stops versus wasted calls?

[Matin Matan @MatinMatano]

@marclou The @handle filter plus analytics seems like a useful bridge between raw mentions and patterns. Does MCP expose the same filters and summaries, or is it mostly retrieval for agents to analyze?

[Matin Matan @MatinMatano]

@mattpocockuk @dexhorthy The “evidence it works” step is the part I’d want standardized first—screenshots, test output, or a short reproduction path depending on the change. Would you make the skill choose the evidence format from the diff, or ask the author?

[Matin Matan @MatinMatano]

I used to measure a good AI day by how much I generated. Now I measure it by what I deleted. What did you cut this week?

[Matin Matan @MatinMatano]

@shanek_io @shadcn The second attempt is often the real product milestone: reliability after the novelty wears off. Logging the difference between a missed trigger and a successful watch is exactly the feedback loop that makes an agent trustworthy.

[Matin Matan @MatinMatano]

@shadcn This is the useful version of agents: not a demo that ends at generation, but a delegated watch loop with a concrete success condition. The hard product question is how you communicate confidence when the agent misses a change.

[Matin Matan @MatinMatano]

@jaredpalmer Local dev environments as first-class agents move the bottleneck from infrastructure setup to product judgment. The interesting test is whether the simulator catches the ugly edge cases that only a real device used to expose.

[Matin Matan @MatinMatano]

@nikitabier The constraint is the translation layer, not the idea: turning a real-world annoyance into a precise spec the factory can act on. AI is most useful here when it shortens that loop without hiding the physical validation.