"I don't want to become a very expensive Enter key" 😹 This hit home!
Honestly, I'm not a great developer. I get a lot of help from AI, and just like you said, I use it to understand. I still don't know everything about how my own two projects are built. I tried reading a bit of your Rust code, but I needed AI to explain it to me 😅
But I truly believe one thing: you have great skills and great ideas, and they will take you further than AI, because you're the one giving the commands.
For me, writing documentation for my projects is how I learn: I understand them better and I keep finding new bugs to fix... and so on! 🐱🦀💚
fnmain(){letme=String::from("JS cat 🐱");ask_ai(&me);// borrowing, not moving! I'm learning 😹println!("{} still here, still learning 🦀",me);}
And honestly, using AI to understand our own code is still learning. What matters is that you keep asking why things work instead of just accepting the output.
Writing documentation is a really good way to do that too. If you can explain the code clearly, you usually understand it much better — and like you said, you often find bugs along the way. 👀
Also, your Rust example made me laugh. 😹🦀
The JS cat survived the borrow checker!
Creator of GithubStarsManager — an open-source, local-first tool that organizes GitHub stars with AI: auto-categorization, semantic search, release tracking, and an MCP server for star library.
The "type-tune got faster, its developer got slower" line is the most honest part of this. AI collapsing the cost of writing Rust while the review cost stays the same is exactly the bottleneck I keep running into too — "compiles and passes tests" doesn't tell me whether the abstraction is right, and the .clone() question is a perfect example of a decision that looks trivial but isn't.
The Scan / Plan / Apply split, with every candidate [ignore] by default, is the design choice I'd steal: it makes the tool's job "surface the decision" instead of "make the decision", and rolling back when tsc validation fails is the safety net that makes apply feel trustworthy rather than scary. Congrats on the release — and on refusing to become a very expensive Enter key.
Creator of GithubStarsManager — an open-source, local-first tool that organizes GitHub stars with AI: auto-categorization, semantic search, release tracking, and an MCP server for star library.
Ha, the expensive Enter key line got me too — and honestly it's the exact tension I keep running into with my own tooling. Automate the mechanical parts, keep the risky parts reviewable. Good luck with the Rust side of it, and thanks for writing the trade-offs up so honestly.
That balance is exactly what I’m trying to keep: automate the mechanical work, but make sure the parts that can actually change the architecture stay visible and reviewable.
And thanks for the encouragement on the Rust side too. I still have a lot to learn, but that’s part of the fun. 🦀
🎓 Self-taught full-stack developer | 💻 Python, JavaScript, Django, React
🇮🇷 From Iran | 💡 Passionate about building tools that empower people
🧠 AI enthusiast | 🤖 Dreaming of ethical, creative,
I research AI tools, APIs & agents — comparing specs, pricing and real user reports — then publish honest rankings so you don't waste money. Research-first, never sponsored fluff.
That line about not wanting to become a very expensive Enter key is the whole thing, isn't it. The clone-vs-borrow review problem is the part that would slow me down most — AI handing you code that compiles but you can't tell if the .clone() is load-bearing. Did the rewrite change how you review AI code in TypeScript too, or is it just the Rust that's slower to check?
Yeah, Rust is definitely the slower one for me to review. 😹
In TypeScript, I can usually spot questionable choices pretty quickly because I already have a strong mental model for the language and ecosystem.
But the Rust rewrite did make me more suspicious of AI-generated code in TypeScript too.
“Compiles” and “passes tests” are useful signals, but they still don’t tell me whether the abstraction, API boundary, dependency direction, or ownership of responsibility is actually right.
So Rust mostly slowed down the review itself, but it probably made my TypeScript reviews a bit more skeptical too. 👀
And yeah, the .clone() question is exactly the kind of thing that makes Rust review take longer for me. 🦀
Memory safety and zero-cost abstractions only deliver their promise when you understand the underlying cache hierarchy and syscall overhead.
Rewriting services without profiling memory allocation patterns just trades one set of bottlenecks for another. Systems engineering always comes down to mechanical sympathy with the hardware.
I am going to ask you 1 thing. where on this planet, earth and sky did you learn rust. It still manages to elude me to this day. and so i got the unholy deity AI to teach me slowly. also,
Keeping the original module as a compatibility re-export is a good safety valve. One edge I’d be curious about: with project references or path aliases, does apply validate only the selected --project, or walk the referenced configs too? A leaf config can pass while a consumer still resolves the old path differently.
Right now, type-tune does not walk TypeScript project references.
If the selected tsconfig has non-empty references, scan/plan/apply are rejected before any rewrite happens. Path aliases are supported, but final validation still runs against the selected --project only.
So your leaf-config example is exactly why I don’t want to claim workspace-wide safety yet.
Supporting the full project-reference graph will need a broader analysis/validation model rather than just following references during apply. 🦀
Great write-up! The line “AI can generate Rust code much faster than I can understand it” names a problem many people feel but few say out loud. 😹
The review bottleneck is real. AI cuts the cost of writing code, but reviewing still costs the same. Reviewing also depends on how fluent you are in the language. In TypeScript you can spot a smell in seconds. In Rust you stop and look things up, as you described with .clone(), pub vs pub(crate), and macros. Seen that way, the bottleneck is the gap between the code you can generate and the code you can judge.
Your port had a big advantage: the TypeScript version. You had a working implementation, design docs, and tests. That means you had an oracle. A few things that could help you learn faster and trust the port more:
Run the old and new implementations against the same fixtures. Run both on a handful of real TS projects and diff the plan.json output. Because your candidate IDs are deterministic, this should be cheap. It checks the port without you having to read every line of Rust.
Let Clippy do the first pass of the ownership and visibility review. cargo clippy flags many of the things you described, such as redundant_clone and needless borrows. unreachable_pub and clippy::pedantic can also be enabled. When AI writes a .clone(), a lint can tell you whether it’s needed. That leaves your attention for the architectural decisions.
Use AI as a tutor, not only as a generator. Instead of “write this function,” try “write it, then explain every .clone() and pub and what would break without it.” Or ask it for a version with zero clones and have it explain the trade-offs. You write that this is the “expensive Enter key” trap, and this is the opposite habit.
The design philosophy is the strongest part. The Scan / Plan / Apply split, with every candidate [ignore] by default, is the same idea as your AI argument: automate the tedious work, not the decisions. Opt-in by default is a good choice for a tool that rewrites source files. It also means AI-assisted review stays advisory and never touches the code. I’d like to see more AI tools put the human decision point in a reviewable artifact like plan.md, instead of an in-chat “approve all.”
Rolling back when TypeScript validation fails is a nice safety net too. It makes apply feel more like a transaction than a codemod.
Some questions:
Can you share numbers on the speed-up? Which phase gained the most (parsing, reference resolution, or I/O) on a large monorepo?
How do you handle cases the tool can’t safely rewrite, such as export * chains, namespace imports, or declare module augmentations? Do they get skipped with a reason in the plan, or flagged?
Is there a “second pass” idea, such as suggesting destinations based on existing folder conventions, while keeping the human in charge?
On the closing request: Rust is mostly learned by reading the compiler’s error messages carefully. The borrow checker error you showed is a good example. You’re already doing the harder part, which is questioning why the code is written the way it is. Keep reviewing the diffs, and the crab will come. 🦀
Thanks for sharing, and congrats on the release! 🎉
I really like the way you described the old TypeScript implementation as an “oracle.” That’s exactly what made this rewrite practical.
I actually went back and benchmarked the old TypeScript version against the current Rust version after reading your question.
On a controlled fixture with 200 packages and 600 TypeScript files, where the scan results matched:
Command
TypeScript
Rust
Speed-up
scan
1,008 ms
90 ms
11.3×
plan
1,002 ms
96 ms
10.4×
reviewed apply --dry-run
1,868 ms
1,012 ms
1.85×
I also tested a real is-kit checkout:
Command
TypeScript
Rust
Speed-up
scan
713 ms
78 ms
9.1×
plan
754 ms
84 ms
8.9×
That said, I don’t think I can honestly say “Rust made phase X faster” yet.
The implementations changed quite a lot too: the TypeScript version used ts-morph / TypeScript findReferences(), while the Rust version uses tree-sitter and its own import/referrer analysis.
On one much larger real application, reference lookup in the old implementation was clearly the hotspot, but the two versions produced slightly different analysis results, so I don’t want to present its timings as an apples-to-apples benchmark.
So for now my conclusion is:
The rewrite is clearly much faster in comparable tests, but I can’t attribute that gain to Rust alone or say which phase improved the most on a large real codebase yet. 👀🦀
Also, I really like your suggestion of asking AI to explain every .clone() and pub instead of just generating the code. That’s exactly the kind of habit I want to build.
I write about Next.js + TypeScript + AI engineering, with a focus on making sites discoverable to LLMs and AI search. Long-form notes at mudassirkhan.me
the line about type tune getting faster while its developer got slower is the honest part of the post. having design docs and tests already in place is probably why the migration took under a month, most rewrites stall because the spec only lives in the old code. we moved one hot path off node once and the test suite as a contract saved us more than the language did. how are you checking the rust output matches the typescript version on large repos, snapshot diffs or something stricter?
Yeah, having the tests and design docs already in place helped a lot. 😸
For parity, I compare candidate IDs, referrers, plan output, and dry-run previews rather than relying on snapshots alone.
The controlled fixtures match well, but on one much larger real app the two versions still disagreed slightly, so I’m not treating that as a valid benchmark yet.
I’d rather expose the mismatch than hide it behind a snapshot. 👀
We're a place where coders share, stay up-to-date and grow their careers.
"I don't want to become a very expensive Enter key" 😹 This hit home!
Honestly, I'm not a great developer. I get a lot of help from AI, and just like you said, I use it to understand. I still don't know everything about how my own two projects are built. I tried reading a bit of your Rust code, but I needed AI to explain it to me 😅
But I truly believe one thing: you have great skills and great ideas, and they will take you further than AI, because you're the one giving the commands.
For me, writing documentation for my projects is how I learn: I understand them better and I keep finding new bugs to fix... and so on! 🐱🦀💚
😁😁
Thank you!! I love your comment. 😸💚
And honestly, using AI to understand our own code is still learning. What matters is that you keep asking why things work instead of just accepting the output.
Writing documentation is a really good way to do that too. If you can explain the code clearly, you usually understand it much better — and like you said, you often find bugs along the way. 👀
Also, your Rust example made me laugh. 😹🦀
The JS cat survived the borrow checker!
Let’s keep learning and building together. 🐱🙌
Thank you so much! 😸💚
The "type-tune got faster, its developer got slower" line is the most honest part of this. AI collapsing the cost of writing Rust while the review cost stays the same is exactly the bottleneck I keep running into too — "compiles and passes tests" doesn't tell me whether the abstraction is right, and the .clone() question is a perfect example of a decision that looks trivial but isn't.
The Scan / Plan / Apply split, with every candidate [ignore] by default, is the design choice I'd steal: it makes the tool's job "surface the decision" instead of "make the decision", and rolling back when tsc validation fails is the safety net that makes apply feel trustworthy rather than scary. Congrats on the release — and on refusing to become a very expensive Enter key.
Thank you! 😸💚
I really like your phrasing: “surface the decision instead of make the decision.”
That’s exactly what I want type-tune to do.
Automate the mechanical work, make the risky parts reviewable, and leave the architectural choice to the developer.
And yes, I’m still trying very hard not to become an expensive Enter key. 😹🦀
Ha, the expensive Enter key line got me too — and honestly it's the exact tension I keep running into with my own tooling. Automate the mechanical parts, keep the risky parts reviewable. Good luck with the Rust side of it, and thanks for writing the trade-offs up so honestly.
Thank you! 😸💚
That balance is exactly what I’m trying to keep: automate the mechanical work, but make sure the parts that can actually change the architecture stay visible and reviewable.
And thanks for the encouragement on the Rust side too. I still have a lot to learn, but that’s part of the fun. 🦀
omg i gave you star and followed you i loved it so cute 😍😍
YAY!!! 😸💚
Thank you so much for the star and the follow!!
I’m really happy you liked it. 🐈🦀✨
That line about not wanting to become a very expensive Enter key is the whole thing, isn't it. The clone-vs-borrow review problem is the part that would slow me down most — AI handing you code that compiles but you can't tell if the .clone() is load-bearing. Did the rewrite change how you review AI code in TypeScript too, or is it just the Rust that's slower to check?
Yeah, Rust is definitely the slower one for me to review. 😹
In TypeScript, I can usually spot questionable choices pretty quickly because I already have a strong mental model for the language and ecosystem.
But the Rust rewrite did make me more suspicious of AI-generated code in TypeScript too.
“Compiles” and “passes tests” are useful signals, but they still don’t tell me whether the abstraction, API boundary, dependency direction, or ownership of responsibility is actually right.
So Rust mostly slowed down the review itself, but it probably made my TypeScript reviews a bit more skeptical too. 👀
And yeah, the
.clone()question is exactly the kind of thing that makes Rust review take longer for me. 🦀Memory safety and zero-cost abstractions only deliver their promise when you understand the underlying cache hierarchy and syscall overhead.
Rewriting services without profiling memory allocation patterns just trades one set of bottlenecks for another. Systems engineering always comes down to mechanical sympathy with the hardware.
Absolutely. 👀
That’s also why I’m trying not to frame the rewrite as simply “Rust made it faster.”
The Rust version changed more than the language itself:
parser choice, reference analysis, data structures, and execution strategy all changed too.
In controlled benchmarks it is clearly much faster, but profiling the real bottlenecks still matters if I want to understand why.
Otherwise I’m just moving the bottleneck somewhere else. 😹🦀
I am going to ask you 1 thing. where on this planet, earth and sky did you learn rust. It still manages to elude me to this day. and so i got the unholy deity AI to teach me slowly. also,
I didn’t really “learn Rust first and then build it.” I built it while learning Rust, with AI acting as my very patient tutor. 😹🦀
So I’m still learning too. 👀
And yes,
May the fluff be with you. 🐈✨
Keeping the original module as a compatibility re-export is a good safety valve. One edge I’d be curious about: with project references or path aliases, does apply validate only the selected --project, or walk the referenced configs too? A leaf config can pass while a consumer still resolves the old path differently.
Good catch. 👀
Right now, type-tune does not walk TypeScript project references.
If the selected
tsconfighas non-emptyreferences, scan/plan/apply are rejected before any rewrite happens. Path aliases are supported, but final validation still runs against the selected--projectonly.So your leaf-config example is exactly why I don’t want to claim workspace-wide safety yet.
Supporting the full project-reference graph will need a broader analysis/validation model rather than just following references during
apply. 🦀Great write-up! The line “AI can generate Rust code much faster than I can understand it” names a problem many people feel but few say out loud. 😹
The review bottleneck is real. AI cuts the cost of writing code, but reviewing still costs the same. Reviewing also depends on how fluent you are in the language. In TypeScript you can spot a smell in seconds. In Rust you stop and look things up, as you described with .clone(), pub vs pub(crate), and macros. Seen that way, the bottleneck is the gap between the code you can generate and the code you can judge.
Your port had a big advantage: the TypeScript version. You had a working implementation, design docs, and tests. That means you had an oracle. A few things that could help you learn faster and trust the port more:
Run the old and new implementations against the same fixtures. Run both on a handful of real TS projects and diff the plan.json output. Because your candidate IDs are deterministic, this should be cheap. It checks the port without you having to read every line of Rust.
Let Clippy do the first pass of the ownership and visibility review. cargo clippy flags many of the things you described, such as redundant_clone and needless borrows. unreachable_pub and clippy::pedantic can also be enabled. When AI writes a .clone(), a lint can tell you whether it’s needed. That leaves your attention for the architectural decisions.
Use AI as a tutor, not only as a generator. Instead of “write this function,” try “write it, then explain every .clone() and pub and what would break without it.” Or ask it for a version with zero clones and have it explain the trade-offs. You write that this is the “expensive Enter key” trap, and this is the opposite habit.
The design philosophy is the strongest part. The Scan / Plan / Apply split, with every candidate [ignore] by default, is the same idea as your AI argument: automate the tedious work, not the decisions. Opt-in by default is a good choice for a tool that rewrites source files. It also means AI-assisted review stays advisory and never touches the code. I’d like to see more AI tools put the human decision point in a reviewable artifact like plan.md, instead of an in-chat “approve all.”
Rolling back when TypeScript validation fails is a nice safety net too. It makes apply feel more like a transaction than a codemod.
Some questions:
Can you share numbers on the speed-up? Which phase gained the most (parsing, reference resolution, or I/O) on a large monorepo?
How do you handle cases the tool can’t safely rewrite, such as export * chains, namespace imports, or declare module augmentations? Do they get skipped with a reason in the plan, or flagged?
Is there a “second pass” idea, such as suggesting destinations based on existing folder conventions, while keeping the human in charge?
On the closing request: Rust is mostly learned by reading the compiler’s error messages carefully. The borrow checker error you showed is a good example. You’re already doing the harder part, which is questioning why the code is written the way it is. Keep reviewing the diffs, and the crab will come. 🦀
Thanks for sharing, and congrats on the release! 🎉
Thank you for such a thoughtful comment! 😸💚
I really like the way you described the old TypeScript implementation as an “oracle.” That’s exactly what made this rewrite practical.
I actually went back and benchmarked the old TypeScript version against the current Rust version after reading your question.
On a controlled fixture with 200 packages and 600 TypeScript files, where the scan results matched:
scanplanapply --dry-runI also tested a real
is-kitcheckout:scanplanThat said, I don’t think I can honestly say “Rust made phase X faster” yet.
The implementations changed quite a lot too: the TypeScript version used ts-morph / TypeScript
findReferences(), while the Rust version uses tree-sitter and its own import/referrer analysis.On one much larger real application, reference lookup in the old implementation was clearly the hotspot, but the two versions produced slightly different analysis results, so I don’t want to present its timings as an apples-to-apples benchmark.
So for now my conclusion is:
The rewrite is clearly much faster in comparable tests, but I can’t attribute that gain to Rust alone or say which phase improved the most on a large real codebase yet. 👀🦀
Also, I really like your suggestion of asking AI to explain every
.clone()andpubinstead of just generating the code. That’s exactly the kind of habit I want to build.Thanks again for such a detailed comment! 🙌
I think porting a working tool is a brave way to learn a new language properly.
Yeah, I think that helped a lot! 😸
Having a working implementation, tests, and a design to compare against made Rust much easier to learn than starting from a blank project.
Still painful sometimes though. 😹🦀
the line about type tune getting faster while its developer got slower is the honest part of the post. having design docs and tests already in place is probably why the migration took under a month, most rewrites stall because the spec only lives in the old code. we moved one hot path off node once and the test suite as a contract saved us more than the language did. how are you checking the rust output matches the typescript version on large repos, snapshot diffs or something stricter?
Yeah, having the tests and design docs already in place helped a lot. 😸
For parity, I compare candidate IDs, referrers, plan output, and dry-run previews rather than relying on snapshots alone.
The controlled fixtures match well, but on one much larger real app the two versions still disagreed slightly, so I’m not treating that as a valid benchmark yet.
I’d rather expose the mismatch than hide it behind a snapshot. 👀