I have seen a couple threads on twitter where people complain about the difficulty with publishing NPM libraries or ask what starter kit they should use (or, people recommended starter packs anyways)

Example threads

One thing that is notable to me in these threads is that people often recommend that you use a bundler (a program that combines multiple src files into a single or fewer output files) when developing a library

Examples of starter packs suggested in these threads that use bundlers

Not using bundlers

In summary 2/15 do not use a bundler, 13/15 do use a bundler. Sidenote: webpack notably absent

Why would you NOT want a bundler for your library?

My main argument is that the consumer of your library is the one that should use a bundler if it is relevant to them. If the library uses a bundler:

• in the best case, it has no impact on the consumer
• in the worst case, it affects the complexity of your library and makes possible limitations for your consumers also.

An example where it can actually create limitations, you might consider code splitting with async import(). If you create a single file bundle, then the consumer of your library may not be able to do code splitting properly via async import()

Why would you MAYBE want a bundler for your library

If you really care about producing a UMD bundle that can be used in a script tag, maybe you want a bundler, but the future does not seem to be in UMD. One other possible bundle type is maybe you like the idea of a single file ESM module. It is similar where you could maybe reference this from a script tag with type module, but this seems like a niche usage. For example, you would still have to consider:

• If you are not bundling dependencies, then what is the benefit of using a bundler?
• If you are bundling dependencies, you are not allowing people to get updates to your sub-dependencies with semver!

Add-on: Another concern brought up by users in discussion thread: There is a cost to having many small files, e.g. in app startup cost on serverless or any nodejs application to loading many small files off disk. To me, this is an app level concern, similar to bundling for the browser though.

My suggestion: no bundler, no starter pack, just tsc

I'd recommend just compiling your code with tsc, no bundler involved. This way, you can develop with typescript, it will output js files, and you can directly deploy a dist folder of js files to NPM.

Example package.json

{
  "name": "yourlib",
  "version": "1.0.0",
  "main": "dist/index.js",
  "scripts": {
    "clean": "rimraf dist",
    "prebuild": "npm run clean",
    "build": "tsc",
    "preversion": "npm run build",
    "postversion": "git push --follow-tags"
  },
  "files": ["dist", "src"],
  "devDependencies": {
    "rimraf": "^3.0.2",
    "typescript": "^4.6.2"
  }
}

Features of the above package.json

• We can use a single command, yarn publish to publish to npm
• The single yarn publish automatically runs clean and build via preversion, then postversion, which pushes the tag to the remote repo
• The "files": ["dist", "src"] refers to publishing the dist and src directories, and src is used for the sourceMap

Example tsconfig.json

{
  "include": ["src"],
  "compilerOptions": {
    "target": "es2018",
    "outDir": "dist",
    "lib": ["dom", "esnext"],
    "declaration": true,
    "moduleResolution": "node",
    "sourceMap": true,
    "strict": true,
    "esModuleInterop": true
  }
}

Features of the above tsconfig.json

• Uses "moduleResolution": "node" - this is not pure-ESM because pure-ESM expects you to import filenames with their file extension, while node module resolution can import extensionless paths, but node module resolution generally works well with consumers that use bundlers themselves
• Uses "target": "es2018" - This is does a small amount of transpilation of super modern features, but would generally not require your users to babel-ify their node_modules if they consume your library

What about testing?

Adding testing is not immediately solved by the above, but bundling doesn't really help testing anyways. It's just a starter pack feature we can add on. Some options you have include

• Use ts-jest
• Use jest on it's own, plus a babel config with @babel/preset-typescript
• Use jest on it's own, run over the compiled output without a babel config
• Use vitest
• Possibly something else? Node now has a built-in test runner. It will be a slow road to adoption but might become more popular over time https://fusebit.io/blog/node-testing-comes-to-core/

Conclusion

It is tempting to have nice zero-config solutions and starter kits, but to me, it is not really beneficial to use the bundler aspect of many of these for publishing to NPM. Am I wrong? Let me know if I am.

Also, these starter kits may not be maintained for perpetuity. Our team used tsdx for some time, but it was not maintained well, and used old typescript version 3.x, and it ended up being hard to remove from our codebase. Learning the basic tools like tsc will help

Footnote 1: Shipping pure-ESM

Do you want to make a pure-ESM package? Then you do not want to use "moduleResolution": "node" in tsconfig.json, and you will want to set "type": "module" in package.json. You may also need to explicitly import with .js extensions in your source code, even if you write .ts. This is awkward, and something the community is still grappling with.

If you have ever stumbled on this topic, you will probably want to see this link https://gist.github.com/sindresorhus/a39789f98801d908bbc7ff3ecc99d99c

Footnote 2: Learning your tools

Several people on the my recent post post suggested that I did not understand my tools, and that I should just learn the tools correctly and I wouldn't have the problem I had.

Indeed, what I am now telling people to do in this post is similar: I am saying "just use tsc by itself! Understand you tools! You may have to do more research and create more boilerplate, but it's better".

Ultimately though, it's up to you to choose your tools and starter packs and whatnot.

Footnote 3: What does it look like when you compile with tsc?

When I refer to compiling with tsc above, I compile a src directory into a dist directory

So if I have:

src/index.ts
src/util.ts
src/components/Button.ts

Running tsc will output:

dist/index.js
dist/index.d.ts
dist/index.js.map
dist/util.js
dist/util.d.ts
dist/util.js.map
dist/components/Button.js
dist/components/Button.d.ts
dist/components/Button.js.map

Then, the dist and src directories are published to npm which enables the sourceMaps to work.

Note: We do not need to explicitly say where the typescript types are with "types" in package.json, many starter packs do this but it is unneeded for this package as the d.ts files are automatically found.

See https://cmdcolin.github.io/posts/2021-12-31-npm-package for my article on creating a typescript package for npm

Footnote 4: Other things people recommend

Other things people recommend in the starter pack threads

• deno - not clear to me how this helps when deploying to npm, but I still gotta try deno
• ts-node - not clear to me how this helps when deploying to npm
• nx - heavy monorepo-based solution
• elm - that's just a different thing
• Typescript-Node-Starter - https://github.com/microsoft/TypeScript-Node-Starter - not a library, full stack framework
• joystick - https://github.com/CheatCode/joystick - not a library, full stack framework

Footnote 5:

Similar things happen if you take on dependencies of starter kits like create-react-app. You become very committed to their particular way of doing things, and can only modify their config with things like rescripts, craco or rewired. If you crafted your setup with just webpack, you may not be so tied down. But, I still use things like create-react-app because they do seem to help me significantly. Now though, the tide seems to be turning other directions like next.js which can do static site generation at a basic level and extend to multiple pages more easily.