The Universe of Discourse

Claude and I write a utility program

Then I had two problems…

A few days ago I got angry at xargs for the hundredth time, because for me xargs is one of those "then he had two problems" technologies. It never does what I want by default and I can never remember how to use it. This time what I wanted wasn't complicated: I had a bunch of PDF documents in /tmp and I wanted to use GPG to encrypt some of them, something like this:

    gpg -ac $(ls *.pdf | menupick)

menupick is a lovely little utility that reads lines from standard input, presents a menu, prompts on the terminal for a selection from the items, and then prints the selection to standard output. Anyway, this didn't work because some of the filenames I wanted had spaces in them, and the shell sucks. Also because gpg probably only does one file at a time.

I could have done it this way:

    ls *.pdf | menupick | while read f; do gpg -ac "$f"; done

but that's a lot to type. I thought “aha, I'll use xargs.” Then I had two problems.

    ls *.pdf | menupick | xargs gpg -ac

This doesn't work because xargs wants to batch up the inputs to run as few instances of gpg as possible, and gpg only does one file at a time. I glanced at the xargs manual looking for the "one at a time please" option (which should have been the default) but I didn't see it amongst the forest of other options.

I think now that I needed -n 1 but I didn't find it immediately, and I was tired of looking it up every time when it was what I wanted every time. After many years of not remembering how to get xargs to do what I wanted, I decided the time had come to write a stripped-down replacement that just did what I wanted and nothing else.

(In hindsight I should perhaps have looked to see if gpg's --multifile option did what I wanted, but it's okay that I didn't, this solution is more general and I will use it over and over in coming years.)

xar is a worse version of xargs, but worse is better (for me)

First I wrote a comment that specified the scope of the project:

    # Version of xargs that will be easier to use
    #
    # 1. Replace each % with the filename, if there are any
    # 2. Otherwise put the filename at the end of the line
    # 3. Run one command per argument unless there is (some flag)
    # 4. On error, continue anyway
    # 5. Need -0 flag to allow NUL-termination

There! It will do one thing well, as Brian and Rob commanded us in the Beginning Times.

I wrote a draft implementation that did not even do all those things, just items 2 and 4, then I fleshed it out with item 1. I decided that I would postpone 3 and 5 until I needed them. (5 at least isn't a YAGNI, because I know I have needed it in the past.)

The result was this:

    import subprocess
    import sys

    def command_has_percent(command):
        for word in command:
            if "%" in word:
                return True
        return False

    def substitute_percents(target, replacement):
        return [ s.replace("%", replacement) for s in target ]

    def run_command_with_filename(command_template, filename):
        command = command_template.copy()
        if not command_has_percent(command):
            command.append("%")

        res = subprocess.run(substitute_percents(command, filename), check=False)
        return res.returncode == 0

    if __name__ == '__main__':
        template = sys.argv[1:]
        ok = True
        for line in sys.stdin:
            if line.endswith("\n"):
                line = line[:-1]
            if not run_command_with_filename(template, line):
                ok = False
        exit(0 if ok else 1)

Short, clean, simple, easy to use. I called it xar, ran

    ls *.pdf | menupick | xar gpg -ac

and was content.

Now again, with Claude

The following day I thought this would be the perfect opportunity to try getting some LLM help with programming. I already had a baseline version of xar working, and had thought through the problem specification. Now I could do it over with the LLM and compare the two results. The program I wanted was small and self-contained. If the LLM sabotaged me somehow, I would probably notice, and if I didn't, it wouldn't matter, because I had a working version already.

So I asked Claude about it. I knew Claude was not the right choice, because Claude is not really designed for this task, and has no integration with the coding editor, but Claude is what I am familiar with so I thought I'd give it a shot.

On the whole it went extremely well. The complete transcript is here. I imagine this was pretty much a maximally good experience, that all the terrible problems of LLM coding arise in larger, more complicated systems. But this program is small and self-contained, with no subtleties, so the LLM could be at its best.

Phase 1

First I just told Claude what I wanted and let it write the program from scratch.

Let's write a program in Python called xar. It will be a stripped-down version of the standard Unix xargs utility, but it will do less and be simpler to use.

It will get command-line arguments that specify a command template. Then it will read lines from stdin. Each line will be appended to the command template to produce a command. The program will run the resulting command, one command for each line of input.

Any questions?

Default Claude is very chatty and I am going to omit most of its chatty comments, which make me impatient.

The program it wrote it was not what I would have written, but it was good enough. If I had just used it right off the bat, instead of writing my own, it would have been enough, and it would have taken somewhere between 2% and 10% as long to produce.

So the one-line summary of this article is: I should stop writing simple command-line utilities, and just have Claude write them instead, because Claude's are good enough, and definitely better than yak-shaving.

Here's Claude's initial version:

    #!/usr/bin/env python3

    import sys
    import subprocess
    import shlex

    def main():
        if len(sys.argv) < 2:
            print("Usage: xar command [args...]", file=sys.stderr)
            sys.exit(1)

        command_template = sys.argv[1:]

        for line in sys.stdin:
            line = line.rstrip('\n')

            full_command = command_template + [line]

            try:
                result = subprocess.run(full_command, check=True)
            except subprocess.CalledProcessError as e:
                print(f"Command failed with exit code {e.returncode}: {' '.join(full_command)}", 
                      file=sys.stderr)
            except FileNotFoundError:
                print(f"Command not found: {full_command[0]}", file=sys.stderr)
                sys.exit(1)

    if __name__ == "__main__":
        main()

Claude's version had numerous comments, which I have omitted. I later told it to stop putting in comments, which it did.

Claude's use of check here was not what I wanted, because that makes subprocess.run raise an exception when the subcommand fails, and then the exception has to be immediately caught and handled. My original control flow had been simpler:

        res = subprocess.run(substitute_percents(command, filename), check=False)
        return res.returncode == 0

        …

        if not run_command_with_filename(template, line):
            ok = False

Claude's program pulled in shlex without noticing that it was unused. But Claude did teach me about str.rstrip('\n') which I had not known about before (or maybe had forgotten), so that was one small win already.

Argument parsing

The next step was a big win. Python's library for command-line argument handling is called argparse and it is really nice. If I were still writing programs in Perl, I would implement a Perl version of argparse because Perl has 29 available argument parsing libraries and they are all loathsome. The one problem with argparse is I never remember off the top of my head how to use it. I think the module is called argparse but it provides a class called Argparser but I often get these backward and try to use argparser and Argparse.

Instead of figuring it out every time I usually dig up some previous Python program and then copy-paste the argument parser from there, amending it to suit the purpose.

But this time I didn't have to do that. Instead, I just said to Claude:

This is good, now please add code at the top to handle argument parsing with the standard Argparse library, even though there are no options yet.

Claude handed me back pretty much the same program, but with the argument parser at the top:

    def parse_arguments():
        """Parse command-line arguments using argparse."""
        parser = argparse.ArgumentParser(
            description="A simplified version of xargs that appends each line from stdin to a command template.",
            prog="xar"
        )

        # Store all positional arguments as the command to execute
        parser.add_argument(
            "command", 
            nargs="+", 
            help="The command template to execute for each input line"
        )

        return parser.parse_args()

Let's pause for a moment. Maybe you kids are unimpressed by this. But if someone had sent this interaction back in time and showed it to me even as late as five years ago, I would have been stunned. It would have been completely science-fictional. The code that it produced is not so impressive, maybe, but that I was able to get it in response to an unstructured request in plain English is like seeing a nuclear-powered jetpack in action.

I make some changes

It bothered me that the command-line arguments were being taken implicitly from sys.argv; that makes the program hard to test. A better way to do this is to have the list of command-line arguments be a function to the main function, and then for the runtime to call main(sys.argv[1:]) instead of just main(). That way you can test the program without having to fake up sys.argv. So I changed that, and while I was making changes I got rid of Claude's comments and the unnecessary exception handling. I pasted the result back at Claude, saying:

Good so far. I have adjusted your source code in several ways:

1. I removed the comments.
2. I removed the exception handling for the subcommand. If one of the subcommands fails, that no longer terminates the main loop
3. Command-line arguments are passed to the main() function rather than being accessed directly

Here is the code. Tell me if you see any problems or if you have any questions about it

I had also sabotaged the code with at least one likely error. Python's subprocess.run returns a “result” structure that includes the subcommand's output and exit status, and I can never remember the name of the component that records the exit status. I either look it up or get it by trial-and-error. This time I didn't bother to either, I just guessed that it was result.exitcode, which I handed off to Claude even though I thought was probably wrong.

Claude's reply said:

Here are a couple of issues I noticed:

There's an error in the error handling code - you're using result.exitcode but in the subprocess.run() function, the attribute is called returncode not exitcode

Awesome, now I don't have to remember that name of that component.

Claude continued:

The parse_arguments() function definition is missing in the pasted code, but it would need to be updated to accept the argv parameter

I had passed sys.argv[1:] to main, and then had main call parse_arguments(argv). But I hadn't changed parse_arguments to expect an argument list. (I hadn't included it in my paste at all, in fact.) Claude was exactly right, and continued with the corrected version:

    #!/usr/bin/env python3

    import sys
    import subprocess
    import argparse

    def parse_arguments(argv):
        …
        return parser.parse_args(argv)

    def main(argv):
        # Parse command-line arguments
        args = parse_arguments(argv)
        …

    if __name__ == "__main__":
        ok = main(sys.argv[1:])
        exit(0 if ok else 1)

Handling NUL-separated values

I said

It's great, thanks for catching those errors. Now let's add a -0 command-line option that tells the program to split the input on "\0" characters instead of at line breaks.

Recall that my handwritten xar did not do this yet; I had left that feature for another day. Claude just added it for me, adding

    parser.add_argument(
        "-0", "--null",
        action="store_true",
        help="Input items are terminated by a null character instead of by newline"
    )

to the argument parser and changing the main loop to:

    # Split input based on the delimiter option
    if args.null:
        # Split on null characters
        items = stdin_data.split('\0')
    else:
        # Split on newlines
        items = stdin_data.splitlines()

    for item in items:
        if not item:
            continue

        …

I was tired of Claude's comments, so I said

Give me the same thing, with no comments

which it did, so I said

From now on leave out the comments. I'm a Python expert and I don't need them. If there is something I don't understand I will ask you.

Claude complied. If I were going to do use Claude again in the future I would include that in the canned instructions that Claude is given up front. Instead I will probably use a tool better-suited to programming, and do whatever the analogous thing is.

Template filling

Now I told Claude to add the % feature:

I changed if not item to if item == "". Now let's make the following change:

• If the command template includes any % characters, each of these should be replaced with the input item.
• Otherwise, if there were no % characters, the input item should be appended to the end of the command as in the current version

Claude did this. It used an explicit loop instead of the list comprehension that I had used (and preferred), but it did do it correctly:

    for arg in command_template:
        if '%' in arg:
            has_placeholder = True
            full_command.append(arg.replace('%', item))
        else:
            full_command.append(arg)

    if not has_placeholder:
        full_command.append(item)

Even without the list comprehension, I would have factored out the common code:

    for arg in command_template:
        if '%' in arg:
            has_placeholder = True
        full_command.append(arg.replace('%', item))

    if not has_placeholder:
        full_command.append(item)

But I am not going to complain, my code is simpler but is doing unnecessary work.

Claude also took my hint to change item == "" even though I didn't explicitly tell it to change that.

At this point the main loop of the main function was 15 lines long, because Claude had stuck all the %-processing inline. So I said:

Good, let's extract the command template processing into a subroutine.

It did this right, understanding correctly what code I was referring to and extracting it into a subroutine called process_command_template. More science fiction: I can say "command template processing" and it guesses what I had in mind!

This cut the main loop to 7 lines. That worked so well I tried it again:

Good, now let's extract the part of main that processes stdin into a subroutine that returns the items array

It pulled the correct code into a function called process_stdin. It did not make the novice mistake of passing the entire args structure to this function. In the caller it had process_stdin(args.null) and inside of process_stdin this parameter was named use_null_delimiter,

YAGNI?

At this point I was satisfied but I thought I might as well ask if it should do something else before we concluded:

Can you think of any features I left out that would be useful enough to warrant inclusion? Remember this program is supposed to be small and easy to use, in contrast to the existing xargs which is very complicated.

Claude had four suggestions:

1. A -p or --parallel option to run commands in parallel

2. A -n or --max-args option to specify the maximum number of items to pass per command

3. A simple -v or --verbose flag to show commands as they're executed

4. A way to replace the command's standard input with the item instead of adding it as an argument

All reasonable suggestions, nothing stupid. (It also supplied code for #3, which I had not asked for and did not want, but as I said before, default Claude is very chatty.)

Parallelization

I didn't want any of these, and I knew that #2–4 would be easy to add if I did want any of them later. But #1 was harder. I've done code like this in the past, where the program has a worker pool and runs a new process whenever the worker pool isn't at capacity. It's not even that hard. In Perl you can play a cute trick and use something like

    $workers{spawn()) = 1 while delete $workers{wait()};

where the workers hash maps process IDs to dummy values. A child exits, wait() awakens and returns the process ID of the completed child, which is then deleted from the map, and the loop starts another worker.

I wanted to see how Claude would do it, and the result was an even bigger win than I had had previously, because Claude wrote this:

   with concurrent.futures.ProcessPoolExecutor(max_workers=args.parallel) as executor:
        futures = [executor.submit(execute_command, cmd, args.verbose) for cmd in commands]

        for future in concurrent.futures.as_completed(futures):
            success = future.result()
            if not success:
                ok = False

What's so great about this? What's great is that I hadn't known about concurrent.futures or ProcessPoolExecutor. And while I might have suspected that something like them existed, I didn't know what they were called. But now I do know about them.

If someone had asked me to write the --parallel option, I would have had to have this conversation with myself:

Python probably has something like this already. But how long will it take me to track it down? And once I do, will the API documentation be any good, or will it be spotty and incorrect? And will there be only one module, or will there be three and I will have to pick the right one? And having picked module F6, will I find out an hour later that F6 is old and unmaintained and that people will tell me “Oh, you should have used A1, it is the new hotness, everyone knows that.”

When I put all that uncertainty on a balance, and weigh it against the known costs of doing it myself, which one wins?

The right choice is: I should do the research, find the good module (A1, not F6), and figure out how to use it.

But one of my biggest weaknesses as a programmer is that I too often make the wrong choice in this situation. I think “oh, I've done this before, it will be quicker to just do it myself”, and then I do and it is.

Let me repeat, it is quicker to do it myself. But that is still the wrong choice.

Maybe the thing I wrote would be sooner or smaller or faster or more technically suitable to the project than the canned module would have been. But it would only have been more be technically suitable today. If it needed a new feature in the future it might have to be changed by someone who had never seen it before, whereas the canned module could well already have the needed feature ready to go, already documented, and perhaps already familiar to whoever had to make the change. My bespoke version would certainly be unfamiliar to every other programmer on the project — including perhaps myself six months later — so would be the wrong thing to use.

I'm really good at hacking this stuff up. Which is a problem. It makes me want to hack stuff up, even when I shouldn't.

Claude tips the balance strongly toward the correct side, which is that I should use the prepackaged module that someone else wrote and not hack something up.

And now I know about concurrent.futures.ProcessPoolExecutor! The world is full of manuals, how can I decide which ones I should read? One way is to read them all, which used to work back when I was younger, but now I have more responsibilities and I don't have time to read the entire Python library reference including all the useless bits I will never use. But here's Claude pointing out to me that this is something I need to know about, now, today, and I should read this one. That is valuable knowledge.

Claude flubs

While testing the program I did

    xar --parallel=2 bash -c '...'

which didn't work because the argparse library tried to parse the -c as a flag to xar, when I wanted it to be a flag to bash. This was my fault, and there is a standard way around it, which is that I should have said

    xar --parallel=2 -- bash -c '...'

which tells argparse to drop the -- and take everything after it literally with no interpretation.

I thought I'd see what Claude had to say about this, and it dropped the ball in a very LLM sort of way.

I asked:

When I did xar --parallel=2 bash -c '...' it misunderstood the arguments. It thought I was trying to pass -c to xar and failed with unrecognized arguments. What should I do about that?

Claude replied:

There are two main approaches to handle this:

1. Use the -- separator to indicate the end of arguments for xar: …
2. Modify your code to handle this case specifically by looking for a command terminator in the argument parser.

Here's how you could modify your code to implement the first approach:

And then it followed with the same, unmodified code…

Which was correct, because -- is already handled by argparse, so no change was actually required to implement the first approach! But Claude's characterization of it as a modification was misleading, wrong, and potentially very confusing.

I said:

I don't see what the difference is in that argument parsing code as compared to what I had before.

and Claude got into the LLM tail-chasing mode where it got more and more confused:

You're right, I apologize for the confusion. There's no difference in the argument parsing code I provided compared to what you had before. The issue is with how argparse handles arguments by default.

To fix the issue with xar --parallel=2 bash -c '...', you need to actually implement support for the -- separator in your code…

   # Find the index of '--' if it exists
   try:
       separator_index = argv.index('--')
       xar_args = argv[:separator_index]
       command_args = argv[separator_index+1:]
   except ValueError: …

Ugh, all wrong in the worst way. The code probably works, but it is completely unnecessary. Claude's claim that “you need to actually implement support for the -- separator” is flat wrong. I pointed this out and Claude got more confused. Oh well, nobody is perfect!

Lessons learned

A long time ago, when syntax-coloring editors were still new, I tried one and didn't like it, then tried again a few years later and discovered that I liked it better than I had before, and not for the reasons that anyone had predicted or that I would have been able to predict. (I wrote an article about the surprising reasons to use the syntax coloring.)

This time also. As usual, an actual experiment produced unexpected results, because the world is complicated and interesting. Some of the results were unsurprising, but some were not anything I would have thought of beforehand.

Claude's code is good enough, but it is not a magic oracle

Getting Claude to write most of the code was a lot faster and easier than writing it myself. This is good! But I was dangerously tempted to just take Claude's code at face value instead of checking it carefully. I quickly got used to flying along at great speed, and it was tough to force myself to slow down and be methodical, looking over everything as carefully as I would if Claude were a real junior programmer. It would be easy for me to lapse into bad habits, especially if I were tired or ill. I will have to be wary.

Fortunately there is already a part of my brain trained to deal with bright kids who lack experience, and I think perhaps that part of my brain will be able to deal effectively with Claude.

I did not notice any mistakes on Claude's part — at least this time.

At one point my testing turned up what appeared to be a bug, but it was not. The testing was still time well-spent.

Claude remembers the manual better than I do

Having Claude remember stuff for me, instead of rummaging the manual, is great. Having Claude stub out an argument parser, instead of copying one from somewhere else, was pure win.

Partway along I was writing a test script and I wanted to use that Bash flag that tells Bash to quit early if any of the subcommands fails. I can never remember what that flag is called. Normally I would have hunted for it in one of my own shell scripts, or groveled over the 378 options in the bash manual. This time I just asked in plain English “What's the bash option that tells the script to abort if a command fails?” Claude told me, and we went back to what we were doing.

Claude can talk about code with me, at least small pieces

Claude easily does simple refactors. At least at this scale, it got them right. I was not expecting this to work as well as it did.

When I told Claude to stop commenting every line, it did. I wonder, if I had told it to use if not expr only for Boolean expressions, would it have complied? Perhaps, at least for a while.

When Claude wrote code I wasn't sure about, I asked it what it was doing and at least once it explained correctly. Claude had written

     parser.add_argument(
        "-p", "--parallel",
        nargs="?",
        const=5,
        type=int,
        default=1,
        help="Run up to N commands in parallel (default: 5)"
    )

Wait, I said, I know what the const=5 is doing, that's so that if you have --parallel with no number it defaults to 5. But what is the --default doing here? I just asked Claude and it told me: that's used if there is no --parallel flag at all.

This was much easier than it would have been for me to pick over the argparse manual to figure out how to do this in the first place.

More thoughts

On a different project, Claude might have done much worse. It might have given wrong explanations, or written wrong code. I think that's okay though. When I work with human programmers, they give wrong explanations and write wrong code all the time. I'm used to it.

I don't know how well it will work for larger systems. Possibly pretty well if I can keep the project sufficiently modular that it doesn't get confused about cross-module interactions. But if the criticism is “that LLM stuff doesn't work unless you keep the code extremely modular” that's not much of a criticism. We all need more encouragement to keep the code modular.

Programmers often write closely-coupled modules knowing that it is bad and it will cause maintenance headaches down the line, knowing that the problems will most likely be someone else's to deal with. But what if writing closely-coupled modules had an immediate cost today, the cost being that the LLM would be less helpful and more likely to mess up today's code? Maybe programmers would be more careful about letting that happen!

Will my programming skill atrophy?

Folks at Recurse Center were discussing this question.

I don't think it will. It will only atrophy if I let it. And I have a pretty good track record of not letting it. The essence of engineering is to pay attention to what I am doing and why, to try to produce a solid product that satisifes complex constraints, to try to spot problems and correct them. I am not going to stop doing this. Perhaps the problems will be different ones than they were before. That is all right.

Starting decades ago I have repeatedly told people

You cannot just paste code with no understanding of what is going on and expect it to work.

That was true then without Claude and it is true now with Claude. Why would I change my mind about this? How could Claude change it?

Will I lose anything from having Claude write that complex parser.add_argument call for me? Perhaps if I had figured it out on my own, on future occasions I would have remembered the const=5 and default=1 specifications and how they interacted. Perhaps.

But I suspect that I have figured it out on my own in the past, more than once, and it didn't stick. I am happy with how it went this time. After I got Claude's explanation, I checked its claimed behavior pretty carefully with a stub program, as if I had been reviewing a colleague's code that I wasn't sure about.

The biggest win Claude gave me was that I didn't know about this ProcessPoolExecutor thing before, and now I do. That is going to make me a better programmer. Now I know something about useful that I didn't know before, and I have a pointer to documentation I know I should study.

My skill at writing ad-hoc process pool managers might atrophy, but if it does, that is good. I have already written too many ad-hoc process pool managers. It was a bad habit, I should have stopped long ago, and this will help me stop.

Conclusion

This works.

Perfectly? No, it's technology, technology never works perfectly. Have you ever used a computer?

Will it introduce new problems? Probably, it's new technology, and new technology always introduces new problems.

But is it better than what we had before? Definitely.

I still see some programmers turning up their noses at this technology as if they were sure it was a silly fad that would burn itself out once people came to their senses and saw what a terrible idea it was.

I think that is not going to happen, and those nose-turning-up people, like the people who pointed out all the drawbacks and unknown-unknowns of automobiles as compared to horse-drawn wagons, are going to look increasingly foolish.

Because it works.

[Other articles in category /tech/gpt] permanent link