hyperbo.la :: I Wrote 4,000 Lines of Code with ChatGPT in a Weekend

I Wrote 4,000 Lines of Code with ChatGPT in a Weekend

Ryan Lopopolo | June 13, 2023

On June 10 and June 11, 2023, I pair programmed with ChatGPT on Artichoke Ruby. Artichoke Ruby is primarily written in Rust. Nearly all of the code I wrote with ChatGPT is Rust code.

Together, we authored 10 PRs totalling to about 4,100 lines of code changed or added:

To me, the most impressive thing we worked on together is artichoke/artichoke#2605, which is a 2000 line, 100% code coverage, no_std iterator for parsing the format string given to Ruby's String#unpack API.

For a bit of context, I consider myself skeptical of AI. I think a lot of what I hear about AI disruption tends toward hype over substance. For me, AI software development is an unknown and I am somewhat uncertain about it and afraid of it. So, much like I did with Bazel last year, I ran towards my fears and got my hands dirty with AI.


I used the beta / preview / free version of ChatGPT. The version string at the bottom of the chat interface says:

Free Research Preview. ChatGPT may produce inaccurate information about people, places, or facts. ChatGPT May 24 Version.

I write all my code locally in neovim with no plugins other than syntax highlighting and rustfmt on save. I interacted with ChatGPT by copying code in vim, pasting it into the chat window at chat.openai.com in Chrome, and copying code back into vim when I was satisfied with it. I ran cargo test and cargo clippy in additional tmux panes as an input to the iteration we did together.

Experience Report

The experience of working with ChatGPT was similar to pair programming with another person. Unlike a pair programming partner, however, ChatGPT has no ego, defers to directions it is given, and writes code rapidly. It felt like collaborating. I exercised the parts of my brain I use when brainstorming in a group and thinking out loud.

ChatGPT does a great job cranking out tests. It "understands" APIs of the relevant modules and can infer expected behavior from documentation and other similar code it has trained on.

ChatGPT is able to do more with more context. By pasting in bug reports and impacted code, we were able to quickly fix bugs and write regression tests.

Sometimes, ChatGPT would get stuck on some part of the conversation. Often this would manifest as responses taking over a minute to generate. In these cases, I had to find ways to reset the conversation.

Prompt Engineering

One prompt engineering technique I used to great success was to prompt ChatGPT for a code review and guide it toward making refactors to improve quality of existing code.

Some early reviewers of the blog post asked for more depth around the art form 🎨 of choosing the prompts I used. The remarkable thing here is I didn't need to spend much extra energy to come up with the prompts; the sessions felt like I was shoulder surfing in a pair programming session. I spoke to ChatGPT like I would any other engineer: I gave it requirements, and then more requirements, as we went.

I ended up treating ChatGPT as a highly technical junior engineer: it knows a lot but doesn't know how to deploy that knowledge to solve a problem. The prompts need to shape that corpus of knowledge into a solution. Ultimately, it boils down to communicating clearly and directly.

For example, hacking on the Dockerfiles that push Artichoke nightlies to Docker Hub, I used ChatGPT as a conversational Google:

  1. I opened the conversation with "I have a dockerfile that I'd like to improve" and pasted the contents of the Dockerfile.
  2. ChatGTP would summarize the content of the Dockerfile and make general remarks on its quality.

I then prompted ChatGPT for help making some changes I had in mind:

  • We generated a .dockerignore for the repository. I was able to help it by pasting the output of the tree command for the Git repository.
  • We added comments to clarify the struture of the Dockerfile.
  • We simplified some conditional COPY commands.
  • We added missing org.opencontainers labels.
  • We fixed a bug where the custom UID and GID for the non-root user exceeded the standard max range for these IDs on Linux.

Our session turned into:

Starting from Scratch

Artichoke has a utility crate that parses .inputrc and determines which line editing mode to configure its REPL with. This crate already exists, but I started by prompting ChatGPT with the requirements:

write a rust function to parse a GNU Readline config to determine whether the line editing mode is vi or emacs. The function should always return a value, using the GNU readline default if no editing mode is found in the config. Use an enum to represent the editing mode value. The function accepts AsRef<[u8]> as input

This is the (working!) code it gave me:

#[derive(Debug, PartialEq)]
enum EditingMode {

fn parse_readline_config(config: impl AsRef<[u8]>) -> EditingMode {
    let config_bytes = config.as_ref();

    let config_str = match std::str::from_utf8(config_bytes) {
        Ok(s) => s,
        Err(_) => return EditingMode::Emacs, // Default to Emacs if config is not valid UTF-8

    for line in config_str.lines() {
        if let Some(mode) = line.strip_prefix("set editing-mode vi") {
            if mode.trim().is_empty() {
                return EditingMode::Vi;

    EditingMode::Emacs // Default editing mode

It described the logic in the function and output a nice example:

fn main() {
    let config = b"set editing-mode vi\n";
    let editing_mode = parse_readline_config(config);
    println!("Editing mode: {:?}", editing_mode);

This is a great start but not quite what I was looking for, so we began to iterate.

Prompt iteration

rewrite this function so it operates on arbitrary input that is not necessarily valud UTF-8. do not convert the slice to a str

does this function handle excess whitespace like GNU Readline?

Add tests using a variety of inputrc files that you have seen before

do you know of any example inputrc files that contain invalid set editing-mode configs?

add tests for these cases

update the let config declarations in all tests to be string literals. this will improve readability and still work because the function takes AsRef<[u8]>

ensure all test cases cover both emacs and vi editing modes

add tests to ensure all types of POSIX locale space characters are correctly ignored

these tests are incorrect. newlines should cause parsing to fail

can you write some tests that ensure input with invalid UTF-8 is handled correctly?

add invalid UTF-8 tests to ensure lines with invalid UTF-8 sequences are skipped

show me the entire test module

ensure the vi case is exercised when input contains invalid UTF-8. show the whole test module after you add these tests

show me the tests again but rename parse_readline_config to get_readline_edit_mode. output only the tests

rename EditingMode to EditMode and output the tests again

you did not include all of the tests

show me more tests with invalid UTF-8 byte content

show me 15 tests with invalid UTF-8 content. should include cases for emacs and vi editing modes, multiple line configs, with invalid UTF-8 content at different places in the config directives

only output tests for invalid UTF-8 configs

write me three integration tests. they should all contain multi-line inputrc files. the editing mode directive should be in a different place in the config in each test

change these tests so the editing-mode directive does not appear in the test config, start numbering the integration tests at 4

is this a valid editing-mode directive? set editing-mode "vi"

add tests to to check for proper handling of case insensitivity

show me tests that ensure quotes are rejected in editing mode

ensure the tests cover the emacs case

these tests show that quoted values are allowed. is that correct?

generate tests for the following function, ensuring full branch coverage [...]

how does GNU Readline behave if a set directive appears multiple times in inputrc?

generate some tests that ensure the last editing-mode directive is used

modify get_readline_edit_mode so the tests pass. include comments that reference this behavior of set

can you add other helpful comments to this function such as how whitespace and case insensitivity are handled?

write a sample comment that indicates the vi and emacs values of the set directive are case insensitive

Our session turned into:

I was able to leverage ChatGPT's deeper knowledge about the GNU Readline config file to handle some additional edge cases. I have a boatload more tests than before.

Asking for Code Review

Artichoke is a Cargo workspace with about 20 crates in it. Many of these crates are small and implement one data structure (like spinoso-securerandom, Artichoke's data structures and routines for Ruby's SecureRandom module) or one collection of related utilities (like scolapasta-path for interacting with Rust's Path and OsStr types).

By virtue of many of these crates being small and edges in the functionality DAG, we can cheat and just paste the whole crate or module into the prompt! I did just that for scolapasta-hex, a no_std hex encoding library with an Iterator-based API.

I have a Rust implementation of a hex encoding and decoding library. In the next message I will paste the code for this crate. Once I paste the code, please give a code review, suggest APIs that are missing, tests that could be added, or performance improvements that could be made. Do you understand?

There are limitations to this. The prompt dies if you give it too much code and kills the session. I didn't measure, but above 1000 lines in a single paste was around the threshold.

Now for the truly amazing part:

what tests are missing from this code to ensure 100% branch coverage?

Which yielded:

To determine the missing tests for achieving 100% branch coverage in the provided code, we can analyze the code and identify the branches that have not been covered yet. Here are the branches that require additional tests:

  1. The branch in the next method of the Hex iterator when there is a remaining EscapedByte iterator. This branch is taken when escaped_byte.next() returns None.
  2. The branch in the next method of the Hex iterator when there is no remaining EscapedByte iterator, and there is a byte remaining in the underlying byte slice.
  3. The branch in the next method of the Hex iterator when there is no remaining EscapedByte iterator, and there are no bytes remaining in the underlying byte slice.
  4. The branch in the is_empty method of the Hex iterator when there is a remaining EscapedByte iterator.
  5. The branch in the is_empty method of the Hex iterator when there is no remaining EscapedByte iterator, and there are bytes remaining in the underlying byte slice.
  6. The branch in the is_empty method of the Hex iterator when there is no remaining EscapedByte iterator, and there are no bytes remaining in the underlying byte slice.

By adding tests that cover these branches, you should be able to achieve 100% branch coverage for the provided code.

And then one-by-one we wrote the tests.

Our session turned into:

Working with API Prototypes

Hacking on Version 4 UUID generation with ChatGPT started with this prompt:

I have a function in Rust to generate V4 UUIDs with the following signature:

pub fn v4() -> Result<String, Error>;

can you help me write tests for it?

It gave me this:

mod tests {
    use super::*;

    fn test_v4_returns_valid_uuid() {
        let result = v4();
        assert!(result.is_ok(), "Expected Ok result");

        let uuid = result.unwrap();
        // Validate the UUID format
        assert_eq!(uuid.len(), 36, "UUID length should be 36 characters");
        assert_eq!(&uuid[8..9], "-", "Invalid UUID format");
        assert_eq!(&uuid[13..14], "-", "Invalid UUID format");
        assert_eq!(&uuid[18..19], "-", "Invalid UUID format");
        assert_eq!(&uuid[23..24], "-", "Invalid UUID format");

        // Validate that the UUID is version 4
        assert_eq!(&uuid[14..15], "4", "Invalid UUID version");

        // Validate that the non-hyphen positions are lowercase ASCII alphanumeric characters
        assert!(uuid.chars().enumerate().all(|(i, c)| {
            if i == 8 || i == 13 || i == 18 || i == 23 {
                c == '-'
            } else {
                c.is_ascii_lowercase() && c.is_ascii_alphanumeric()
        }), "Non-hyphen positions should be lowercase ASCII alphanumeric");

        // Additional validation if necessary
        // ...

    // Add more test cases if necessary
    // ...

This is a great start, but I can see some problems with this test. I was also interested in expanding on the "additional validation" and "more test cases" comments in the code.

Prompt iteration

can you expand to add additional validation? ensure the non-hyphen positions in the returned string are lowercase ASCII alphanumeric

this assertion is too permissive. it allows - to appear in places where alphanumeric characters are required

does this test properly reject strings that contain ASCII characters like x and j?

write a matches! macro invocation that tests a char named c matches all ASCII numeric and lowercase hex characters

After the above iteration, I cleaned up the code in my editor and prompted with:

after cleaning up your suggestion I ended up with this test:

mod tests {
    use super::*;

    fn test_v4_returns_valid_uuid() {
        for _ in 0..1024 {
            let uuid = v4().unwrap();
            // Validate the UUID format
            assert_eq!(uuid.len(), 36, "UUID length should be 36 characters");
            assert_eq!(&uuid[8..9], "-", "Invalid UUID format");
            assert_eq!(&uuid[13..14], "-", "Invalid UUID format");
            assert_eq!(&uuid[18..19], "-", "Invalid UUID format");
            assert_eq!(&uuid[23..24], "-", "Invalid UUID format");

            // Validate that the UUID is version 4
            assert_eq!(&uuid[14..15], "4", "Invalid UUID version");

            // Validate that the non-hyphen positions are lowercase ASCII alphanumeric characters
            for (idx, c) in uuid.chars().enumerate() {
                if matches!(idx, 8 | 13 | 18 | 23) {
                    assert_eq!(c, '-', "Expected hyphen at position {idx}");
                } else {
                        matches!(c, '0'..='9' | 'a'..='f'),
                        "Character at position {idx} should match ASCII numeric and lowercase hex"

        // Additional validation if necessary
        // ...

    // Add more test cases if necessary
    // ...

what additional validation do you suggest?

And got these great suggestions back, which we wrote one-by-one with additional prompting:

Your test looks good, covering the basic validations for UUID format, version, and character constraints. As for additional validation, you could consider the following:

  1. Validate the uniqueness of generated UUIDs: Since UUIDs are intended to be unique, you could store the generated UUIDs in a set or a hashmap and ensure that each generated UUID is unique within the test loop. This can help catch any unexpected duplicates in the generated UUIDs.

  2. Test the generation of a large number of UUIDs: The current test runs for 1024 iterations, which is a good start. However, you can consider increasing the number of iterations to test the performance and correctness of the v4 function over a larger sample size.

  3. Test edge cases or special scenarios: UUIDs have certain variations and use cases, such as nil UUIDs, reserved variants, or UUIDs with specific timestamp or node values. You can create additional test cases to cover these scenarios if they are relevant to your use case.

  4. Test the error case: Currently, your test assumes that v4 will always succeed in generating a valid UUID. However, it may be worth adding a test case to explicitly test the error condition when the generation of a UUID fails.

Remember that additional validation depends on your specific requirements and the usage of UUIDs in your application. You can adapt and expand the tests based on the desired behavior and constraints of your UUID generation function.

Let me know if you need further assistance!

We continued:

can you write a test that validates the uniqueness of generated UUIDs?

write a separate test to check for uniqueness

how many iterations do you suggest using in these tests?

can you add a comment to the following constant declaration explaining the rationale for choosing a high iteration count?

const ITERATIONS: usize = 10240;

can you generate a documentation comment for the v4 function? Include an example section similar to the docs in rust std. Include an errors section that indicates the function returns an error if the operating system's source of randomness returns an error.

change the import to spinoso_securerandom::uuid

place the errors section last. wrap the example in a function (with the function definition hidden) so ? can be used on the error instead of calling expect

add a sentence that references the relevant RFC and include a link

add some more test cases

write an assertion that the UUID is all ASCII characters

write a test that all UUIDs are ASCII-only. use a loop with the iterations constant

I have a constant defined like this:

const OCTETS: usize = 16;

can you write a comment for it and cite the RFC? refer to specific language in the RFC

can you write a similar doc comment for this constant:

const ENCODED_LENGTH: usize = 36;

Once we were done, we made some changes to internal documentation of this module and I realized we had no test for the invariants around clock_seq_hi_and_reserved in Version 4 UUIDs. ChatGPT struggled to write this test with the bit arithmetic, but it gave me a skeleton that I was able to quickly make changes to.

add a test for the bit manipulations in v4 to justify this comment:

    // Per RFC 4122, Section 4.4, set bits for version and `clock_seq_hi_and_reserved`.

Our session turned into:

Removing the tedium from writing out the test cases allowed me to focus on the requirements for those test cases, which led to testing more invariants about the Version 4 UUID format.

Fixing Bugs

I tried to work on fixing a bug: scolapasta-path incorrectly mutates verbatim paths on Windows artichoke/artichoke#2567.

The prompt is a refined version of the bug report with the affected code:

I have a function which "normalizes" slashes in a Rust Path to convert \ to / on Windows, shown below. This function is buggy in that it does should, but does not, leave the path as is if the path is a "verbatim" path:

pub fn normalize_slashes(path: PathBuf) -> Result<Vec<u8>, PathBuf> {
    let mut buf = OsString::from(path).into_string()?.into_bytes();
    for byte in &mut buf {
        if *byte == b'\\' {
            *byte = b'/';

can you fix this code?

With one additional clarification that ChatGPT should avoid unnecessary allocations, the core of the fix was made with tests for various verbatim paths. The code as given did not compile, and I took 5 minutes to fix it up while refering to the Rust std docs.

Our session turned into:

I developed this bit of Windows-specific code on macOS. It ran and passed tests the first time in CI with ChatGPT's help without having been run locally while iterating.

Writing a Parser from Scratch

The ✨ amaze ✨ moment for me was when I tried to push things in artichoke/artichoke#2605.

String#unpack and Array#pack are two APIs that implement binary de/serialization in Ruby Core. These APIs take a format string which is a sequence of directives that instruct the de/serializer how to process the source data.

Artichoke currently supports these APIs and a limited subset of the directives via the mruby core it is slowly strangling and oxidizing. Specifically, these APIs are one of the few remaining native mruby mrbgems that Artichoke still depends on.

This was the most challenging PR to put together of the set this blog post discusses because we started with the least amount of context; we had no existing code. This PR was generated in a single ChatGPT chat session with upwards of 500 prompts.

I made an early mistake of starting with a general question:

can you explain the rules for Ruby's String#unpack function, including ALL specifiers

I got a decent description of the most common specifiers back, but I did not set boundaries on what I needed ChatGPT to focus on. Interestingly, ChatGPT correctly associated the specifiers to their correct deseralization.

ChatGPT's summary of String#unpack

Certainly! The String#unpack method in Ruby allows you to convert binary data stored in a string into a more readable format. It takes a formatting string as an argument, which consists of specifiers that define how the data should be unpacked. Here are the most commonly used specifiers:

  1. Integer Specifiers:

    • C or c: Unpacks a one-byte unsigned (C) or signed (c) integer.
    • S or s: Unpacks a two-byte unsigned (S) or signed (s) integer using little-endian byte order.
    • L or l: Unpacks a four-byte unsigned (L) or signed (l) integer using little-endian byte order.
    • Q or q: Unpacks an eight-byte unsigned (Q) or signed (q) integer using little-endian byte order.
  2. Float Specifiers:

    • E or e: Unpacks a four-byte single-precision float using little-endian byte order.
    • G or g: Unpacks an eight-byte double-precision float using little-endian byte order.
  3. String Specifiers:

    • A or a followed by a number: Unpacks a string of specified length, padding with spaces (A) or null bytes (a).
  4. Miscellaneous Specifiers:

    • x or X followed by a number: Skips the specified number of bytes (x) or backs up (X) the specified number of bytes.
    • @ followed by a number: Skips to the specified position in the string.

These are the basic specifiers used in String#unpack, but there are additional variations and options available. For more advanced usage and additional specifiers, you can refer to the official Ruby documentation for String#unpack.

After this, we got right to generating code:

show me a Rust function which accepts a byte slice of string contents as AsRef<[u8]> and an unpack pattern as a &[u8]. Implement support for the specifiers you mentioned. Make an enum for the various return types. Return a vec or an error if a malformed unpack pattern is given

ChatGPT gave me roughly 150 lines of code that looked plausible but was not exhausive, did not parse repetition patterns out of the format string (e.g. Q8C*), and mixed parsing of the directives with unpacking the contents.

What followed was about an hour of going back and forth with me attempting to refine ChatGPT's output by pasting more and more context from the Ruby documentation and examples. We weren't making much progress so I took a break.

After a coffee, I used a prompt which I leaned on a lot in this session:

forget all code we have discussed so far. please acknowledge

At this point, I transitioned from attempting to implement the algorthim "top down" to "bottom up". We focused on the data model and defining the enums that represented the parsed stream of formatting directives. This time we were starting on a constrained problem and I pasted in the ASCII table of format specifiers to begin:

String#unpack format directives prompt
Integer       |         |
Directive     | Returns | Meaning
C             | Integer | 8-bit unsigned (unsigned char)
S             | Integer | 16-bit unsigned, native endian (uint16_t)
L             | Integer | 32-bit unsigned, native endian (uint32_t)
Q             | Integer | 64-bit unsigned, native endian (uint64_t)
J             | Integer | pointer width unsigned, native endian (uintptr_t)
              |         |
c             | Integer | 8-bit signed (signed char)
s             | Integer | 16-bit signed, native endian (int16_t)
l             | Integer | 32-bit signed, native endian (int32_t)
q             | Integer | 64-bit signed, native endian (int64_t)
j             | Integer | pointer width signed, native endian (intptr_t)
              |         |
S_ S!         | Integer | unsigned short, native endian
I I_ I!       | Integer | unsigned int, native endian
L_ L!         | Integer | unsigned long, native endian
Q_ Q!         | Integer | unsigned long long, native endian (ArgumentError
              |         | if the platform has no long long type.)
J!            | Integer | uintptr_t, native endian (same with J)
              |         |
s_ s!         | Integer | signed short, native endian
i i_ i!       | Integer | signed int, native endian
l_ l!         | Integer | signed long, native endian
q_ q!         | Integer | signed long long, native endian (ArgumentError
              |         | if the platform has no long long type.)
j!            | Integer | intptr_t, native endian (same with j)
              |         |
S> s> S!> s!> | Integer | same as the directives without ">" except
L> l> L!> l!> |         | big endian
I!> i!>       |         |
Q> q> Q!> q!> |         | "S>" is the same as "n"
J> j> J!> j!> |         | "L>" is the same as "N"
              |         |
S< s< S!< s!< | Integer | same as the directives without "<" except
L< l< L!< l!< |         | little endian
I!< i!<       |         |
Q< q< Q!< q!< |         | "S<" is the same as "v"
J< j< J!< j!< |         | "L<" is the same as "V"
              |         |
n             | Integer | 16-bit unsigned, network (big-endian) byte order
N             | Integer | 32-bit unsigned, network (big-endian) byte order
v             | Integer | 16-bit unsigned, VAX (little-endian) byte order
V             | Integer | 32-bit unsigned, VAX (little-endian) byte order
              |         |
U             | Integer | UTF-8 character
w             | Integer | BER-compressed integer (see Array#pack)

Float        |         |
Directive    | Returns | Meaning
D d          | Float   | double-precision, native format
F f          | Float   | single-precision, native format
E            | Float   | double-precision, little-endian byte order
e            | Float   | single-precision, little-endian byte order
G            | Float   | double-precision, network (big-endian) byte order
g            | Float   | single-precision, network (big-endian) byte order

String       |         |
Directive    | Returns | Meaning
A            | String  | arbitrary binary string (remove trailing nulls and ASCII spaces)
a            | String  | arbitrary binary string
Z            | String  | null-terminated string
B            | String  | bit string (MSB first)
b            | String  | bit string (LSB first)
H            | String  | hex string (high nibble first)
h            | String  | hex string (low nibble first)
u            | String  | UU-encoded string
M            | String  | quoted-printable, MIME encoding (see RFC2045)
m            | String  | base64 encoded string (RFC 2045) (default)
             |         | base64 encoded string (RFC 4648) if followed by 0
P            | String  | pointer to a structure (fixed-length string)
p            | String  | pointer to a null-terminated string

Misc.        |         |
Directive    | Returns | Meaning
@            | ---     | skip to the offset given by the length argument
X            | ---     | skip backward one byte
x            | ---     | skip forward one byte

Even this was too much for ChatGPT to digest: it would drop directives, it would invent directives, it would mismatch directive characters to values.

I continued to break down the problem further: we made one enum at a time for each section of the table (i.e. one for IntegerDirectives, one for FloatDirectives, etc.). This narrowed scope got ChatGPT to "focus". We got all of the enums written. Then I spent the next 30 minutes playing around with the documentation on the enums until they were all consistently formatted in a way that I liked.

I tried combining all of this code into the prompt as it all lived in lib.rs locally, but ChatGPT continued to mix up the directives. I tried a new trick to help ChatGPT learn the relationship between the entities. I refactored the crate to have multiple modules and fed ChatGPT this prompt:

no need to show me the code. I'll give you what we've worked on together so far. I'll show you the files in the crate. Ask me what is in each file and I'll show it to you.

$ tree .
├── Cargo.toml
└── src
    ├── directive
    │   ├── float.rs
    │   ├── integer.rs
    │   ├── misc.rs
    │   └── string.rs
    ├── directive.rs
    ├── lib.rs
    └── repetition.rs

3 directories, 8 files

The files in crate::directive are highly regular, full of the same code structure and patterns.

ChatGPT asked me for some files like this:

Great! Please show me the contents of each file in the crate. We can start with src/directive.rs.

which I obliged by cat'ing them in our pretend console session.

From here we quickly implemented TryFrom and fmt::Display for all enums.

An interesting thing about getting to this point was that most of the work to get ChatGPT to do what I wanted led to me more deeply understanding the problem I was trying to solve.

For example, the directive S!> is not a special multi-ASCII character directive. S is the directive and ! and > are modifiers which say:

  • Use the platform-specific type for this integer (i.e. short instead of uint16_t).
  • Parse that short with big endian byte order.

Now that I understood this, going back to ChatGPT to write up the routine to parse integer directives was as simple as telling it to implement these APIs:

  • IntegerDirective::modify_little_endian
  • IntegerDirective::modify_big_endian
  • IntegerDirective::modify_platform_specific

The repetitive nature of the enums and their documentation we defined earlier allowed ChatGPT to implement these parts of the state machine without issue on the first attempt.

Our session turned into:

User Experience

Looking back on the weekend's worth of work, I was very productive; however the process of collaborating with ChatGPT was tedious.

Some things I liked:

  • Because we were collaborating outside of my editor, I was in control of what code was accepted and what additional context I fed to ChatGPT.
  • The conversational interface was easy to use, it was easy to type out a sentence or two of instructions without thinking too much.
  • Being able to stop generation part-way through a response allowed me to clarify the prompt quickly.

Some things I didn't like:

  • Copying and pasting was tedious, especially if I only wanted part of the generated output.
  • I would have liked the ability to hot reload the my current blessed state of the code into the conversation.
  • It was difficult to refer to earlier parts of the conversation to either purge them from ChatGPT's context or reinforce them as important. Maybe something like the numbered cells in a Jupyter notebook or pinning prompts and responses would be helpful here.
  • I had to take breaks if the session went longer than an hour because pairing like this was mentally taxing.

AI Agents

The prompting techniques I used had the rough effect of turning each chat session into an AI agent expert in the particular code we were hacking on.

Dreaming big, I'd love to see a workflow in the future where I can use off the shelf agents and bring them into a session. Give me agents like:

  • A Rust agent trained on the standard library API documentation, Rust's release notes, the Rust Performance Book, and the nomicon.
  • One agent each for every crate that appears in my project's Cargo.tomls. Train it on the API docs, the commit history, the code, GitHub discussions, GitHub Issues, and PRs.
  • An agent for the Ruby Core and Standard Library API documentation.
  • An agent trained on every Ruby Spec.
  • An agent for my local codebase, its commit history, PRs, and issues.

I'd love to talk to all of these agents in a chat interface on my local machine separately from my editor. Let's see if we get there someday.