[noupdates]

Quite frankly, most seasoned developers should be able to write their own Claude Code. You know your own algorithm for how you deal with lines of code, so it's just a matter of converting your own logic. Becoming dependent on Claude Code is a mistake (edit: I might be too heavy handed with this statement). If your coding agent isn't doing what you want, you need to be able to redesign it.

[nicetryguy]

It's not that simple. Claude Code allows you to use the Anthropic monthly subscription instead of API tokens, which for power users is massively less expensive.

[tibbar]

this is the real reason why people are switching to claude code.

[co_king_3]

Drug dealer business model. The first bag is free. Don't act surprised when you get addicted and they 10x the price.

[bradfa]

Yes and no. There are many not-trivial things you have to solve when using an LLM to help (or fully handle writing) code.

For example, applying diffs to files. Since the LLM uses tokenization for all its text input/output, sometimes the diffs it'll create to modify a file aren't quite right as it may slightly mess up the text which is before/after the change and/or might introduce a slight typo in text which is being removed, which may or may not cleanly apply in the edit. There's a variety of ways to deal with this but most of the agentic coding tools have this mostly solved now (I guess you could just copy their implementation?).

Also, sometimes the models will send you JSON or XML back from tool calls which isn't valid, so your tool will need to handle that.

These fun implementation details don't happen that often in a coding session, but they happen often enough that you'd probably get driven mad trying to use a tool which didn't handle them seamlessly if you're doing real work.

[noupdates]

I'm part of the subset of developers that was not trained in Machine Learning, so I can't actually code up an LLM from scratch (yet). Some of us are already behind with AI. I think not getting involved in the foundational work of building coding agents will only leave more developers left in the dust. We have to know how these things work in and out. I'm only willing to deal with one black box at the moment, and that is the model itself.

[vjerancrnjak]

It's quite tricky as they optimize the agent loop, similar to codex.

It's probably not enough to have answer-prompt -> tool call -> result critic -> apply or refine, there might be a specific thing they're doing when they fine tune the loop to the model, or they might even train the model to improve the existing loop.

You would have to first look at their agent loop and then code it up from scratch.

[chasd00]

I bet you could derive a lot by using a packet sniffer while using CC and just watch the calls go back and forth to the LLM API. In every api request you'll get the full prompt (system prompt aside) and they can't offload all the magic to the server side because tool calls have to be done locally. Also, LLMs can probably de-minimize the minimized Javascript in the CC client so you can inspect the source too.

edit: There's a tool, i haven't used it in forever, i think it was netsaint(?) that let you sniff https in clear text with some kind of proxy. The enabling requirement is sniffing traffic on localhost iirc which would be the case with CC

[sergiotapia]

Claude Code has thousands of human manhours fine tuning a comprehensive harness to maximize effectiveness of the model.

You think a single person can do better? I don't think that's possible. Opencode is better than Claude Code and they also have perhaps even more manhours.

It's a collaboration thing, ever improving.

[noupdates]

Challenge accepted.

[mikert89]

The model is being trained to use claude code. i.e. the agentic patterns are reinforced using reinforcement learning. thats why it works so well. you cannot build this on your own, it will perform far worse

[noupdates]

Are you certain of this? I know they use a lot of grep to find variables in files (recall reading that on HN), load the lines into into context. There's a lot of common sense context management that's going on.