“Claude Code builds its prompts on the fly. So, there is a sparse main prompt (I spent an hour looking for it, its there somewhere clone the repo and Fable can tell you a ton, but Fable had only dl the repo to a scratchpad and had erased it by the time I thought to ask where the main system prompt is in the repo.”
“I noticed this because I customized Claude Code to just read all source at startup (if the project was only a few thousand lines of code). And it would still read the stuff it had already read! Because the system prompt explicitly told it to...”
“In the end I had a serious talk with Claude Code and we agreed on him making the prompts he wants exactly and I used ChatGPT for the artwork so Claude could cut it out.”
“I used the VSCode extension a lot for months but now use the CLI almost exclusively. Ultimately it feels cleaner and is, by nature, closer to the Claude Code harness.”
“You can also just delegate this to subagents with Claude Code (though you have a more limited choice of models unless you swap the cheaper models via OpenRouter).”
“Question on the multi-client setup (Cursor + Claude Code simultaneously), are you assuming users run Palinode as a background HTTP daemon that both clients connect to over streamable HTTP? Or if both spawn it via stdio, how do you handle concurrency across two processes touching the same SQLite index and git repo without lock contention?”
“Claude Code ships LSP support but only if packaged within a plugin. So there are hoops you have to jump through to get it rather than getting it out of the box. But it might be long term more reliable than ad hoc fixes?”
“For example, Claude Code likely collected a substantial number of real-world coding trajectories through its CLI. However, trajectories involving tools such as LSP, MCP, or AST-grep were probably scarce in the dataset.This lack of representation may also indirectly limit the effectiveness of subsequently generated synthetic data.”
“There’s been an interesting co-evolution that I’ve been experiencing with Claude Code. I’ll ask it to do a task, I’ll watch what it’s doing (often lots of find and grep and ripgrep) and then after the task is complete I’ll ask it if there are any tools that would’ve made the job easier.”
“Claude has the exact opposite problem, especially opus-5, where I literally can’t trust it to print hello world without taking a shortcut, or just simply lying and saying it printed it when it didn’t, behind a giant wall of inscrutable text.”
“Hi HN. We built an API context registry to help coding agents (like Claude Code) generate production-ready API integration code without blowing through token limits.We build a lot of API integrations.”
“I have less than 10 skills, maybe 2-3 hooks, only 4-5 sub-agents, tight output style, a concise claude.md and agents.md since I run claude code and codex in parallel in vs code, no MCPs, down to under 25k context on new chats, a parity-audit skill to keep the models from drifting apart, don't use anyone elses skills/hooks/plugins, etc.”
“The honest bit: hooks cost about 2 seconds per tool call, and a 14-agent audit of the harness once had its own verification fail silently (matched 0 of 154 verdicts and reported zero refuted). That's in the post too.”
“If I use the same memory repo in Claude Code and Cursor on the same day, and both write, do I end up resolving git conflicts by hand? Or does the executor queue/serialize somehow?”
“The thing that finally worked for me: keep the "what we tried, why we rejected it, what's next" on GitHub Issues themselves, as a map of decision tickets the agent resolves one at a time, so any fresh session (Claude Code or Codex) rehydrates from the Issue, not from chat history.”
“You pick somewhere, browse for a bit, and the moment Claude finishes or needs something from you, it pings you and brings you back to the terminal.Instead of me opening X on my own and checking to see if claude is done 50 times and adhdmaxxing, brb will facilitate this in a much smoother way.”
“Like if I'm working in Claude Code and Cursor at same time and both agents try to save something related. Does it just fail the second commit and make the agent retry?”
“whats so bad about it when the guy in question is bolting gui elements ontop of a fucking tui under the assumption that claude app and claude code cli use a materially different harness core”
“The parts I found it genuinely good at were the investigations, not just the typing: it measured 8,396 output gaps across eight live sessions to find the idle threshold for the "waiting on you" detection, tracked a crash on every pane kill down to node-pty's ConPTY console-list helper and fixed it with a config flag, and worked out why touch scrolling died on text (xterm rebuilds the span under your finger) and fixed it with pointer capture.”
“Claude code can you create a local web site that can train me to use agents, I'd like there to be about 1 hours content tests some examples and eli5 it.... I'll need a batch file to start the environment and share the URL when done too please. Ta.”
“We saw Retool pivot (probably not the right word) into a vibe-coding based approach along with a few lesser known competitors.We didn't and still don't think vibe-coding is right for our users.”
“Claude code is easy. Use the CLI. Talk to it about what you are doing. You will hit limits faster early on, but will eventually get the hang of what you need from it”
“claude code’s learning curve is basically a vertical line at first, then it flattens out real quick once you get the loop. i wouldn’t drop money on a course. just set up a dummy project, something stupid, like a to-do app or a weather fetcher, and force yourself to do everything through the agent.”
“everyone on my team is shipping with claude code and i still open files by hand like a caveman. is there a real course for this or do i just watch people on youtube until it clicks.”
“Claude code (or any other agent) would then generate the guided tour, and give you a command to run in your terminal. That command opens a TUI where you can review the code - stepping through organised chapters, checking things off as you review them, leaving comments on lines (github style gutter drag).”