Quick background if you haven't used them: hooks in Claude Code are your own scripts that run automatically at set points, like right before a command runs or right when the agent finishes its answer. A hook can block that step or send the answer back to be rewritten. CLAUDE.md is something the model reads and can forget after a long session. A hook just runs, every time.
Here are the ones I'd miss most. Each started as a rule in CLAUDE.md that the agent kept breaking.
Replies drift into English after long sessions. A hook checks every answer and sends it back if it isn't in Japanese.
I had my agent hallucinate the wrong day of the week three times in a row simply because I let it calculate timestamps on its own instead of using a basic system script. Moving all structured data generation like JSON fields file paths and dates away from the model and into deterministic Python scripts completely fixed our pipeline stability.
LLMs are great for writing plain text and copy but terrible at programmatic math. Do you still let your agents handle raw config values or do you lock those down with scripts?
It doesn't even have to be a big dramatic failures, more the small moments where something clicked and you went from trusting the output by default to double checking everything. For me it was watching an agent confidently rename a function across twelve files, then leave the original function untouched in a thirteenth file it apparently didn't search, with zero indication anything had been missed. It wasn't even a hard case, the file just wasn't in the directory it happened to grep first. What was your moment? And did it actually change your workflow afterward, or did the trust creep back in after a week like it always seems to for me?
Daily reality: Claude Code in one pane, Codex in another, Cursor in a third, plus a few shells. An agent stops and waits for "yes". By the time I find the right pane, 20 min gone.
Just me? What's your setup notifications, layout tricks, or you've accepted the tax?
I've been experimenting with AI coding assistants recently, and something interesting happened.
I asked an AI agent to make a relatively small change to an existing project. Instead of modifying the necessary section, it generated a completely new implementation with hundreds of lines of code.
It technically worked, but maintaining it would have been another challenge.
I tried spinning up multiple AI coding tabs last weekend to knock out a bunch of bug fixes in one go thinking it would make me four times as productive. Instead I spent two hours staring at different git diffs completely losing track of what logic belonged where.
The agents spat out code way faster than I could read, let alone test or understand. Code generation speed doesn't mean much if your own review bottleneck slows everything down to a crawl.
How many parallel tasks do you actually let your tools run before you start missing bugs?
I might be missing some but I've been pretty much in love with @Lovable, @Cursor, @bolt.new and have been trying to use @Replit more and I honestly haven't touched @BASE44 too much but have heard good things. @chrismessina has nudged me to use @Windsurf for whenever I build another Raycast Extension! Currently I use: - @bolt.new / @Lovable - @Cursor - @Warp Curious what everyone thinks is the top one so far!
I logged the raw requests from one Claude Code session on a real repo to see where the input tokens actually go. I expected the files the agent read to dominate.
Input was about 123k tokens:
~74% tool definitions
~21% conversation, including a 40 KB file the agent read
Two of mine: "Never edit requirements.txt by hand, it's generated with hashes." Added after an unpinned dependency jumped to a new version on deploy and caused an incident in production. "Any API change updates the docs, the MCP tool descriptions and the slash commands in the same step." Added after an audit against production found the three had drifted apart. Curious what's in yours, and what happened before you wrote it.
