holapabs

6 months ago i started treating claude code like a real coworker, not a chatbot. now it drafts my prs, runs my repros, writes my slack messages. i built a system around it: rules, hooks, slash commands, evals. here's what's inside.

@DhruvJain08 @awilkinson @gregisenberg @danshipper for me the mess wasn't visual, it was state bleed between parallel sessions. a hook stamps every prompt with a session id and logs each tab to a tracking file, so they can't step on each other. self-identifying sessions helped more than any visual layer.

@awilkinson @gregisenberg @danshipper The tab management point is underrated. When you're running multiple agent workflows in parallel, context switching between Claude Code sessions in terminal gets messy fast. A clean visual layer makes the difference between productive and chaotic.

The Codex Mac app rocks. Visually I find it way easier to manage 10-15 tabs than Claude Code Mac or Terminal. The battle rolls on! I was a hardcore Claude Code user and when @gregisenberg and @danshipper pushed me to try it I was skeptical. Impressed. A few pieces of feedback that would make Codex sing (CC: @sama and @fidjissimo): 1. Not having the AskQuestionTool available in work mode (only plan mode) is a travesty! Being able to quickly reply vs get a wall of 15 text based questions that I have to type answers to totally takes me out of my flow. (I updated my settings so that it always switched to plan mode whenever it needs my input, but many users won't do this). 2. I can't explain it, but something about the way it updates on its activity / loads / visually thinks, makes it feel slower. 3. Giving the sub-agents names (like human names) is actually distracting. I would prefer to be able to infer what the agent is/does based on its name (Legal Whiz, NextJS Master, etc). 4. If you could solve preference/environment syncing across multiple Macs, that would be incredible. Current Git-based solutions are very hacky and cause all sorts of errors. If I change my settings on my Mac Studio, I'd love it if it synced to my MacBook. 5. It seems weird that it can't control its own integrated browser and use it to click around sites (unless I'm missing something?) Great work! Super impressed!

@sergiopreira @eng_khairallah1 the drift problem is real. my hacky version: every correction i make gets logged, and repeated ones get promoted into the rules file or a hook. manual versioning of the charter. crude but it keeps CLAUDE.md from rotting. curious how bitloops makes that testable.

The hard part isn't knowing the stack — it's maintaining it. CLAUDE.md files that nobody owns. Context that drifts between sessions. No way to verify your agent actually followed the rules you wrote. Tooling for building agents is ahead of tooling for keeping them correct. Bitloops is working on the second part. github.com/bitloops/bitlo…

Andrej Karpathy: "10x engineers are normal. real agentic engineers are 100x." Context engineering is becoming the foundation of agentic software. Tool routing, memory systems, evals, and orchestrator-subagent workflows are the new stack. Make sure to bookmark it before it gets lost in your feed. This is the article I wish existed 6 months ago.

@bettercallsalva @Jmoon_174 @boringmarketer the feedback loop is the part that compounded most for me. every correction i make gets logged, and the repeated ones get promoted into a written rule or a hook. the tool stopped repeating the same mistakes once that was wired in.

@Jmoon_174 @boringmarketer same here, a few months in. memory across sessions and scheduled jobs are the two that changed the most. the Claude Code ceiling point is right: the base tool is a great default, and the real gains come from the layer you wire around it. hooks, memory, cron.

some of the smartest devs I know aren't using Claude Code, they're customizing their own coding agents. opinionated, built in tools, model flexibility, built on Pi. we're about to see huge growth of open source CODING agents the terminal might be where we see the next hermes

Wrote a PreToolUse hook that rejects any file write containing an em-dash. before it touches disk. 20 lines of python. a hard gate, not a polite request in a system prompt. Prompts are suggestions. hooks are laws.

asked claude to audit its own work setup. instead of agreeing with everything, it told me i had 1657 lines of session log nobody reads, 24 stale tasks, and no evals. the value isn't that the AI is smart. it's that the system is designed to push back.

4/ none of this is theoretical. it's what i use every day, in production, at a real job. i'm cleaning up the public version of the repo. dropping it soon. if you're building something similar, follow + drop a reply. happy to compare notes.

3/ every time i correct claude, i run /violated <description>. it logs the correction with timestamp + canonical case. after 30 days i review the log. if a correction repeats, it gets promoted to a written rule. or a hook if it's mechanically detectable. compounding.

6 months ago i started treating claude code like a real coworker, not a chatbot. now it drafts my prs, runs my repros, writes my slack messages. i built a system around it: rules, hooks, slash commands, evals. here's what's inside.