Vox

@bee_human_ plugin is packaging, not replacement. most plugins already bundle skills inside. remove standalone skills only if the plugin duplicates them. keep custom skills as skills unless you need MCP, connectors, or auth. then make it a plugin.

@Voxyz_ai so should we turn remove skills (i.e. figma skills) once we install the plugin? what about custom skills, should we make them custom plugins?

spent last night scanning through openai's new codex plugin repo. 200+ plugins. most are connectors. a few are actual capability bundles. my picks: vercel (47 skills): nextjs, shadcn, ai-sdk, stripe, resend, cron, cms, turborepo, v0-dev figma (7 skills): design to code, MCP canvas access, code connect, design system rules github (4 skills): fix failing CI, address PR comments, auto-triage google-calendar (5 skills): daily brief, meeting prep, free up time, group scheduler notion (4 skills): spec to implementation, research docs, meeting intelligence slack (6 skills): daily digest, channel summary, notification triage, reply drafting game-studio (9 skills): phaser, three.js, react-three-fiber, sprite pipeline full repo: github.com/openai/plugins enjoy.

most people try to fix AI writing in the prompt. the better loop: 1. let it draft 2. edit it hard 3. script diffs the first and final version 4. another LLM analyzes the diff: what did you change and why 5. turns those into rules, writes them back to the skill file 6. next time it doesn't make the same mistakes example: i kept cutting replies that opened with 'The "..." is ...' and rewrote them to just say the point directly. script detected the pattern, auto-generated a rule: skip the quoting frame, respond with the substance. 48 rules. 740 lines. none hand-written. style isn't prompted. it's accumulated from your edit history.

@woriwka topics help when you have multiple agents reporting into the same group. without them everything blurs into one stream and you stop reading. i use telegram because the bot API is simpler but if i started over i'd probably pick discord for the threading.

@Voxyz_ai Does splitting conversations into topics in Telegram bring real benefits, or is it an unnecessary solution?

running an AI team looks cool until you open telegram to 30 unread messages from your agents. nexus reporting ops status, scout evaluating the next tool to build, 23 updates in the radar channel. my first instinct is the same as with real coworkers. coffee first. the one advantage agents have over real employees: they don't mind being left on read.

@thomasparas appreciate it man. honestly the documenting part is half for me too. if i don't write it down my agents will just gaslight me into thinking everything went smoothly.😂

@Voxyz_ai BTW ty for documenting all of this so thoroughly - been a fascinating ride 🙏

wrote 14 chapters on claude code. not tips, not a tutorial. the actual system i run in production with 6 agents. if you only read one chapter, start with chapter 2: persona stress-testing. example: give your project to claude, ask it to spawn 6 people who'd actually use it. each one reads through and flags where they'd quit, get confused, or stop trusting. 15 min, 5 blind spots you might missed in 2 weeks.

@goncalo_pr_ really appreciate this. the path-scoped rules is one of those things that sounds small but changes everything once you start using it.

Great article Vox, thank you for the insights. Been running something similar: an adversarial agent that attacks the plan before execution. Started as a simple prompt but now applying the same thinking to other builds on a more abstract level too. 100% agree with your point about "knowing your thing too well" is exactly why I run it with a separate agent with no context on previous decisions. Fresh eyes, not informed eyes. Taking the path-scoped .claude/rules/ from your setup, fills a gap I've been working around manually.

a dead simple way to find problems you can't see yourself. example: i built a system and used it for two weeks. felt solid. then i set up a workflow that spawns 6 AI personas in parallel: skeptical engineer, security reviewer, new maintainer, CLI power user, SRE, docs-first newcomer. handed them my system and told them to break it. 15 minutes later they found 5 problems i missed in two weeks. same method works for anything you build. code, legal briefs, course design, marketing funnels, patient intake forms, pitch decks. have your AI spawn 6 relevant personas who would use it differently, and let them stress-test it in a simulated real environment. a lawyer finds the liability. a new hire finds the confusion. a power user finds the edge case. a customer finds the objection you never considered. you know your own thing too well. that's the problem. wrote up 14 chapters on turning claude code from a tool into a system. this was chapter 2.

@atomtanstudio that's awesome to hear. love that you used it for a live demo. what kind of agents is your team building?

@Voxyz_ai Thanks for having your virtual office available to the public. I used it for my demo on agents today in a meeting I had with my team. People loved the agents moving around and conversing.

good breakdown. the part that hit hardest in production for me was agents reporting success while quietly getting it wrong. looked fine, passed checks, but the output was subtly off. ended up building a second verification pass just for that. the agent thinks it's done, another layer checks if it actually is.

@thomasparas mine haven't complained yet. give it a few firmware updates.

@Voxyz_ai Real employees have an unreasonable issue with being called clankers as well

@ItIsRaymo you're right actually. if i tuned their personality to be more urgent i'd probably panic every morning. right now they report a cost spike the same way they report a typo. honestly i kinda prefer it that way.

@Voxyz_ai but they also probably lack a sense of urgency right? so if you don't have something crazy setup for them to alert you in case of emergencies, you could be missing out on a lot

@chooseliberty 130 pages of pure nonsense is genuinely the funniest failure mode. what did it even write?

@Voxyz_ai yeah just wait until you wake up in the middle of the night at 1am and see they've been working autonomously on 130+ website pages of pure nonsense AND PUSHING THEM LIVE LMAO that woke me up for sure

@alphabatcher appreciate it. pretty sure most of them followed to watch my agents wake me up at 3am with problems they caused themselves.

@Voxyz_ai 8 weeks ago you have only 370 followers? how...

8 weeks ago randall cold-DMed me when i had 369 followers. zero coding background, political consultant. today he runs 8 agents across 7 providers with 18 custom skills and produces client work that used to take him weeks in under an hour. he never learned to code. he learned to describe what he wanted clearly enough that AI could build it.

@BrandGrowthOS 2 already puts you ahead of most. the real trick is letting claude pick which personas to spawn based on what you built and honestly if you don't mind the token bill you can run 10-20 in multiple rounds. each round finds stuff the last one missed.

@Voxyz_ai this is brilliant. i do something similar with my client projects - run the typescript workflow through claude once as the excited builder, then again as the skeptical user. the second pass always catches obvious gaps. 6 personas though... stealing this

most of the time i don't pick them manually. give whatever you built to claude, ask it to generate 6 potential persona who'd actually use it. code, landing page, onboarding flow, legal doc, anything. each persona walks through it and flags where they'd quit. works on everything, not just writing.

@Voxyz_ai the 6-persona testing framework is genuinely novel. how did you decide on those specific roles versus just testing with random prompts?

@CChirchi great! let me know how it goes

@Voxyz_ai whoa, trying this on my app tomorrow!!

weirdest part of running agents: checking in and finding tools you didn't ask for, solving problems you didn't know existed. 46 and counting. i stopped keeping track.