Kodzilla

I GAVE 7 AI BOTS A JOB. THEN MADE THEM WORK TOGETHER. One finds ideas. Another judges them. Others build, test, and review. I broke down the entire system, including the handoffs that make it work.

NOW THE SPEED SIDE Same question, same document. Reader answers in 80 ms. Claude Opus 5.5 needs ~5 s and Haiku 5.5 ~7 s just to start writing. Per 1M questions: $6.50 vs $172 vs $6,880. Waitlist is open: platform.xerj.ai

XERJ Reader: a neuro-symbolic AI model for data work. It doesn’t generate text. It highlights the answer in the source document, or says there isn’t one. $0.01 per 1M input tokens. $0 for output. ~1,050x cheaper than Claude Opus 5.5. 26x cheaper than Claude Haiku 5.5.

STOP BUILDING AI AGENT TEAMS FROM SCRATCH. 15 READY GROK BOT REPOS ALREADY EXIST. most people still burn weeks wiring roles by hand. these 15 are already sitting on github: code, research, marketing, trading, even the personal shit. the play: 01 / Agent Org Chart — ready structure, eng + marketing + support + sales github.com/serenakeyitan/... 02 / Grok Factory — one coordinator dumps work on specialist agents github.com/jaredtrichard/... 03 / multiBot — spins up teammates and hands them straight to CLI github.com/simo255/multiBot 04 / Grok Bot Second Brain — 5 agents split coordination, memory, research, ops github.com/.../grok-bot-seco… 05 / Foreman — watches the team, spots stuck tasks, checks what actually got delivered github.com/Archive228/forema… 06 / grokbot-x — research, drafts, posts, rewrites for X in one chain github.com/YannisKiefer/grok… 07 / PhoneZero — makes calls and books tables, you still have to nod on the key steps github.com/function1st/Phone… 08 / BOTROSTER — self-hosted crew, shared files, browser, scheduled jobs github.com/mandarwagh9/botro… 09 / Grok Bot Profiles — 4 roles: write specs, repro bugs, open PR, review PR github.com/HAEGONG/grok-bot-… 10 / HyperGrok Trading Desk — 7 agents for market, research, strategy, risk, orders github.com/galleonlabs/hyper… 11 / Cryptofeed — live crypto trades, prices, order books github.com/.../cryptofeed 12 / CCXT — one API across 100+ exchanges github.com/ccxt/ccxt 13 / Freqtrade — backtest and run strategies, paper + live tracking github.com/freqtrade/freqtra… 14 / VectorBT — test thousands of strategy combos on history in one go github.com/polakowo/vectorbt 15 / QuantStats — one-click report: returns, drawdown, volatility github.com/ranaroussi/quants… pick one team. give it a clear job. wire the specialists with API, MCP or script. define what each owns and what still needs your yes. stop starting from zero every time.

TWO MODELS. ONE BLACK HOLE. SOL ON TOP. GPT-6.2 Sol and Opus 5.5 were given the same scene. The difference is visible in the first few seconds. Sol holds the disk and the lighting. Opus still blows out the center. That is the part people notice when they stop scrolling. Most people still route 3D and design work through Opus by default. The ones who switch early are the ones who test it on one real job instead of a leaderboard. Astra is not out yet. Fable is the next one people will actually run. Until then the catch is simple: the model you already pay for is not always the one that holds up on the scene. bookmark this and reply which one you would hand the next 3D job to.

HOW TO LET CLAUDE CODE CODEX OR CURSOR CONTROL YOUR DESKTOP APPS Vercel just open-sourced agent-desktop. your agent no longer stays inside the terminal. it can click and type in real apps on your machine. the play: install agent-desktop > connect it via MCP to Claude Code / Codex / Cursor > give it a task that needs the desktop > watch it run in the background steps: 1. clone the repo github.com/vercel-labs/agent… 2. run the setup so your agent can see the accessibility tree of open apps (Notes, Reminders, browsers, etc.) 3. start Claude Code, Codex or Cursor and ask it to do something that requires the desktop (update a reminder, fill a form, debug an app) 4. leave it running. it works in the background while you do other things. multiple agents can run at the same time, each with its own cursor. > the catch it needs real desktop permissions. once you give them, the agent can click anything the logged-in user can click. > what not to do do not run it with full permissions on a machine that has production logins or secrets open. start with a test user or a clean session. you asked it to help with the work. now you are the one watching the screen. bookmark this and reply with what desktop task you would hand off first.

MY 7 BOTS WERE WAKING 1176 TIMES A WEEK AND MOST OF THEM DID NOTHING Most of them opened their eyes, found nothing new, burned a run and went back to sleep. So I turned the whole team into a relay. One prompt. Only Scout keeps a clock. Monday morning. The other 6 stay asleep until the bot in front writes DONE in a shared LEDGER file. No DONE, no wake. Scout finds what people in your niche actually complain about. Judge kills weak ideas before anyone builds them. Builder drafts the PR but never merges. Pitch writes the launch posts but never posts. Buyer gets $40 and waits for your yes on every single buy. Same work. 60 wakes a week. Roughly 95% fewer runs. It keeps running with the laptop closed. The full setup and the exact prompt are in the article.

STOP BUILDING YOUR CLAUDE SECOND BRAIN BY HAND Obsidian’s CEO already open-sourced the skills and four more repos finished the rest. Most people still copy broken tutorials from last month. The play: Clone kepano/obsidian-skills. Point Claude Code at your vault. Drop any file into /raw and type ingest. The map writes the page, links it to five existing ones and updates the rules with how you actually think. After 90 days the search bar cites your own notes with dates. Google cites strangers. Your vault cites you. The people running this don’t open ChatGPT anymore. They open their own search. The catch: without a source date on every ingest the vault starts inventing. One missing date and it fills the gap with whatever sounds right. One 5-minute setup. Runs forever on the laptop in front of you. Compounds for years. Bookmark this before you open another empty chat.

YOUR AGENT FORGETS EVERYTHING WHEN YOU SWITCH TOOLS Living Brain keeps one persistent memory across ChatGPT, Claude, Cursor, Codex and your own agents. Write in one, retrieve in any other. the play: → connect via plugin or MCP → save a decision or note in one tool → open the next tool and ask for it → it pulls the memory in ~464ms > the catch the memory lives outside the model. check what gets stored and who can access it > what not to do don't dump every chat into it on day one. start with decisions and project context only one brain that travels with you beats rewriting the same context every time you switch tools bookmark this

HOLY SH*T. YOU’VE BEEN GOOGLING STRANGERS FOR 20 YEARS Karpathy’s 5 minute second brain setup crossed 16m views. one folder, one Claude.md every article, podcast, meeting and note you already consume becomes searchable by you. Install Obsidian, open Claude Code in the vault, drop anything into /raw and type ingest. it writes the wiki page, links it to 5 existing ones and updates the rules with how you actually think. After 90 days the search bar cites your own past notes with dates. Google cites strangers. your vault cites you. The people running it don’t open ChatGPT anymore. they open their own search. One extra rule i added: every ingest must cite the original source date. without it the vault starts inventing. with it, after month 3 it drafts in your voice better than you do at 2am. Your coworker started this on a saturday in february. by august his tweets landed 4x your engagement on the same topics. by october his first draft of any doc was already better than your third. Compounds for 10 years. after year 3 your vault knows more about what you believe than you do. 5 minutes to wire. runs forever on the laptop in front of you. Bookmark this before you open another empty chat

NO F*CKING WAY SPACEXAI HANDED CODERS 345 TOOLS FOR FREE 17 official Grok Build plugins that turn one agent into a full stack team. all of them already on github people drop $19 on a single skill. here you get 150 skills, 5 agents, 59 browser tools and 131 web guides. roughly $3k in skills, zero to you already sitting at 380k stars combined. late, but not too late your agent writes code. it still can't open your figma, hit the database or read the error killing signups 17 plugins worth installing first: 1. superpowers → 15 skills. plan first, test first, debug properly. 304k stars github.com/obra/superpowers 2. pstack → 58 skills + 2 agents from cursor. investigate, design, review, verify github.com/cursor/plugins 3. modern-web-guidance → 131 how-tos from google so it stops writing old web code github.com/GoogleChrome/mode… 4. figma → reads real designs and turns them into code github.com/figma/mcp-server-… 5. chrome-devtools → 59 tools. live chrome for the agent. clicks, network, performance github.com/ChromeDevTools/ch… 6. browser-use → agent clicks through the app in a real browser and qa’s it github.com/browser-use/plugi… 7. supabase → database, auth, storage straight from the agent github.com/supabase-communit… 8. mongodb → 8 skills for schema, queries, search, slow-query fixes github.com/mongodb/agent-ski… 9. stripe → official payment best practices + mcp server github.com/stripe/ai 10. firecrawl → any site into clean llm-ready markdown github.com/firecrawl/firecra… 11. exa → real-time web search with sources github.com/exa-labs/exa-grok… 12. tavily → search, extract, crawl, deep research github.com/tavily-ai/tavily-… 13. vercel → 37 skills, 3 agents, 4 commands for deploys and logs github.com/vercel/vercel-plu… 14. cloudflare → 14 skills for workers, durable objects, wrangler github.com/cloudflare/skills 15. netlify → 12 skills for functions, edge, blobs, deploys github.com/netlify/context-a… 16. sentry → real prod errors and stack traces inside the agent github.com/getsentry/plugin-… 17. datadog → logs, metrics, traces in plain english github.com/datadog-labs/grok… pick 3 today. one that thinks, one that ships, one that catches bugs test each on one real task. if the output isn’t better, uninstall it --trust lets a plugin run code on your machine. two minutes on the repo saves a bad week bookmark this. the tool you were about to build is probably already here

HOW TO RUN SONNET AS LEAD HAIKU AS SWARM AND OPUS AS ADVISOR ON CALL Anthropic just shipped the cleanest multi-model split in Claude Code. one command, three models, way fewer tokens. the play: sonnet 5.5 drives the session > haiku 5.5 swarms search/grep/audit in parallel > opus 5.5 sits on call and only reviews the final diff steps: 1. open terminal and run claude --model sonnet --advisor opus code.claude.com/docs/en/sub-… 2. create three haiku 5.5 subagents in .claude/agents: - search (maps the repo) - grep (finds call sites) - audit (checks the diff in a throwaway worktree) 3. add one rule to CLAUDE.md: send every repo search and test run to the haiku 5.5 subagents so the main context stays small 4. route fast deterministic forks (which file, which tool, retry or stop) through JEV so the big models never see them 5. paste this prompt so Claude rebuilds the tree and shows every change as a diff first: Rebuild my Claude Code setup around this tree: 1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor. 2. Create three Haiku 5.5 subagents in .claude/agents: search (maps the repo), grep (finds call sites), audit (checks the diff). 3. Give audit isolation: worktree so its checks never touch my working branch. 4. Add one rule to CLAUDE.md: send every repo search and test run to the Haiku 5.5 subagents so the main context stays small. 5. Route fast deterministic forks (tool choice, path selection, retries) through JEV. Show me every change as a diff first. Nothing gets edited until I say go. > the catch opus still burns tokens if you call it on every small decision. keep it strictly on-call. > what not to do do not run the whole session on opus or let haiku write code. wrong model = wasted context. sonnet builds. haiku swarms. opus reviews. bookmark this and reply with your current claude setup if you want the exact agents folder structure.

HOW TO BUILD A 200+ AGENT RESEARCH TEAM IN GROK 20 analysts, one portfolio manager, one risk desk. Running from a single interface. the play: portfolio manager → 20 analysts → risk desk → parallel agents > first steps - open Grok and create a portfolio manager agent - spin up 20 analyst agents with clear research scopes - add a risk desk agent that challenges findings - wire the handoffs so they run in parallel > the catch 200 wallets is a starting point, not proof. High profit numbers can be noise, luck, or something else. The agents flag patterns. You still have to verify. > what not to do - don’t treat the output as confirmed insider trading - don’t skip the risk desk. Unchecked agents just amplify noise - don’t scale the team before the handoffs actually work one command, 20 analysts, zero manual wallet checks. reply with the first research task you’d give the team bookmark this before you build the next agent