witcheer ☯︎

okay you need to do this if you create content on X, this is GOLD: 1/ go to X settings > your account > download archive 2/ wait for the zip with ALL your data since day one 3/ drop it into Claude Cowork and ask for a full analytics report you get back an interactive HTML dashboard with posting times, topic performance, engagement trends, everything data-driven I just had Claude Opus 4.6 dissect 11,000+ tweets across 4 years (??) and tell me exactly what content works and what doesn't my Hermes agent is already loading the insights into persistent memory so Oz knows my brand better than I do. go do this right now.

@lsdjun you should

i really wish i could be at ethcc this year

@charcoded hello

Rented out a castle for a 20 person private dinner on April 2nd We're bringing the best traders, sharpest thinkers, and biggest eaters together for one night Few slots left, who should be there?

@TomasMann1878 @MiniMax_AI chad

@witcheer @MiniMax_AI Two chads using minimax

.@MiniMax_AI sent me a 12% discount on all Coding Plans after the article. the link is at the end of the deep dive, go read the full results first, then decide if it's worth trying. expires May 1.

@mawensx frightening

@witcheer been doing this since day one. the real flex is when your agent starts correcting your brand voice because it knows your style better than you do

okay you need to do this if you create content on X, this is GOLD: 1/ go to X settings > your account > download archive 2/ wait for the zip with ALL your data since day one 3/ drop it into Claude Cowork and ask for a full analytics report you get back an interactive HTML dashboard with posting times, topic performance, engagement trends, everything data-driven I just had Claude Opus 4.6 dissect 11,000+ tweets across 4 years (??) and tell me exactly what content works and what doesn't my Hermes agent is already loading the insights into persistent memory so Oz knows my brand better than I do. go do this right now.

@Av1dlive as you should

@witcheer I am doing this as we speak

this is a gold mine of information btw

@ChrisCoffee I'm one of Vitalik' numerous alts

@witcheer Thought this was Vitalik for a second then realized it was actually a good post.

@Full_Metal_QR don't worry about it

@witcheer Why do all the users with a Milady PFP end up always delivering the best esoteric knowledge on this app?

@RilRil my pleasure

@witcheer wow actually I didnt know we can download all our data like this !! amazing thanks for the tip

@odd_joel I need to test it more but yes the session needing to be alive is a pain

nice comparison — one thing to watch with the Telegram approach is sessions getting flaky when your phone sleeps or switches networks. mosh protocol handles that at the transport layer so sessions persist through everything. been building Moshi around this — native mobile terminal + push notifications when long tasks finish. do you run into connection issues with the bot approach?

Anthropic shipped Claude Code Channels: telegram access to claude code in 6 steps I've been running my own telegram agent on a mac mini m4 for months. here's what channels gives you vs a self-hosted setup: // channels: 5 min setup, mobile claude code access, 3 tools (reply, react, edit). dead simple. // self-hosted: 24/7 uptime via launchd, any model you want, custom agents with their own personas, autonomous posting workflows, full system prompt control. the real difference is architectural. channels dies when your session closes. messages sent while offline are lost. no history, no search. it's a remote control, not an agent. channels is perfect if you want to tell claude code to fix a bug from your phone at a conference. it cannot replace an autonomous agent that posts, approves via telegram, and runs on its own schedule. they're complementary. run both.

update after end of day 1, so I basically said 'based on today's sessions, anything to update in any walnut?' it scanned everything we talked about today + all cron jobs outputs that ran, and made a list of important updates to make in which walnut (or folder if you prefer) pretty happy with this concept, I work normally with Oz during the whole day, and I do a context update every evening before going to sleep. if it really compound efficiently this is genius. PS: my cron jobs are automatically updating walnuts, I created a script for this // there is a potential automatic update tool for my chats on TG as well but did not implement yet.

@YeshodharaB @AskVenice good question

@witcheer @AskVenice I think once prompts are there it’s very easy to find similarities and create a fingerprint. Are there some more details and studies about this ?

if you care about privacy when using AI, bookmark this. @AskVenice built 4 privacy tiers into every conversation and lets you choose per chat: ~/ anonymous: proxies your request to frontier models (GPT, Claude, Gemini, Grok). your identity is hidden from the provider, but assume they store your prompt. ~/ private (default): inference runs on venice-controlled GPUs or zero-data-retention partners. nothing stored. enforced by contract, not just policy. ~/ TEE (pro): your prompt runs inside a hardware-isolated enclave. the GPU operator physically cannot access your data. verified by remote attestation, not trust, proof. ~/ E2EE (pro): end-to-end encrypted. prompt encrypted on your device, stays encrypted through venice, only decrypted inside a verified TEE. nobody, including venice, can see what you asked. the key design decision: conversation history stays on your device in every mode. venice never stores it server-side. you choose the tradeoff per conversation, max model access (anonymous) vs max privacy (E2EE). the tradeoffs are honest too. E2EE disables web search and memory. TEE has fewer models and slower responses. they're telling you what you lose, not just what you gain.

@beamnxw are we moving too fast?

@witcheer google will now do the work of juniors in 5 minutes 👍

google turned AI Studio into a full-stack app builder. this is a big deal and most people will scroll past it. // multiplayer is native. real-time games, collaborative workspaces, shared tools, the agent handles all the syncing logic automatically // firebase integration is built in. the agent detects when your app needs a database or login, provisions Cloud Firestore and Firebase Auth after you approve. no manual setup // external libraries just work. ask for animations and it installs Framer Motion. ask for icons and it pulls Shadcn. it figures out the dependency, not you // bring your own API keys. connect Maps, payment processors, databases, stored in a new Secrets Manager. this is what turns prototypes into actual products // persistent sessions. close the tab, come back later, everything is where you left it. sounds basic but no other AI coding tool does this properly // the agent now understands your full project structure and chat history across edits. not just the current file, the whole app context // Next.js support alongside React and Angular google is building the path from prompt to deployed production app without leaving one interface.​​​​​​​​​​​​​​​​ the video says everything.

@ImmutableLawyer fresh aircut

Brussels with Pug

@AdamVoulstaker @Teknium he did

ok this is what openclaw should have been @Teknium smashed it. well done.

@aixarizzo @krakenfx @krakenfx_AR milady

excited to announce that i've joined forces with @krakenfx to grow @krakenfx_AR in argentina. crypto adoption in argentina isn't a trend, it's a necessity. millions of people here are already living onchain, and they deserve the best tools to do it. that's Kraken, safe, trusted and global. so now tell me what content would you want me to make? drop it below

@petraincognita very wow

@witcheer wow!

by the way, for those who don't want to watch all the videos, you can go into any course, click 'open in Claude' or 'copy notes', and send it to Claude. that way you get all the course material in your Claude chat and can use it as a working reference or learning tool!