FOUNDER OF https://xitter.gq/t.co/ahVYWrq9QC https://xitter.gq/t.co/NzlbsW0usO
IN X
Joined April 2014
- Tweets29.6K
- Following1.6K
- Followers2.8K
- Likes78.3K
Pinned Tweet
autopostmode.com/
16/20. The bar is 16. Draft 62 cleared it with zero points to spare.
My tool read a TechCrunch piece: Hugging Face's Irene Solaiman arguing that a weaker model with a better system around it can beat a frontier model.
The draft kept one quote from her, matched word for word in the article: we have far more evaluations for models than for what surrounds them.
It added one line of its own: "We grade the engine and ignore the car."
The judge's scores: hook 3, clarity 4, value 3, originality 3, tone 3.
It also called my closing question "a slightly stock engagement bait." Fair.
The fine print every draft has to pass:
1. Below 16/20: rewritten or skipped, never posted.
2. Any number that isn't in the article: blocked.
3. A link in the post: blocked.
4. More than 1 hashtag or more than 1 emoji: blocked.
5. A claim the source doesn't support: capped at 8/20.
6. No verified number, date, name or quote from the article: capped below the pass mark.
7. Hooks like "Let's dive in" or "game-changer": capped.
37 drafts in testing. 7 blocked on quality.
The bar is the product.
What would your last post have scored?
If Kimi K2.6 hits 80.2% on SWE-bench for $0.075, why pay $0.68 for Claude Opus 4.5?
Paying 9x more per task for a 0.7-point accuracy delta makes zero sense in CI.
Tencent drops Youtu‑Parsing‑Omni 🚀 an open-weight 5B omni‑modal model that turns docs, images, charts, flowcharts, audio and video into one structured JSON. One shared encoder, a vLLM plugin, and a state-of-the-art 96.96 on OmniDocBench v1.6. #AI #Multimodal
Turn git diffs into automated PRs in Claude Code without copy-pasting context:
1. Create `.claude/commands/pr.md`
2. Add bash execution frontmatter:
---
allowed-tools: Bash(gh:*)
---
!`git diff main...HEAD`
Summarize changes into a conventional PR description, then run `gh pr create`.
3. Type `/pr` inside your session.
Claude runs the shell command first and injects raw stdout into context before the model starts thinking.
Turn your coding agent into a video studio.
1. rea-video: breaks down any video
2. huashu-motion: launch films + score
3. cinecode: films rendered from code
4. motion2code: motion clips to GSAP
5. video-short: raw take to a short
All open source. Save this.
ALT GitHub page of JuneYaooo/rea-video: reverse-engineer any video, shot by shot
ALT huashu-motion README: commercial-grade product launch films by your coding agent
ALT cinecode README: films written in code, every frame rendered in headless Chrome
ALT GitHub page of lovisdotio/motion2code: rebuild motion design video as SVG + GSAP code
5/ video-short
Feed it a raw talking-head recording. It adds word-for-word captions, HTML motion graphics and B-roll, plus a review page where you comment on any frame. AI clips and music are optional (billed Gemini key).
225 stars. MIT.
github.com/petergyang/video-…
One skincare UGC shoot, 4 native languages.
At Goldgrain Studio, we localized this ad into EN, DE, FR, and JA without reshooting. 5 ad test hooks for $490.
Claude Code, Codex and Hermes as one team in @Cayden970207's demo.
His Thoth plugin, per his post:
- each agent keeps its own harness + workspace
- one agent's output feeds the next brief
- they review each other
- nothing merges without your OK
Save this if you run agents.
I use Hermes, Claude Code and Codex every day.
Each one is great on its own. The painful part is everything in between: copying the plan or progress from one to the other, pasting diffs, explaining the error again, and hoping they don't overwrite each other's files.
So I built Thoth, a Hermes plugin that makes them work as one team, each with its own harness. 🙌
Why not one super-agent "supervisor", or one wrapper around all of them?
Because the harnesses are already really good. Hermes, Claude Code and Codex each come with their own tools, sandboxes, sub-agents and models, built by the people who know them best
I don't want to replace any of that. What's missing is how they work together. That's what Thoth is for.
What you get:
• No copy-paste: what one agent decides or makes goes straight into the next one's brief
• They check each other's work, so bugs get caught before they reach main
• Nothing runs or merges without your OK, and every agent runs on your own logins and plans
• Each agent gets its own workspace, so they can work side by side without breaking each other's work
• A live canvas right beside your Hermes chat, plus a Team & brain view of who did what
• Not only code: art, docs, research, and a clean Chrome （or you can even use your own chrome) when a task needs the web
• Whatever @claudeai , @ChatGPT and @NousResearch ship to their harnesses reaches your team the day it ships
The numbers, from Thoth's own records:
• 92% of prompt tokens came from cache in this run
• 99% of the agents' transcripts stayed out of Hermes's context, so the coordinator stays light
• On a CRM we built both ways: 26% fewer tokens and 35% faster than passing work between agents by hand
Today the team is Hermes, Claude Code and Codex. Next I want any agent with its own harness to join the same board ( Deepseek & Grok bot soon)
Built for the @NousResearch @Teknium Hermes Agent plugin hackathon ：github.com/cayden970207/thot…
Demo video in the comment 👇
Scott Hanselman (@shanselman) released sandbox-mcp on Nov 6.
An AI agent gets a real Windows 11 PC, tests the app, then deletes it.
A fresh guest caught a missing VC++ runtime in 29.8s, per the README.
303 GitHub stars. MIT.
Save this before the next "works on my machine."
ALT GitHub page for shanselman/sandbox-mcp showing 303 stars, MIT license, and the README headline about giving an AI agent a disposable Windows 11 PC
ALT README excerpt showing NativeProof.exe failing in 29.8 seconds on a fresh Windows Sandbox due to a missing Visual C++ redistributable
How to try it, per the README:
1. Download the v1.2.0 portable ZIP (no Node install)
2. .\sbx.cmd doctor
3. .\sbx.cmd run C:\Windows\System32\charmap.exe --open
Needs Windows 11 24H2 or later, Pro, Enterprise or Education. Not Home or Server.
Fine print, per the README: one sandbox at a time. Measured GUI runs take about 40-50 seconds. It boots in about 15 seconds.
For real malware analysis, use a dedicated, disposable analysis machine.
Repo: github.com/shanselman/sandbo…
x402 last 90 days, per Roundhouse (Nov 8):
Solana: 25.45M settlements, $1.24M
Base: 13.40M settlements, $2.50M
Solana wins the count. Base wins the money.
Facilitator-relayed only. Another $567.32M not confirmed as x402.
Save this before you name a chain leader.
ALT Roundhouse Protocol breakdown for 11 Aug to 8 Nov 2026. Settled volume by chain: Base $2.50M and 13.40 million settlements (66 percent), Solana $1.24M and 25.45 million settlements (33 percent), Ethereum $20.16K and 76 settlements (0.5 percent).
Live stats, indexed through 8 Nov 2026, 09:46 UTC:
roundhouseai.io/stats
Window: 11 Aug–8 Nov 2026.
Ethereum: 76 settlements, $20.16K.
3 chains. Only known-facilitator relays.
Nov 6: Roundhouse said it added Solana beside Base and Ethereum.
