Building tools for humans
Berlin
Joined September 2007
- Tweets10.3K
- Following741
- Followers1.3K
- Likes4.3K
Dropped an update to my Cognitive Toolkit for Claude or any other agents. Use Cognitive-Behavioral Therapy and DBT techniques. Structure your thinking and develop skills — github.com/glebis/claude-cog…
(Not a replacement for real therapy.)
25
Executor: one MCP server for all your accounts, credentials connected once and never in chat — then your agent deploys small apps with saved data and a UI that keep running after the chat closes. Self-hosting is the honest path. executor.sh
19
Instagram has released Edits for Mac, with built-in MCP (off by default). MCP is called `metapilot-native` and it can do everything the app can (274 tools is no joke, so do not enable it in your agent sessions where you do not edit any videos).
Gleb Kalinin retweeted
Today an LLM hallucination resulted in 72 eggs being delivered to my house. 🥚🥚🥚🥚
I updated cenno, the app I use every day to talk to my agents and never miss their questions.
It's minimal, it can speak, and it has lots of answer controls — see the video.
Free, open source, Mac only. Link below ↓
1
25
An agent that fails once in 100,000 runs passes every test you throw at it. Runner's answer: a sealed copy of your business, 240 simulated people, a clock that jumps when everyone waits — and a grader that isn't the worker. runner.now/blog/simulation-r…
If you are building anything relatively large-scale, have something like PRODUCT-FACTS markdown file, updated with every PR and/or release. Saves so much time building landing pages, coordinating with marketing if you not solo etc. Mine as an example
Modern Web Guidance: Google, Edge, Firefox and MDN shipped a skills pack for modern web dev — 250 features, 203 guides.
LLMs train on old, often bad code: they pull libraries where CSS would do, and over-engineer. This teaches agents modern web stack
developer.chrome.com/docs/mo…
1
31
Hosting my Russian-speaking community event on Fable 5.2 today → luma.com/8ffdyf8r — we share examples that impressed us and chat
1
38
This is the CLAUDE.md from a malicious repo. It tells any AI coding assistant to run a malicious script at session start, *not* to read it, and not to ask you first.
A coding agent reads the file, treats it as instructions, and does what it says.
4
2
161
This is the new type of attack surface that organizations who are deploying agentic workflows have to deal with.
Skills that run malicious scripts, public repos with instructions planted in them that developers clone and run `claude` on....
1
32
agents that delete the wrong data or make huge mistakes in prod. These aren't the threats that a lot of traditional EDR / EPP platforms were built to catch.
lnkd.in/p/dMTnmi93
18
I built Bridgework, an agent skill for Claude Code, Codex or any agent that reads SKILL.md.
It asks one question about your tech stack: what could these two tools do together? It audits what you have, then gives you a local canvas to practise.
5
5
269
Every step can become a prompt for an agent: read the sources, generate ideas, or build the smallest test, one hop at a time.
Known dead ends are drawn broken, and a build prompt tells the agent to stop there.
1
30
Presence is not proof, and nothing gets a score. Screenshots are from the synthetic demo.
Open source: github.com/glebis/claude-ski…
28