Welcome back! Skills turn a long prompt into one command, and they save your usage limit twice over. We also compared three serving engines, found an agent setup that discovers tools at runtime, and broke down why your agent starts from zero every session.
In today's Generative AI Academy Newsletter:
33 free Claude skills you don’t know: Build the instruction once, call it with a short command, and stop burning tokens on context you already wrote.
Ollama vs vLLM vs SGLang: Three engines, three traffic patterns, and one prompt trick that speeds up all of them.
Build cross-app AI agents with a single MCP: One connection instead of a server per app, with the agent finding the action it needs at runtime.
Short-term vs long-term memory in AI agents: A 1M token window still resets after every call. Here’s what fills it back up.
33 free Claude skills you don't know
Don't write every single prompt from scratch!
Instead, do this ↓
Skills are essentially pre-built shortcuts for Claude, saved instructions you trigger with a short command instead of typing the same context every time.
Need a LinkedIn post drafted in your voice?
One command handles it.
Want AI-sounding writing cleaned up, meeting notes summarized or a spreadsheet fixed?
Same idea, one command each time instead of a fresh explanation.
The logic is very simple: build the instruction once, save it as a skill, then call it whenever you need it.
Retyping context also eats into your usage limit, so saving it pays off twice.
A Claude course from Remus Ranca covers 24 moves for making every session last longer.
They span Claude.ai, Cowork and Claude Code.
Master Claude Code in 6 sessions
You already know the developers who seem to move twice as fast as everyone else.
They're not typing faster. They're:
Orchestrating teams of agents
Running parallel Claude instances across git worktrees
Wiring MCPs into their whole stack
Scheduling loops that keep working while they sleep
Almost none of it is written down in one place.
Mastering Claude Code course puts it all in one place.
6 live sessions, each one hands-on:
You start with the agentic loop and model selection
You build Skills and full agent teams from scratch
You apply Boris Cherny's 13 tips plus 6 new April 2026 techniques
You set up all 3 tiers of CLAUDE.md as a second brain that gets smarter every session
You wire MCPs into Slack, GitHub, Gmail and Supabase
You leave having built, tested and shipped your own custom Skill as a capstone
This is the deep track, the habits that separate casual users from power users.
Pricing is locked in for early sign-ups.
Ollama vs vLLM vs SGLang
They all run open models. Each one is built for a different kind of traffic.
Ollama: one user, one machine
→ Requests wait in a queue, 1 per model at a time by default
→ Runs quantized models: GGUF via llama.cpp, and MLX by default on Apple Silicon since v0.40
→ Speaks both the OpenAI and Anthropic API formats, so Claude Code can run on a local model
vLLM: many users, mostly unique prompts
→ Continuous batching adds new requests to the running batch at every step
→ PagedAttention stores the KV cache in pages, and prefix caching is on by default
→ Supports the widest range of models and GPUs
SGLang: agents and multi-turn chats
→ Its router sends each request to the worker that already holds its prefix
→ RadixAttention keeps shared prefixes in a radix tree and reuses them
→ Fast at JSON and regex-constrained outputs
How to pick:
→ Prototyping on a laptop or Mac: Ollama
→ Serving lots of users with varied prompts: vLLM
→ Agents, RAG and long chats that repeat the same context: SGLang
One tip that helps on all three:
Keep your system prompt and tool definitions identical at the start of every call, and put whatever changes at the end. That's what lets the prefix cache hit.
Build cross-app AI agents with a single MCP!
Most agent setups wire every integration upfront.
GitHub gets a server, Slack gets a server, Notion gets a server, and the agent carries all of them whether the task needs them or not.
This one connects to a single MCP server and finds what it needs at runtime.
Liner Actions MCP sits in as the action layer. The agent searches for a capability when the task calls for it, then executes through the same connection.
Here’s what that looks like. Ask it to create a GitHub issue and post it to Slack.
The agent searches for the GitHub action, runs it, takes the returned issue URL, then searches for the Slack action and passes the URL into it.
The agent still decides what comes next. Liner handles finding the specific action, connecting to the service and executing against it.
Account connections become just-in-time too.
If Slack isn’t authenticated yet, the user connects it at that moment and the task continues, without the agent ever holding a direct connection to a Slack MCP server.
What you get:
One MCP connection covering actions across many apps
Runtime discovery of the exact action a task needs
Cross-app sequences where one result feeds the next
Authentication only when a service is actually required
Execution against the underlying app without exposing every integration to the agent
The change is in when the agent learns what it can do.
Handing it every tool at startup becomes handing it a way to find the one tool the task requires.
You don't need to pay to learn Claude
Seven Claude courses at GenAI Academy. All recorded, and all free.
Start with "What Can AI Actually Do For You?" if you're new, 30 minutes, no setup required.
Then the Claude Starter Course gets you running in 35.
Past that, pick by the problem you have. Hitting usage limits in Claude.ai, Cowork or Claude Code has its own course.
Building no-code agents has one.
So does rolling Claude out across an ops team in 30 days, from pilot to adoption data to full deployment.
The AI Portfolio Builder turns whatever you built into a case study you can show someone.
World Wide Vibes Hackathon is live too, $5,000 prize pool, beginner-ready, 100% online.
One login. Watch them in any order.
Short-term vs long-term memory in AI agents
Your agent handles a conversation well. The next session it starts from scratch, no context, no history.
That's a memory problem, and it comes down to two types.
Short-term memory (the context window)
Working memory for a single call. System prompt, conversation history, tool outputs and retrieved content all go into the window so the model can reason over them.
The window resets when the call ends. Inside a long conversation, the model can also lose track of details it saw earlier in the same window.
Long-term memory (external storage)
Lives outside the model and persists across sessions. In practice it splits three ways: episodic memory for past interactions, semantic memory for facts and preferences, procedural memory for rules and processes.
When a new call starts, the agent retrieves the relevant records and injects them into the context window through RAG. When the session ends, the useful parts get extracted and written back.
They aren't alternatives
Long-term memory doesn't replace the context window. It fills it.
A 1M token window still resets after every call.
Single isolated tasks need only short-term memory. Anything that has to remember users, past decisions and outcomes across sessions needs both.
Everything else you shouldn't miss
1Password will log agents in without showing them your password: It verifies the agent’s identity with a unique digital key, sends you an approval request, then fills the credentials behind the scenes. Works with Claude Duet, Dots, ChatGPT Work, Grok Bot, Muse and Town’s assistants. The waitlist is open.
Canva trained its own image model: Canva Craft plans the layout and text before rendering, up to 2K across 30+ aspect ratios, which Canva says produces legible text and cleaner edits. Available through the public API now, with Canva MCP support coming.
China-linked hackers used an open-source AI tool to run an attack: Microsoft says attackers used ARTEX to break into IT firm ConnectWise, then hijacked its trusted apps to attempt logins across 2,250+ customer environments, reaching high-level access at some within three hours.
Claude’s text now carries an invisible watermark: Anthropic’s latest models embed a SynthID-Text watermark in all output to comply with the EU AI Act. It works by nudging word choices between near-equivalent options, with no hidden characters and nothing identifying the user.
Learn more about AI from the experts building it
Follow us on Instagram for fast, visual AI updates in 30 seconds.
Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
Learn how to build your next AI application with practical resources and expert guidance.
Explore investment opportunities in the future of AI and join our community-backed growth journey.
