Welcome back! Mods landed in Claude Code, so you can now change the app itself with one prompt. We also mapped the full AI engineering stack, found a repository blueprint worth stealing, and tested a library that makes typed decisions in a single forward pass.
In today's Generative AI Academy Newsletter:
Anthropic finally added mods to Claude Code: Five steps and two prompts to build a feature Claude ships in about 3 minutes.
Calling an LLM API doesn’t make you an AI engineer: The 7 layers that keep agents running once they hit real traffic.
A production repo blueprint for AI applications: Nine folders, one job each, from request handling to safe rollouts.
Structured decisions without generating a token: Typed classification in one encoder pass, 6 to 7x faster than the generate-and-parse approach.
Anthropic finally added mods to Claude Code
A skill tells Claude how to do a job. MCP connects your tools. A mod changes the app itself.
How to build your own in 5 steps:
1. Switch one on
Update to v2.1.287+, turn on usage data sharing, then run:
/plugin enable cc-plugin-you-should-know@builtin
2. Run the audit
Paste Prompt 1. Claude reads your last 30 sessions and lists what you ask for again and again.
3. Build it
Pick one repeat. Paste Prompt 2, say yes to live updates, and Claude builds and tests the fix in about 3 minutes.
4. Keep it
Your mod disappears when the chat closes. Tell Claude "save this mod as a plugin" and it loads every session.
5. Share it
Put it on GitHub. Your team adds it with one line:
claude plugin marketplace add you/your-mod
Check your idea before you build. Do you ask for this every day? Can you see it without typing a command? Does it run by itself? Can you read it in one glance? Does it keep your keys hidden? Five yeses, build it. Any no, skip it.
Mods worth copying: a panel that tracks every ask until you prove it's done, a bar showing your other sessions, the plan as a live progress bar, a ping when Claude finishes, a guard that blocks key files, deletes that go to the Bin instead of vanishing, an outbox that holds email until you approve it.
Both prompts are in the graphic.
GenAI Academy covers all of it: setting up Claude properly, building custom skills and orchestrating past copy-pasting the same prompt over and over.
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.
Calling an LLM API doesn’t make you an AI Engineer
Most people jump straight to agents and skip everything that keeps them running in production.
The 7 layers:
𝟭. 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀
Python, Git, Linux, data structures, system design. Weak fundamentals show up the moment your agent hits real traffic.
𝟮. 𝗗𝗮𝘁𝗮 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴
Pandas, PySpark, Polars, APIs, cleaning, Airflow. Your RAG system is only as good as what you feed it.
𝟯. 𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴
Scikit-learn, PyTorch, ONNX, model evaluation. Knowing how models are trained helps you debug how they behave.
𝟰. 𝗟𝗟𝗠𝘀, 𝗥𝗔𝗚 & 𝗔𝗴𝗲𝗻𝘁𝘀
OpenAI, Anthropic and Gemini APIs, Ollama, LangGraph, LlamaIndex, MCP. Frameworks change fast. Retrieval, tool use and orchestration don’t.
𝟱. 𝗗𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀
PostgreSQL, Redis, and FAISS or Pinecone for vectors. Picking the wrong store is expensive to undo.
𝟲. 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 & 𝗠𝗟𝗢𝗽𝘀
FastAPI, Docker, Kubernetes, MLflow, monitoring, one cloud. A model that never ships has zero business value.
𝟳. 𝗥𝗲𝗮𝗹 𝗣𝗿𝗼𝗷𝗲𝗰𝘁𝘀
Testing, docs, open source. Build a chatbot, a RAG system, an agent.
You don’t need all of it at once.
Go layer by layer, and build something at each step.
From Claude fundamentals to full content workflows, you learn to work with AI without sounding like it wrote everything for you.
One of the most realistic repository blueprints for production-ready AI applications
It covers everything around the LLM, from request handling to evaluations and deployment. Each folder has one job:
api/ — request handling, typed schemas, auth, rate limits
application/ — workflows, business rules, budgets, output validation
providers/ — model integrations, deadlines, bounded retries, cancellation
security/ — identity, tenant isolation, access control, tool permissions
prompts/ — templates, input contracts, model config, version references
observability/ — tracing, latency, errors, model usage, cost
evals/ — versioned datasets, quality checks, held-out acceptance tests
tests/ — unit, integration, security, failure paths
deploy/ — builds, infrastructure, health checks, safe rollouts
RAG, agents, caching, tool execution and background workers sit as separate capabilities rather than requirements. A retrieval-heavy product leans on RAG.
An agentic one leans on tool execution. The structure fits both without forcing one architecture on everything.
Here's why it matters. The model calls time out. Users request data they aren't authorized to see. A prompt change degrades answer quality without breaking a single unit test.
Your application has to handle all three, and your engineers need enough tracing, tests and evals to catch them before and after deployment.
A folder structure won't do that work for you. It gives the work a place to live.
Start small. Add capabilities when the product needs them.
BTW, GenAI Academy covers this landscape in depth, from Claude fundamentals through full agent orchestration and enterprise-scale deployment.
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.
Run structured decisions over any text without generating a single token
Most LLM classification pipelines generate text, parse it into a structured answer, then handle every case where the output doesn’t match the expected format.
Laya removes that layer. Typed decisions in one encoder forward pass. No generation, nothing to parse, nothing to hallucinate.
Three decision types cover most classification and routing work.
Choice returns a top label with probabilities across all options. Score returns an expected level on an ordinal rubric.
Bool returns a calibrated probability of yes or no. All of it runs in a single forward pass, however many questions you ask at once.
A built-in Router picks between three checkpoints per request: one for English, one for 100+ languages, one fine-tuned for typed decisions.
Language detection takes under 0.5ms, so the right checkpoint gets selected without you deciding.
What you get:
33ms per question on GPU, 7.2ms batched, measured on a T4
6 to 7x faster than TypeSafe Jev at 236-276ms p50
LangChain and LangGraph integration for routing and guardrails
MCP server to expose decisions as tools in Claude Code, Cursor and other agents
Presets for prompt guardrails, content moderation, ticket triage and model routing
Self-hosted server with a Jev-compatible wire protocol, so existing clients change only a URL
TypeScript package that returns identical answers to the Python one
pip install laya. Apache-2.0.
Everything else you shouldn't miss
OpenAI cannot make AI safe on its own: Three fired safety researchers say the dismissals will chill internal dissent and outside collaboration. They want independent evaluator access preserved, chain-of-thought monitorability protected and clear rules on talking to external safety groups.
OpenAI’s fastest Sol ships today: Ultrafast mode for GPT-6.1 Sol lands in the API, Codex and ChatGPT Work, running up to 8x faster than Sol Standard at near-Astra intelligence. Pricing is $12 per million input tokens and $60 per million output.
Google’s Gemini agent works across your whole stack: A universal agent that handles knowledge work, media, coding and multi-step tasks across Workspace and enterprise tools. It adds persistent context, multi-agent orchestration, model routing, governance and cost controls.
Midjourney is testing a thinking mode: Image generation that reasons before it renders is now in testing on the Alpha site. No word yet on how it changes output quality or how much slower it runs.
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.
