Armand Ruiz

Director Data Science & AI Elite

VP of Product - AI Platform @IBM

As the Director of AI Engineering at IBM, I lead the watsonx team, a global group of over 1,000 AI Engineers who work on cutting-edge Generative AI solutions for various industries. With more than 12 years of experience in AI, I have a passion for democratizing practical AI skills and empowering professionals to leverage AI to transform their business.

Previously, I managed IBM's Data Science and AI Elite, delivering complex AI implementations for hundreds of clients. 

I also led the product management of IBM Watson Studio and Watson Machine Learning, an AI platform used by hundreds of customers and SaaS apps powered by AI. I hold a Master of Science in Telecommunications from Universitat Politècnica de Catalunya and speak three languages: Spanish, Catalan, French, and English.

Connect with me!

✉️ JOIN MY NEWSLETTER: Every Saturday morning, you'll get 1 actionable tip to launch and grow your AI strategies in business in less than 4 minutes. → GO HERE: https://www.nocode.ai/

📧 Email: armand@nocode.ai
X: @armand_ruiz


"It is kind of fun to do the impossible"
All opinions are my own and don't represent my employer.

Director watsonx team - Client Engineering

I am building the world's largest team of 𝐀𝐈 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐬 by

- Recruiting hundreds of new engineers around the globe
- Reskilling our existing talent
- Offering them top-quality training in new-generation AI

Responsible for Pre-Sales Generative AI Pilots with the new IBM watsonx.ai Platform.

I’ve been practically living in Claude Code lately. If you build with AI daily, you know that your orchestration workflow is just as important as the model itself.

Recently, Boris Cherny (the creator of Claude Code at Anthropic) shared the internal best practices and workflows his team actually uses on a daily basis. 

Someone brilliantly distilled those threads into a structured CLAUDE.md file that you can drop straight into the root of any project.

It essentially acts as a system prompt that turns Claude into a much more autonomous, rigorous engineering partner.

Here is what the file enforces:
1/ Workflow Orchestration: Mandates a "Plan Node Default" for any task over 3 steps and uses subagents liberally to keep the main context window clean.

2/ The Self-Improvement Loop: This is the real magic. After ANY correction, it updates a tasks/lessons.md file. You are building a compounding system where the mistake rate drops over time because it actively learns from your feedback.

3/ Verification Before Done: It cannot mark a task complete without proving it works (diffing behavior, running tests, checking logs). "Would a staff engineer approve this?"

4/ Autonomous Bug Fixing: Zero hand-holding. You point it at failing CI tests or error logs, and it goes to work without requiring constant context switching from you.

5/ Strict Task Management: Forces a "Plan First" approach written to a todo.md with checkable items before implementation even starts.

It forces the AI to prioritize simplicity, find root causes instead of temporary fixes, and minimize the blast radius of its changes.

If you are spending hours a day in the terminal with AI, setting up a strong .md instruction file like this will save you a massive amount of time and frustration.

It takes time and effort to set this up. If you do, you are ahead of everyone else.

If your employer isn’t providing unlimited Claude Code + secured MCPs to connect your enterprise repos, email, work-chat, meeting AI transcripts, calendar, and docs... they are likely moving at legacy speed. 

Without that context and tool access, you’re stuck in a cycle of manual copy-pasting and reinventing the wheel while the rest of the market moves in parallel workflows.

My advice? Don’t wait for the corporate roadmap to catch up. If you aren't getting these tools at work, force yourself to learn these agentic workflows on the side. 

When companies move this slowly, they eventually run into trouble and you don't want to find yourself at a disadvantage in today's market because your professional environment kept you in the past.

Stay curious and keep your skills ahead of your company's tech stack.

Everyone's talking about which LLM is best for agents. But after reading Lance Martin's deep dive into Manus's architecture, I'm convinced the real differentiator isn't the model. It's the context engineering around it.

The problem: A typical Manus task makes ~50 tool calls. Without careful management, results pile up in the context window and the LLM gets dumber the longer it runs. 

Anthropic calls it "attention budget depletion." 

Chroma calls it "context rot."

Manus solves this with three strategies:
𝟏. 𝐑𝐞𝐝𝐮𝐜𝐞: Older tool results get compacted to file path references. When that's not enough, the full trajectory gets summarized via structured schema.

𝟐. 𝐎𝐟𝐟𝐥𝐨𝐚𝐝: Fewer than 20 atomic tools bound to the LLM. Everything else runs in a sandbox via bash. The model stays lean.

𝟑. 𝐈𝐬𝐨𝐥𝐚𝐭𝐞: Sub-agents exist to give each task a clean context window, not to mimic human roles. No context bleed.

On top of that, Manus doesn't commit to one model. It routes tasks; Claude for code, Gemini for multimodal, OpenAI for math. 

Model-agnostic by design.

The takeaway: Manus has been refactored 5 times since launch. If a stronger model doesn't improve your agent, your engineering is the bottleneck.

Stop obsessing over which model to pick. Start obsessing over the implementation around it.

Full article: https://lnkd.in/gaY3rAV5

Prediction: Token budgets will soon surpass human salaries.

We are entering an era where the cost of compute for reasoning is becoming a primary line item, and the numbers are scaling fast.

I am already seeing developers consuming $100k+ a week in tokens.

It sounds astronomical, but when you look at the output, the ROI is there. They aren't just running a single chat thread; they are spinning up massive parallel processes. Swarms of agents coding, debugging, testing, and architecting simultaneously. The velocity is incredible.

But at that scale, cloud inference costs become the bottleneck.

The big unlock? Local Compute.

Putting a Mac Mini (or equivalent high-performance local hardware) on every developer's desk to run local models is the inevitable next step. We need to move the heavy lifting to the edge to make this sustainable. 

Invest $10,000 on a local computer and you have unlimited AI.

This creates a massive new need for the Enterprise: Fleet Management for Models.

Just as we manage laptops with MDM, companies will need a way to manage these distributed AI systems; securely pushing the latest Open Source models to local machines, updating weights, and orchestrating the workflow between local "thinking" and cloud "verification."

The future is hybrid, and the infrastructure to support it is being built right now.

I’m building my Personal OS to automate my mental overhead. For now, it’s simple, but the potential is already showing. I’ve moved away from "checking notifications" to a Pulse-based system.

How it works: 
1️⃣ Orchestration: Claude.md acts as the central hub, invoked by a launchd scheduler. 
2️⃣ Sequential Logic: Every "Pulse" (Morning/Midday/EOD) triggers a sequence of my agents: Scout → Horizon → Focus. 
3️⃣ Multi-Channel Delivery: I get a unified briefing via macOS notifications, an HTML digest, and my terminal.

The Agent Roster:
- Scout (/scout): Google Chat & Email monitor. It knows my VIP list and drafts replies before I even see the message. If something is important it will notify me, and even give me a call.
- Horizon (/horizon): The calendar expert. No more walking into meetings cold.
- Focus (/focus): The filter. It turns the noise into a top actionable list.

I feel spend less time "managing" my day and more time executing it. 

I also turn all the context for the day ahead into a fun podcast, and that gets me ready for the day and fully prepared.

I've been using computers in a completely different way lately. No much clicking. No browsing through tabs. No filling out forms. I describe what I want done, and agents go do it.

Many are still building software for humans. Replicating existing apps with AI. 

But that's not where this is going. The moment I stopped thinking of myself as a "user" and started thinking of myself as an orchestrator, everything changed. I set goals. I define constraints. I verify outcomes. That's it. The entire middle — the actual work — I don't touch it anymore.

And now I can't unsee it: most software we're building today is already obsolete. It's designed for humans to operate. But humans aren't going to be the operators anymore.

The new generation of software won't have dashboards. It won't have onboarding flows. It'll have APIs, state machines, and feedback loops built for AI agents, not for people.

The real question isn't "how do we develop this product/feature for humans using AI?" --> It's "if no human ever had to touch this, what would we build from scratch?"

if you are in a leadership position, you are punishing your team by keeping them busy.

If your team is too busy doing their "normal jobs" to spend hours a week experimenting with AI, you are failing them.

You are setting them up to be replaced by people who have the time to learn.

They will get disrupted, either by another department within the company or by a competitor, not because of AI, but because of your lack of urgency and understanding of the current landscape.

You have to give people the time, because integrating AI into daily workflows isn't something that happens overnight.

You have to actively remove work from their plates so they can learn the new way to work.

Because we are now all beginners again.

If you don't give them the time, they will never learn it, and your company will suffer a slow, painful decline into inefficiency.

No matter how big you are.

you need to clear your calendar and invest in AI implementation for your own productivity asap

The hardest part isn't the technology. It's making the time to actually sit down and try it.

Block 2 hours a day. Thank me later

Parallelism is the biggest unlock in AI. The real shift isn't speed; it's concurrency.

 I can now run 10 "hard things" simultaneously while I'm in meetings.

My workflow has flipped:
- Old way: Block time to do the work. Maybe use AI to do it faster.
- New way: Block time to review the results and give more context knowledge to the AIs.

We are moving from being "doers" to orchestrators. Your output is no longer limited by your hours, but by the number of parallel streams you can manage.

I get 100s of emails and chats daily. The only way to manage it is to stop everything, dive in, and manually filter the signal from the noise. It's inefficient and a total focus-killer.

So i built Mark, my Communications AI Agent. Its job is simple:
1/ Monitor Everything: It watches all my inboxes and chats 24/7.
2/ Surface What Matters: It identifies critical messages and ignores the rest.
3/ Notify Intelligently: I get a clean Terminal and Desktop notification with a summary file. No more context switching.

But that's just the start. The real magic is next → I'm connecting Mark to a shared knowledge base of all my documents, past conversations, and daily meeting notes.

Soon, it won't just notify me. It will draft context-aware replies for almost everything.

That's my way to solve this. I'm curious...have you developed something similar?

I moved from the EU to the US more than 10 years ago, and while I miss many things about Europe, I have to agree with the man of the hour 🦄 Peter Steinberger : the US is an incredible place for builders right now. 

There's an undeniable energy and enthusiasm here that's hard to find elsewhere. 

It feels like a special time to be building here.

Manus just shared why most people are still stuck in the "chat" era: the real value has shifted from conversational bots to custom superagents.

Chatbots answer questions; superagents execute workflows. 

But they aren't magic, you need to understand the underlying mechanics and put in the time to customize them for your specific needs. If you treat them like a standard search bar, you’re leaving 90% of the value on the table.

Here is how to actually get them to work for you:
1/ Move from Prompting to Engineering: Stop asking "Can you do X?" and start defining the architecture of the task.

2/ Invest in the Setup: Spend the time upfront to define personas, data sources, and specific tool-use constraints. The more "narrow" the agent's focus, the higher the quality.

3/ Iterative Refinement: A superagent is never "done." You have to monitor the execution and tune the instructions based on real-world failures.

Swipe through for the tips that actually matter for high-stakes execution.

Manus Intro

What’s the point of buying software when you can just create it on the fly?

It might take our generation a minute to adapt, but the next generation is growing up expecting to build exactly what they need. For them, the only thing that matters is the infrastructure to power it.

I’ve been building non-stop lately. I’ve never felt this empowered in my career. I can literally turn my ideas into reality in a matter of hours.

yesterday at the gym I heard someone pitching their new AI tool that helps you “do more, faster.” i feel like doing less and slower. am i the only one?

ok, so you moved from a chatbot to an AI agent. and you think you are ahead already.

but if you are still the one triggering all the actions...

if the agent isn't working autonomously for you all day long...

and if you don't have 10 of them running in parallel...

you are nowhere close to calling yourself AI-native. 

i recommend you spend the weekend learning. this is a good place to start: https://lnkd.in/eW4Sam_6

Every PM should understand evals. Not "be an expert." Not "build the eval framework yourself." Just understand them.

Here's why.

A/B testing changed product management forever. Before it, PMs guessed. After it, they measured. Evals are doing the same thing for AI products and systems but most PMs have heard the word but they don't know what they're talking about. 

An eval is how you answer one simple question: "Is my AI actually good?"

Not vibes. Not "it seems better." Not "the demo looked great." Actual, measurable, repeatable quality assessment.

Here's what you should tell tell your product leads:

1. Start with a golden set: a curated dataset where you know the right answers. This becomes your source of truth.

2. Define your graders: what does "good" look like? Accuracy? Tone? Factual correctness? Be specific.

3. Run your model against the golden set and score it. That's your baseline.

4. Every time you change something - new prompt, new model, new data — run evals again. Did it get better or worse?

5. Automate this. The speed of your eval loop determines how fast your AI improves.

That's it. That's the core.

The teams that ship reliable AI products aren't the ones with the best models. They're the ones with the best evals practice.

If you're a PM working on anything AI-related and you can't answer "how do we know it's good?"... start there.

Evals are the new A/B testing. And just like A/B testing, the PMs who get it early will have a massive advantage.

Manus Agents just launched on Telegram and it changes how you interact with AI agents entirely. I've been using this all weekend non-stop.

This isn't a chatbot bolted onto a messaging app. It's the full Manus — reasoning, multi-step task execution, tool use — running inside a conversation you already have open.

Here's what makes it different:
1/ Zero setup friction. Scan a QR code. You're connected in under a minute. No API tokens. No config files. No terminal.

2/ Full task execution from chat. Research, data processing, structured reports, PDFs — initiated and delivered without ever leaving the conversation.

3/ Multimodal by default. Send a voice message, a photo, a document. The agent transcribes, interprets, and acts. No manual file handling.

4/ It remembers how you work. Set your communication style — concise, structured, conversational — and it adapts across every interaction.

5/ You pick the model. Manus 1.6 Max for deep reasoning and creative work. Manus 1.6 Lite for fast, everyday tasks. Both go end-to-end.

But here's what got my attention even more — the cofounder just shared what's coming next:

1/ Create your own specialized agents and plug them into any group chat.

2/ WhatsApp, LINE, Slack, Discord support landing soon.

3/ Native Windows and Mac apps that let Manus operate your computer — like Browser Operator, but way more powerful.

And a lot more already in the shipping pipeline. Big updates dropping over the next 30 days.

Your AI agent shouldn't live behind a login screen on a separate platform. It should be wherever you already are — one message away, ready the moment you need it.

Learn more here: https://lnkd.in/gSuAx5UX

trust your AI overlord and... just hit accept 🤣

Recently, I stumbled across a YouTube video that completely stopped me in my tracks.

An urban explorer infiltrates the abandoned IBM center on the Côte d'Azur in La Gaude in South of France. Hearing the narrator describe it as a "vestige of Silicon Valley" and seeing that massive concrete colossus sitting empty in the middle of the forest is absolutely heartbreaking 💔 

For me, that place wasn't just a building. It was pure magic. It was where my professional life truly began and where my life completely transformed.

Early in my career, I had the chance to visit this unbelievable, high-energy campus while I was interning for IBM in Belgium. I knew immediately I belonged there, and I did absolutely everything in my power to get hired and eventually, I earned my spot.

When I look at the footage of those abandoned offices and empty reception halls, I don't see decay. I see endless, vibrant memories:

- Going from a nervous intern to earning my first full-time role.
- Working late into the night and battling those early-career nerves before major presentations.
- Learning a completely new language (French!).
- The incredible lunches, coffee breaks and walks around the campus with colleagues who became amazing, lifelong friends.
- Spending hours on the campus soccer & tennis courts, both competing in tournaments and playing with colleagues. I also learned to play la pétanque!
- And most importantly, it’s where I met my incredible wife.

That campus was the launchpad that gave me the opportunity to eventually relocate to the US. 

Looking back, I can draw a direct line from walking through those doors in France to my time as VP of AI Platform at IBM, and to do now the most exciting work in my career in Silicon Valley.

The doors might be closed, but the energy of what we experienced there is still very much alive. 

I will forever be grateful for that chapter, those people, and those memories.

Link to the video: https://lnkd.in/gA-Wja2s

Open video: INFILTRATION d’IBM, le plus gros site abandonné de la Côte D’Azur mais aussi le plus Surveillé !!!  on youtu.be

INFILTRATION d’IBM, le plus gros site abandonné de la Côte D’Azur mais aussi le plus Surveillé !!! 

I thought I was a heavy AI user. Then my usage 3x'd in the last 30 days. The pace of change is accelerating and what was cutting-edge a month ago is now the baseline. The new default is managing AI agents, not just using them.

This chart below is my AI usage (tokens)

Collabs: collab.shivanivirdi@gmail.com
Hi, I’m Shivani.
AI Engineer | Specialising in RAG, Context Engineering & Agentic Systems | Founder, NeoSage

I help engineers and founders understand and build production-grade LLM systems that actually work.

What I write about:
- RAG architectures, retrieval, ranking & evaluation
- Context engineering (the layer most teams overlook)
- Agentic AI & tool-using systems
- AI system design, patterns, and reasoning frameworks

🎯 Who I am
Ex-Microsoft/AWS/Adobe engineer building secure distributed systems and user-facing intelligence features. Now full-time in AI engineering education and NeoSage — where I teach the systems behind reliable, scalable LLM applications.

🚀 What I do
I break down advanced AI engineering into actionable mental models, design blueprints, and repeatable frameworks.

If you're building RAG, improving accuracy, or designing agents, my content will make you better.

🌍 Why I do it
Because AI isn’t magic, it’s systems engineering.

And the world needs more people who can bridge the gap between research and actual application.

Let’s build the future of intelligent systems together.

Shivani Virdi

I taught myself AI from scratch, got promoted at Microsoft, and quit to start my own thing. If I had to start again in 2026, this is the exact path I'd follow.

Step 1: Understand LLMs Before You Build
→ Deep Learning Specialisation by Andrew Ng 
https://lnkd.in/gwqSPS4J
→ Karpathy's Neural Networks: Zero to Hero 
https://lnkd.in/gZumxBZw
→ Karpathy's deep dive into LLMs https://lnkd.in/gcaiyVJz
→ "Attention Is All You Need" https://lnkd.in/gXUccydp
→ Stanford CS25 https://lnkd.in/gUzsiN_e
→ NeoSage Issues 1 & 2 for faster intuition
  https://lnkd.in/g2pXsyHn
  https://lnkd.in/gPimp6mv

Step 2: Prompting as a Developer Skill
→ OpenAI Prompt Guide https://lnkd.in/gqjcPxcQ
→ Anthropic Prompt Engineering Guide https://lnkd.in/gsnasAV6
→ The Prompt Lifecycle Every AI Engineer Should Know (NeoSage) https://lnkd.in/g5UJaSed

Step 3: Get Your Hands Dirty
→ Build a chatbot
→ Fine-tune a model
→ Build a RAG pipeline from scratch https://lnkd.in/gbdiAkPD
→ RAG Techniques repo by Nir Diamant https://lnkd.in/gA7maM5Y
→ GenAI Agents repo by Nir Diamant https://lnkd.in/g5HHZrJK

Step 4: Think in Systems
→ Frameworks: LangChain, LangGraph
→ Protocols: MCP (get started: https://lnkd.in/gfdJqGU4), A2A
→ Agents: planning, memory, inter-agent coordination https://lnkd.in/gFfxtN-c
→ Evaluations: if you can't measure it, you can't improve it

Books I'd Read:
→ AI Engineering by Chip Huyen https://lnkd.in/g-MRviYk
→ The LLM Engineering Handbook by Iusztin & Labonne https://lnkd.in/gyA4vFXz
→ Designing Machine Learning Systems by Chip Huyen https://lnkd.in/guN-UhXA

Learning Hubs:
→ DeepLearning.AI https://lnkd.in/gAYmJqS6
→ Anthropic Academy https://lnkd.in/gZANmHVk
→ Google x Kaggle GenAI Intensive https://lnkd.in/ga5X7tVJ
→ NVIDIA GenAI/LLM Learning Paths https://lnkd.in/gCtDnhni
→ NeoSage by yours truly blog.neosage.io

---
👋 P.S. If you want to skip the wandering and get the structured version, my AI Engineering cohort covers most of this path hands-on. 

First cohort sold out with 50+ engineers from Microsoft, Adobe, Amazon, Shopify and more.
To join the next cohort, click "𝗩𝗶𝘀𝗶𝘁 𝗺𝘆 𝘄𝗲𝗯𝘀𝗶𝘁𝗲"
---

♻️ Repost to help someone starting their AI journey.

AI Engineering | Founder @ NeoSage | ex-Microsoft • AWS • Adobe | Teaching 70K+ How to Build Production-Grade GenAI Systems

I've watched hundreds of hours of AI content. If I could keep only 5 YouTube playlists and delete the rest, these are the ones that stay.

Most people will save this post. Only 1% will actually commit to finishing even one. Be the 1%.

1. Neural Networks: Zero to Hero (Karpathy)
https://lnkd.in/gd2JZ5Wt
Karpathy walks you from micro-gradients to nanoGPT. 
Pure code-first intuition.

2. Stanford CS336: Language Modelling from Scratch (2025 edition)
https://lnkd.in/gzW4JkW9
Full-stack LLM engineering: data, training, serving, eval.

3. MIT 6.S191: Intro to Deep Learning (2025)
https://lnkd.in/gNGV2kbB
Fast-track tour of modern DL, transformers + diffusion packed into 1h 45m.

4. CS25: Introduction to Transformers w/ Karpathy
https://lnkd.in/gs4iTSpt
Turns "Attention Is All You Need" into deployable code.

5. Stanford CS230: Deep Learning (Autumn 2025)
https://lnkd.in/e_khEgxV

Some of these overlap in content, but these are some of the finest deep learning lectures you'll find.

The best playlist is the one you actually finish.
___________
P.S. My AI Engineering cohort provides a structured learning path to upskilling in AI with hands on projects.

The first cohort already sold out with engineers from Microsoft, Adobe, Amazon, Shopify and more. 

If you're looking for the guided version of this path, check out the cohort (Visit my website).
___________

♻️ Repost to help someone start their AI journey.

If I had to explain how RAG, the data layer of AI, matures in production, these are the 6 RAG patterns I'd walk through.

Each level builds on the previous.
The skill is knowing when to stop.

Level 1: Naive RAG
• Embed, chunk, retrieve, generate.
• Chunking is your only quality control.
• Without reranking, irrelevant chunks weigh the same as relevant ones.
• Retrieval gets noisier as data scales.
• Fine for prototypes. Not for production.

Level 2: Hybrid Search + Reranking
• Combine keyword search with semantic search.
• Then rerank with a cross-encoder that reads each chunk against the query.
• 48% better retrieval. 35% fewer hallucinations.
• This is where production actually starts.

Level 3: Sub-querying + Routing
• Some questions span multiple topics or sources.
• One retrieval pass won't cover it.
• Break the question into parts.
• Retrieve for each. Route to the right index. Merge.
• Extra LLM calls. Only layer on when Level 2 isn't enough.

Level 4: Graph RAG + Multimodal
• Graph RAG answers questions about connections.
• "Which suppliers share compliance violations?"
• Isolated chunks can't answer that.
• You need a knowledge graph over your data.
• 
• Multimodal handles tables, charts, diagrams.
• Text embeddings can't parse a chart.
• Each format needs its own extraction and retrieval path.
• Both cost 3-5x more. Add them when your data is relationship-heavy or visually rich.

Level 5: Corrective RAG
• A lightweight evaluator scores retrieval confidence before generating.
• High confidence: proceed normally.
• Low confidence: re-retrieve, reformulate, or search the web.
• Catches bad retrieval before it reaches the user.

Level 6: Multi-agent RAG
• Separate agents for different retrieval strategies.
• Coordinated by an orchestrator.
• Justified when each source needs a fundamentally different approach.

If Levels 2-3 solve it, this adds overhead without value.
Most production systems live at Level 2 or 3.

The patterns are a ladder, not a checklist.
Climb only as high as your problem requires.
___

P.S. My AI Engineering cohort goes into most of these patterns in much more hands-on depth. 

The first cohort already sold out with engineers from Microsoft, Adobe, Amazon, Shopify and more. 

If you're looking for hands-on structured learning, check out the cohort https://lnkd.in/gCTA2tvp
___

♻️ Repost to help someone building RAG for production.

Cursor just published how they actually use their own agent internally. 8 practices I wish I knew 6 months ago

These are the practices that separate using Cursor from actually getting results with it.

1. Plan Before Code
↳ Press Shift+Tab.
↳ Agent researches your codebase first.
↳ Asks clarifying questions.
↳ Creates a detailed plan with file paths and references.
↳ You approve before it writes anything.
↳ Plans save as editable Markdown.
↳ When output goes wrong, revert and refine the plan.

2. Rules vs Skills
↳ Rules (.cursor/rules/): always-on project context.
↳ Commands to run, patterns to follow, pointers to canonical files.
↳ Reference files instead of copying contents. They go stale otherwise.
↳ Skills (SKILL.md): load dynamically when the agent decides they're relevant.
↳ Custom commands, hooks, domain knowledge.
↳ Context window stays clean.
↳ Start simple. Add rules only when you notice repeated mistakes.

3. Hooks for Self-Correcting Loops
↳ Configure a hook that runs every time the agent stops.
↳ It checks a scratchpad file.
↳ Goal not met? Sends a follow-up and the agent continues.
↳ Cap it at N iterations.
↳ Goes from "one pass and done" to "iterate until tests pass."
↳ Same pattern works for matching UI to a mockup.

4. TDD with Agents
↳ Write tests first. Confirm they fail.
↳ Then let the agent implement.
↳ Explicitly tell it not to touch the tests.
↳ Agents perform best with a clear target to iterate against.
↳ Without tests, they have no signal for whether their code works.

5. Parallel Agents with Worktrees
↳ Cursor creates isolated git worktrees per agent.
↳ Each agent edits, builds, and tests independently.
↳ No file conflicts. Click Apply to merge back.
↳ Run the same prompt across multiple models simultaneously.
↳ Compare results side by side.
↳ Best for hard problems where models take different approaches.

6. Debug Mode
↳ Standard agent interaction guesses at fixes. Debug Mode doesn't.
↳ Generates multiple hypotheses.
↳ Instruments your code with logging.
↳ Asks you to reproduce the bug while collecting data.
↳ Then makes targeted fixes based on evidence.
↳ Systematic, not guess-and-patch.
↳ Best for race conditions, timing issues, and regressions.

7. Context Discipline
↳ Switch conversation when switching tasks/ agent repeats mistakes.
↳ Long conversations accumulate noise.
↳ Stop tagging every file manually.
↳ Agent has grep and semantic search. It finds what it needs.
↳ Use @Branch to orient it to current work.
↳ Use @Past Chats to pull earlier context cleanly.

8. Cloud Agents
↳ Describe the task. Agent clones your repo.
↳ Creates a branch. Works autonomously.
↳ Opens a PR when done.
↳ Trigger from Slack with @Cursor.
↳ Best for bug fixes mid-flow, tests, and docs you'd otherwise skip.

Full article: https://lnkd.in/gdNksnkT

👋 P.S. Looking for a structured way to upskill in AI engineering? Check out my cohort (link in comments).

♻️ Repost to help someone get more out of their Cursor setup.

Anthropic spent $10M on Super Bowl ads to throw shade at OpenAI. And everyone's talking about it.
But no one's asking the real question

A little recap on the drama :)

OpenAI added ads to ChatGPT's free tier. 
First major AI product to go ad-supported.

Anthropic saw the opening and took the perfect slapback. 
Four satirical Super Bowl spots. 

(one of them is attached, enjoy with popcorn XD)

Tagline: "Ads are coming to AI. But not to Claude."

Honestly? It's funny. Gotta give them that.

But wait there's more,
Then Sam Altman posted a 420-word rant on X.

Started with: "they are funny, and I laughed."

Then: "clearly dishonest." 
Then: "deceptive." 
Then: "authoritarian company." 
Then: "Anthropic serves an expensive product to rich people."

And that was all the fuel any media house needed.

TechCrunch: "Sam Altman got exceptionally testy."
Futurism: "Sam Altman Is Spiraling."
SF Standard: "Can OpenAI take a joke?"

But here's the real question underneath the memes.

When you ask AI about your health, finances, or business, do you want the answer optimised for you or for an advertiser?

That's what Anthropic wants you to ask.

And regardless of the Super Bowl stunt, it's the right question.

What's your take?

taplio

The Battle

Armand Ruiz

Shivani Virdi

Followers

Posts

Engagements

Likes

Performance ScoreArmand Ruiz's LinkedIn Performance

Performance ScoreShivani Virdi's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Armand Ruiz

Shivani Virdi

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Armand Ruiz's LinkedIn Performance

Performance ScoreShivani Virdi's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreArmand Ruiz's LinkedIn Performance