Marcelo Skaba

This is wild. red.anthropic.com/2026/mythos-pr…

I love working with Claude Code, but today it got stuck looping on a problem. So I stopped it and tried a simple prompt: “Are you having problems? Why don't you take a step back, analyze, plan, and then execute?” It actually did exactly that — stepped back, made a plan, and solved the issue in one shot. Good reminder: the AI "genius" sometimes get stuck like a junior engineer.

Lovely seeing PMs/TAMs/Sales folks using Claude Code. Why? Because the context lives in GitHub. When they realize code is the SSOT, their faces light up — and the fear evaporates. Add MCPs to Jira/Kayako/Notion and suddenly everyone can drive with real context, not vibes.

“Agentic AI adoption is blocked by data availability.” — Salesforce Of course they’d say that. It is convenient, coming from a data hoarder. Agentic AI fails because of poorly manage context: goals, constraints, recency, priority. Data is fuel. Context is steering. No steering = crash.

Professional services engineer asked if a legacy API was still supported—I fed the question to Claude Code with my brainlift and code map, and it surfaced the deprecation analysis plus a step-by-step migration guide in 20 seconds. AI doesn't replace expertise; it scales it. #AI #CloudArchitecture

RIP @windsurf 💀 After the latest planning mode changes, it’s basically dead. It ignores instructions, rewrites code I didn’t ask for, and produces gibberish. What used to be great is now unusable. By contrast, Claude Code delivered: I used continuous work across several days, comparing multiple branches, running endless analysis. Task planning was flawless, caching kept costs extremely low, and the output was excellent.

Does anyone know how to disable planning mode from @windsurf ? It is on by default now and causing some pain.

The takeaway: LLMs aren’t oracles — they’re echo chambers with compression. If you want bias, fine. If you want contrarianism, prime them with counter-narratives or antibias constraints. Next time you prompt for strategy: * Decide if you want consensus or disruption * Prime with examples of the style you want * Ban default answers you already know Otherwise… you’re building yesterday’s architecture with tomorrow’s tools.

Same thing happened with other SPOVs Claude repeated: “Enterprise AdTech should consolidate services rather than embrace microservices.” “Hybrid OLTP/OLAP is superior to cloud-native data lakes.” These aren’t bugs — they’re mirrors. If you don’t define a new direction in your prompt or in intermediate reasoning, the model will optimize toward dominant thought patterns it’s absorbed. That’s great if you want to know what most experts think. It’s risky if you want to disrupt what most experts think.

I’ve been testing Claude Code to extract for spiky points of view (SPOVs) on a product that use a data center. I let it search the web, wikis, and code. Every run, it keeps spitting out: “On-Premise Deployment is a Strategic Advantage, Not Legacy Technical Debt.” Is this a crowd consensus? Maybe not, but it might be that most people still think like that. And when that mindset is encoded into model weights, it will surface unless you explicitly push the model away from it in your prompts or reasoning scaffolds (“brainlifts”).

Your LLM is your new RCA weapon—no extra tools needed: 1️⃣ Tech debt: Prompt your LLM to analyze diffs & flag risky legacy code 2️⃣ Signal debt: Feed logs & let the model surface missing alerts 3️⃣ Process debt: Chain-of-thought prompts route incidents to the right stakeholder Cut the fluff—let your LLM close the debt loop. #LLM #RCA #DevOps

Stop dancing around failures. A killer RCA digs into 3 debts: 1⃣ Tech debt that broke the tool 2⃣ Signal debt—alarm was there, we didn’t hear it 3⃣ Process debt—wrong person, wrong time Dig deep, fix these, and banish repeat SEVs for good. 🚀 #RCA #DevOps #IncidentMgmt

@jczapovics Thats awesome. I wish we could use claude4 in windsurf teams subscription.

That's a jump. Now it includes both Cascade and Tab. 70% of this is Cascade, that brings more productivity improvements compared to Tab, moving step by step toward a more agentic future. Cascade has a level of abstraction and control that static step-by-step methods can't reach. 78% of Cascade reasoning model usage is Claude. @windsurf_ai

Teaching an AI to replace myself has been far tougher than I imagined—crafting a robust “brainlift” is no small feat. Have you tried building AI to take over your own role? I’d love to compare notes. #AI #ProductDefense #CTOthread

CEOs and CPOs can converse with the Assistant via @ephor's MCP interface to understand the rationale behind every recommendation—and debate priorities just as they would with me. That transparency is critical when AI steps into a CTO’s shoes.

As CTO, I’m charting our top product priorities—and defining our Product Defense Plan: proactive strategies to safeguard product health before issues ever surface. To operationalize this, I’m building an AI-driven Assistant with direct access to our codebase and data streams feeding four specialized agents.