February 26, 2026

"Vibecoding" (氛围编程) has exploded in China as a cultural phenomenon. ByteDance launched the TRAE IDE, Tencent built CodeBuddy, Alibaba released Qoder — all racing to dominate the AI coding space. Yet Chinese devs are still looking for ways to access Claude Code. A 12-year-old named Guoguo publicly criticized the big tech companies' tools at a conference in Beijing. And a former Meituan employee with no technical background built the best-selling app on the Chinese App Store using Cursor in 1.5 hours.

The landscape

"Vibecoding" (氛围编程) has become a phenomenon in China. The term describes casual, creative programming with AI — and it's democratizing software creation.

The tools race

Chinese big tech companies are competing aggressively: ByteDance launched the TRAE IDE with both traditional and AI-native modes. Tencent built CodeBuddy, integrating its Yuanbao models. Alibaba released Qoder in August 2025. Even so, Chinese developers are still finding workarounds to access Western tools like Claude Code.

Notable cases

Guoguo, a 12-year-old, publicly criticized the big companies' coding tools at a Beijing conference in 2025, showcasing the democratization of development. "Peanut," a former Meituan employee with no programming background, built the "Little Kitten Colored Lightbox" — a ¥1 app that became the best-selling app on the Chinese App Store in December 2024, built entirely with Cursor in just 1.5 hours.

Implication

Vibecoding is lowering the barriers to software creation, letting non-programmers build functional apps and reshaping how coding tools are developed in China.

Peter Steinberger, creator of the viral AI agent OpenClaw, shared advice on the OpenAI Builders Unscripted podcast. His core message: AI builders should take a more "playful" approach and give themselves time to improve gradually. He champions experimentation over perfectionism, creative energy through play, iterative improvement as a natural process, and learning by doing rather than extensive planning.

Who he is

Peter Steinberger is the creator of OpenClaw, an AI agent that went viral. He talked about his development philosophy on the OpenAI Builders Unscripted podcast.

The message

AI builders should be more playful — light, experimental. Instead of chasing perfection right out of the gate, the path is to iterate with a light touch. Rigidity and excessive seriousness stall the creative process.

Practical principles

• Experimentation > Perfectionism
• Keep creative energy alive through playfulness
• Accept gradual improvement as natural
• Learn by doing, not by over-planning

Anthropic published a comprehensive guide to building "skills" — reusable workflows that teach Claude to perform specific tasks consistently. The target audience: devs looking for consistency, MCP connector builders, power users automating tasks, and organizations standardizing operations. It covers skill architecture, standalone and MCP-based patterns, testing methodologies, and distribution strategies. Estimated time: 15-30 minutes for your first working skill.

What Skills are

Skills are reusable workflows that teach Claude to follow specific processes consistently. They're like recipes Claude follows — including tools, steps, and response patterns.

Who they're for

• Devs who want consistency across workflows
• MCP connector builders
• Power users automating tasks
• Organizations standardizing Claude operations across teams

What the guide covers

Technical requirements and architecture best practices. Standalone and MCP-integrated skill patterns. Testing and iteration methodologies. Distribution strategies. Estimated time for your first working skill: 15-30 minutes if you already know the 2-3 workflows you want to automate.

Anthropic released enhanced web search tools that use code execution to filter results before they enter the context window. Claude now automatically writes and runs code to post-process results. On the BrowseComp benchmark, Sonnet climbed from 33.3% to 46.6% and Opus from 45.3% to 61.6%. On DeepsearchQA, Sonnet's F1 score went from 52.6% to 59.4% and Opus from 69.8% to 77.3%. On average: +11% accuracy while using 24% fewer input tokens.

The problem

Traditional web search pulls full HTML into the context, generating irrelevant data that degrades response quality and wastes tokens.

The solution

Claude now automatically writes and runs code to post-process search results, keeping only the relevant information and discarding the rest.

Results

BrowseComp (finding hard-to-locate information): Sonnet 4.6 climbed from 33.3% to 46.6%; Opus 4.6 from 45.3% to 61.6%. DeepsearchQA (systematic answers): Sonnet's F1 from 52.6% to 59.4%; Opus from 69.8% to 77.3%. Overall average: +11% accuracy with -24% tokens consumed.

Available tools

Code execution, persistent memory storage, programmatic tool calling, dynamic tool discovery.

Claude Code on desktop now integrates the full development workflow. Live App Preview: it launches servers and shows the app running inside the interface, monitoring logs and iterating automatically. Automated Code Review: a "Review code" button examines local changes with inline comments on the diff. GitHub PR Management: it tracks PR status, CI checks, auto-fixes failures, and auto-merges. Cross-platform continuity: sessions move between CLI, desktop, and web/mobile without interruption.

Live App Preview

Claude Code launches development servers and displays the running app right inside the desktop interface. It monitors console logs, spots errors, and iterates automatically. Users can visually select UI elements to give refinement feedback.

Automated Code Review

The "Review code" button has Claude examine local code changes and generate inline comments in the diff view. Immediate feedback on potential bugs and improvements before committing.

GitHub PR management

Tracks pull request status, CI check results, and failures using the GitHub CLI. Optional auto-fix corrects CI failures automatically. Auto-merge completes PRs when checks pass — all in the background.

Cross-platform continuity

Sessions move from CLI → desktop via /desktop, or desktop → web/mobile via "Continue with Claude Code on the web." Available immediately for all users.

Anthropic announced significant updates to Cowork and the plugin ecosystem for enterprise. A unified "Customize" menu consolidating plugins, skills, and connectors. Admins can build plugins from templates or from scratch. Private plugin marketplace and per-user provisioning. Slash commands with structured forms. OpenTelemetry support for tracking usage, costs, and tool activity. New connectors: Google Workspace, DocuSign, Apollo, Clay.

Plugin management

A new unified "Customize" menu consolidates plugins, skills, and connectors. Admins build plugins from templates or from scratch, with Claude guiding the setup. Organizations can create private marketplaces and manage per-user provisioning.

User experience

Slash commands now open structured forms. Company branding appears throughout Cowork, including a redesigned home. OpenTelemetry support for tracking usage, costs, and tool activity per team.

New integrations

Google Workspace connectors (Calendar, Drive, Gmail), DocuSign, Apollo, Clay. Partners like Slack, LSEG, and S&P Global built plugins for shared customers.

Anthropic released finance-focused updates. Claude now moves between Excel and PowerPoint without losing context. 5 new plugins: financial analysis, investment banking, equity research, private equity, and wealth management. Two institutional data connectors: FactSet and MSCI. Partners LSEG and S&P Global launched complementary plugins.

Cross-app workflow

Claude moves between Excel and PowerPoint without losing context. Finance professionals complete multi-step tasks in a single session.

5 specialized plugins

• Financial analysis: market research, modeling, presentations
• Investment banking: deal review, comparable analyses, pitch materials
• Equity research: earnings analysis, model updates, research notes
• Private equity: due diligence, scenario modeling, opportunity scoring
• Wealth management: portfolio analysis, drift identification, rebalancing

Data connectors

FactSet: real-time market data and research. MSCI: proprietary index information. LSEG and S&P Global launched complementary plugins.

"Intelligence Yield" (IY) is a metric that measures compute efficiency: (p50 time horizon × average score) / work time. Opus 4.6 shows the highest IY: it solves complex tasks (p50: 870 min) with strong performance (78.86%) using only ~228 min of work. OpenAI models (o3, GPT-5.2) reach impressive horizons but with massive compute investment, resulting in a lower IY.

The metric

Intelligence Yield (IY) measures a model's compute efficiency: (p50 time horizon × average score) / work time. It reframes progress from "how hard can it solve" to "how intelligently does it use resources."

Opus 4.6 in the lead

Opus 4.6 achieves the highest IY: it solves complex problems (p50: 870 minutes) with a score of 78.86%, using only ~228 minutes of work time.

Comparison with OpenAI

OpenAI's models (o3, GPT-5.2) reach impressive time horizons, but they demand massive compute investment — 326k+ and 512k+ seconds — resulting in a significantly lower IY.

Trend

Data from 2023 to 2026 shows that Anthropic's models consistently optimize the compute-to-capability ratio.

A practical guide from You.com for organizations to discover and implement AI use cases. The core argument: successful transformation requires identifying specific business challenges rather than chasing vague technological capabilities. It splits into two areas — internal (back-office automation, workflows, knowledge management) and external (customer-facing products, support, personalization). It emphasizes organizational alignment.

The problem

Organizations often frame AI projects around "shiny" capabilities instead of delivering measurable value. That leads to wasted investment.

Two implementation areas

Internal: operations automation, employee workflows, knowledge management. External: customer-facing products, support systems, personalization.

Critical factor

Organizational alignment. Unified discovery ensures the company doesn't miss high-impact opportunities and enables scalable, repeatable processes.

Wealthsimple is recruiting a select team of AI builders for full-time positions. The twist: candidates submit working prototypes instead of traditional résumés. They're looking for people who "look at a broken process and can't help imagining how it should be." Remote-first across North America. Applications until March 2, 2026, offers by March 7.

What it is

Wealthsimple's recruiting program for AI builders. Full-time, paid positions with competitive compensation.

Selection process

No résumé. Candidates must design and prototype an AI system that expands human capability. They can use AI tools during development.

Who they're looking for

People who look at a broken process and can't help imagining how it should be. They move fast under uncertainty and think critically about where AI should and shouldn't step in.

Details

Remote-first across North America. 18+, work authorization required. Applications until March 2, 2026, offers by March 7.

OpenAI's Codex team has ~40 people with minimal hierarchy: 1 PM, 2 designers, and various engineers. Teams of 2-3 people own features end-to-end. Codex is used internally for everything: in one bug bash they processed 100+ issues in 1 hour, most fixed within 24h. It works as an "onboarding buddy" — new engineers ship meaningful features on day one. Automatic PR review with custom instructions.

Team structure

~40 people. 1 PM, 2 designers, engineers with varied expertise. Minimal hierarchy. Teams of 2-3 people own features end-to-end.

AI baked into the workflow

Codex is used at scale internally. In one bug bash, they processed 100+ issues in 1 hour. It works as an "onboarding buddy" for new engineers — setup and codebase comprehension are automated. Engineers ship meaningful features on day one.

Culture

Meetings kept to a minimum. Spontaneous collaboration and accessible leadership. Decisions prioritize speed — Codex enables fast course correction.

Hiring

Two extremes: exceptional generalists who multiply the team's effort, and specialists who think outside the box.

Martin Fowler shares insights from the Thoughtworks retreat on the future of development with LLMs. Senior devs keep their value by focusing on architecture. Hands-on exercises converted 1/3 of LLM-resistant senior devs into advocates. Mid-level developers face the biggest challenge. "Cognitive debt" (Margaret-Anne Storey): teams lose the ability to explain design decisions. Practices that improve developer experience also improve LLM effectiveness. Executives now prioritize smoothing the path for agents when they never did it for humans.

Impact by seniority level

Senior devs keep their value by focusing on architecture. Hands-on exercises converted 1/3 of the resisters into advocates. Mid-level devs face the biggest challenge: they built careers without LLMs and lack the senior expertise to leverage them.

Cognitive debt

Margaret-Anne Storey introduced the concept: when no one on the team can explain why certain design decisions were made. Fowler compares it to technical debt — teams pay interest through slower development.

DevEx = AgentEx

Laura Tacho observed that practices improving developer experience simultaneously improve LLM effectiveness. The irony: executives now prioritize smoothing the path for agents when they neglected to do it for humans.

Concerns

AI adoption accelerates the pace of work and expands task scope, potentially causing burnout. IDEs must orchestrate between deterministic features and LLM capabilities.

Margaret-Anne Storey argues that teams using generative AI face a bigger threat from "cognitive debt" than from technical debt. Even when AI produces clean code, devs lose track of design intent and system logic. A real case: in an entrepreneurship course, a team got stuck in weeks 7-8 — no one could explain why certain design decisions had been made.

The concept

Technical debt lives in the code. Cognitive debt lives in the developers' minds. It's the loss of shared understanding of what the software does and how it works.

Real case

In an entrepreneurship course, a team of students got stuck in weeks 7-8. The real problem: no one on the team could explain why certain design decisions had been made.

Recommendations

• Recognize speed that's unsustainable without understanding
• Require human verification of AI changes
• Document decision rationales, not just the code
• Regular knowledge-sharing sessions

Warning signs

Hesitation to modify existing code. Reliance on tribal knowledge. Treating the system as a black box.

Boaz Barak, an OpenAI alignment researcher, analyzes Anthropic's Claude constitution. He identifies three poles of alignment: general principles, operational policies, and personality. He argues the constitution leans toward personality while minimizing policies. He praises the emphasis on preventing authoritarian control and the honesty standards. But he questions the anthropomorphization, the self-determination granted to Claude, and the apparent reluctance to have clear rules.

What it is

Boaz Barak, an OpenAI alignment researcher, gives a detailed analysis of the constitution Anthropic published for Claude, comparing it with OpenAI's Model Spec.

Three poles of alignment

General principles (ethical axioms), operational policies (specific rules), and personality (empathy/character). Barak argues the constitution leans toward personality while minimizing policies.

Strengths

Emphasis on preventing authoritarian control by humans. Honesty standards that exceed typical human ethics.

Concerns

Risks of anthropomorphization. The document grants Claude the authority to seek "true universal ethics" — Barak finds this surprisingly permissive. He advocates for explicit rules and human oversight rather than trusting the AI's moral reasoning.

Greg Brockman, OpenAI's president, describes a transformative shift in software engineering. Since December, tools like Codex have reached a "step function improvement." Engineers report that Codex now "writes essentially all the code" and assists with operations and debugging. Goal by March 31: agents becoming the primary tool for technical tasks. The 6 recommendations include appointing an "agents captain," documenting in AGENTS.md, and keeping human accountability on merges.

The thesis

Greg Brockman claims there's been a "step function improvement" in AI coding tools since December. Engineers at OpenAI report that their jobs have fundamentally changed.

Strategic goal

By March 31: agents becoming the primary tool for technical tasks, replacing traditional editors and terminals.

6 recommendations for teams

• Set aside time to experiment; appoint an "agents captain"
• Create and maintain AGENTS.md files tracking performance
• Inventory internal tools and make them agent-compatible
• Structure codebases with agents in mind — fast tests and quality interfaces
• Keep human accountability for merged code
• Build observability and tracking tools