From "Vibe Coding" to "Software 3.0": How to Turn Claude into a Development, Research, and Execution Team Running on Natural Language
When discussing Claude's utilization in 2026, one cannot overlook Andrej Karpathy. As a founding member of OpenAI, the former Director of AI at Tesla who led the Autopilot vision team, and an educator who influenced countless engineers through CS231n and LLM lectures, Karpathy is a pivotal figure. His profile explicitly mentions his involvement with OpenAI, Tesla, and Stanford.
Furthermore, in May 2026, Karpathy joined Anthropic. According to Reuters, he joined the pretraining team that shapes the foundational capabilities of Claude models. Therefore, using Claude "the Karpathy way" isn't just a prompt technique borrowing a famous name; it's a philosophy of treating LLMs as a new type of computer, OS, and development environment, connected to Claude's latest features.
Note: The "Karpathy version" of Claude discussed here is not an official mode distributed by Anthropic or Karpathy himself. It is a practical framework reconstructed from his public lectures, notes on Claude Coding, the Software 3.0 philosophy, and the latest official Anthropic information.
1. The Essence of Karpathy-style Claude: "Build with AI," Not "Ask AI"
Most people use Claude for text generation, summarization, research, or emails. While useful, Karpathy-style utilization goes much further.
In his 2025 lecture "Software Is Changing Again," he discussed how the map of software has shifted. Software 1.0 is code written by humans; Software 2.0 is neural network weights; and Software 3.0 is prompts for LLMs, where natural language is becoming the new programming language.
Applying this to Claude, it is not a "chat partner." Claude is a new execution environment that runs on English or Japanese.
It writes code, writes tests, reads specifications, uses browsers, edits files, connects to Slack/Google Drive/Jira, delegates to sub-agents, saves outputs as Artifacts, and remembers the context of entire projects.
By July 2026, Claude has evolved from an AI that answers questions to an AI that advances work. The Karpathy-style approach is to embrace this shift head-on.
2. The Latest Claude Map as of July 2026
Choosing a Claude model isn't just about "which is smartest." Claude Sonnet 5, Opus 4.8, Fable 5, Mythos 5, and Haiku 4.5 each serve different roles.
Anthropic announced Claude Sonnet 5 on June 30, 2026. It is the default for Free and Pro users and is available across Max, Team, Enterprise, Claude Code, and the Platform. Pricing is set at $3/$15 per million tokens (input/output) after an initial introductory rate.
Sonnet 5's key strengths are its price-performance ratio and agentic performance. Anthropic describes it as the "most agentic Sonnet," capable of autonomous tool use, planning, and terminal work at a level that previously required much larger models.
Claude Opus 4.8 is positioned as the high-end model for complex agentic coding and enterprise tasks, emphasizing honesty, alignment, and reduced hallucinations in long-duration tasks.
In June 2026, Claude Fable 5 and Mythos 5 also debuted. Fable 5 is the highest-performing general model for software engineering and scientific research. Mythos 5 is a limited-release model via Project Glasswing, focused on defensive cybersecurity.
For Karpathy-style resource allocation:
- Routine and most development: Sonnet 5.
- Heavy design and complex reviews: Opus 4.8.
- High-difficulty, long-duration tasks: Fable 5.
- Defensive cybersecurity: Mythos 5.
- Fast classification and sub-agents: Haiku 4.5.
3. Karpathy's Turning Point: From "80% Hand-written" to "80% Agentic"
In January 2026, Karpathy noted that his workflow shifted drastically. Previously, he did most coding manually with some AI completion; by late 2025, agents did the bulk of the coding while he focused on editing and finishing.
The essence isn't that "AI can write code," but that the human role has changed. Humans define goals, constraints, and success conditions. We read the implementation, prune bad abstractions, and judge design beauty. We stop Claude if it goes off-track and decompose problems when it gets stuck.
Karpathy warns that because agents still make mistakes, we must watch them "like a hawk." Errors have shifted from simple syntax bugs to subtle conceptual misses and over-abstractions.
4. Drive Claude with "Success Conditions," Not "Commands"
AI agents excel at looping toward a concrete goal. Instead of micro-managing steps, provide success conditions and have it write tests to verify them.
Bad: "Fix this bug."
Better: "Login redirect is failing. Find the cause and fix it."
Karpathy-style:
- Goal: Ensure users return to the original destination after login.
- Success Conditions: Existing cases don't break; unauthorized access redirects back after login; no open redirects; unit tests added; iterate until tests pass.
- Process: Explore files -> Plan -> Write tests first -> Implement -> Run tests -> Report risks.
5. Claude Code: The Center of the Karpathy Workflow
Claude Code is a development execution environment connected to terminals, IDEs, Git, and MCP. The human sets the goal, and Claude explores, plans, tests, and implements.
The human's job is to read the code and remove "slop"—AI-generated code that looks functional but is low quality. Reading skills become more important than writing skills.
6. CLAUDE.md: The Project Constitution
Since Claude Code sessions reset context, CLAUDE.md serves as the project's constitution. It should contain design philosophies, prohibitions, quality standards, and review perspectives rather than just snippets.
7. Use XML Tags for Context
Structure your prompts like Software 3.0 code. Use <task>, <context>, <success_criteria>, and <constraints> tags to eliminate ambiguity.
8. "Vibe Coding" is Not Lazy Coding
Vibe Coding means developing by conveying the "vibe" or intent in natural language. However, this requires stricter human oversight. As generation is delegated to Claude, the human's ability to identify quality and design flaws must be sharpened.
9. The Value is "Expansion," Not Just "Time-Saving"
Claude allows you to build things you previously wouldn't have bothered with: internal dashboards, automated reports, or custom learning apps. The true leverage of AI is expanding your range of action.
10. The Triangle: Claude Code + IDE + Human
Karpathy suggests keeping Claude Code on one side and a large IDE on the other. Claude is the worker; the IDE is the inspection room; the human is the supervisor.
11. Basic Claude Code Workflow
- Explore: Understand the repo without changing code.
- Plan: Propose changes and risks.
- Test-First: Write tests that fail initially.
- Iterate: Implement until tests, lint, and builds pass.
12. Mandatory "Simplification Review"
LLMs tend to over-engineer. Always ask: "Can this be simpler?" A 1000-line complex structure can often be reduced to 100 clear lines.
13. Team Up with Sub-agents
Divide roles: a Security Reviewer agent, a Simplifier agent, a Documentation agent. Separating the "creator" from the "critic" improves quality.
14. Skills: The Company Manual
Use Skills to store reusable instructions for specific tasks like "Enterprise Proposals" or "Code Reviews," evolving Claude from a personal assistant to a business foundation.
15. Connect to Reality with MCP
Model Context Protocol (MCP) allows Claude to see Jira issues, read Slack, and check Sentry logs. It turns Claude into an agent that performs "investigative work" in the real world.
16. Claude Projects: The Workspace
Projects act as context-rich rooms. Upload company profiles, personas, and brand tones to ensure Claude always works within your specific context.
17. Artifacts: Focus on Deliverables
Don't just ask for answers; ask for Artifacts—interactive dashboards, prototypes, or structured documents that serve as actual tools.
18. File Creation and Editing
Use Claude to build financial models in Excel or structured reports. It's about creating intermediate assets for human decision-making.
19. Claude Tag and Connectors
With Claude Tag in Slack and Google Workspace Connectors, Claude becomes a teammate you can delegate tasks to directly within your communication channels.
20. Claude Science and Design
Specialized workbenches for scientists and designers allow for the creation of auditable research artifacts and interactive design adjustments.
21. The 7 Principles of Karpathy-style Claude
- Treat it as a computer, not a chat box.
- Give success conditions, not commands.
- Put tests first.
- Shift from writer to reader.
- Prune over-abstractions.
- Build the environment (CLAUDE.md, MCP).
- Expand your scope of action.
Conclusion: A Design Philosophy for Working with AI
Karpathy-style Claude usage is not about "writing with AI" but "designing an AI workforce." It is the democratization of Software 3.0. Don't just use Claude; work with it.
