Claude Code's Creator Just Revealed How Coding Will Change in the Next 6 Months
Boris, one of the engineers who built Claude Code at Anthropic, just dropped some serious insights about where AI coding is headed—and it's way more hands-off than you might expect.
In a new interview, Boris explained how coding has completely transformed in just the past year. We've gone from copying and pasting code between ChatGPT and your IDE to having AI agents that autonomously write entire applications. "A year ago, you directly manipulated text in an IDE," Boris said. "Now when you code, you use an agent—you don't directly manipulate text anymore."
How They Actually Test New Models
Here's the wild part: instead of running formal benchmarks, Boris evaluates new Claude models by simply... doing his regular work. "My perfect day is I'm just coding all day," he explained. "I'll just code using that and see what the vibe is."
Anthropic has tried building sophisticated evaluation systems, but Boris admits "it's just so hard to build evals. By far the biggest signal is just the vibes. Like, does it feel smarter?"
Claude Code's Secret Sauce: Hackability
Claude Code was designed to be as "hackable" as possible, with multiple ways to extend it:
- CLAUDE.md files that give Claude context about your codebase.
- MCP (Model Context Protocol) for connecting external tools.
- Slash commands for reusable workflows (Boris has one for making git commits).
- Subagents with forked context windows for complex tasks.
The Future: Goals Over Tasks
Looking ahead 6-12 months, Boris predicts coding will shift toward giving Claude higher-level goals rather than specific tasks. "Maybe Claude will be more about goals and these higher level things it needs to do, and less about the specific tasks that go into it."
How to Prepare for This Future
Boris's advice for engineers? Learn the fundamentals, but start thinking bigger. "You still have to learn the craft—languages, compilers, system design. But also just start to get more creative. If you have an idea for a startup or product, you can just build it now."
Pro Tips from the Creator
Boris shared his personal Claude Code workflow:
- Easy tasks: Tag @Claude on GitHub issues and let it write the PR.
- Medium tasks: Start in plan mode, align on approach, then auto-accept.
- Hard tasks: Stay in the driver's seat with Claude as a pairing partner.
His biggest tip for newcomers? "Don't use it to write code yet. Use it to ask questions about the codebase first."
Oh, and fun fact: Boris' coding journey began by programming a TI-83 Plus calculator in BASIC to cheat on middle school math tests, an early example of using code to automate tasks.
Our favorite moments from the episode
If you want to jump straight to some of our favorite moments from the episode, use the timecodes below!
- Insight: The last year has seen a fundamental shift in AI-assisted coding, moving from simple autocomplete and copy-pasting code snippets to using AI agents as a core part of the inner development loop. (0:57)
- Prediction: The future trajectory of software development involves engineers doing less direct text manipulation and instead directing models to perform those manipulations for them. (1:57)
- Insight: Agentic coding failed in the past for two main reasons: the AI models weren't good enough, and the "scaffolding" (the tools and systems built around the models) was also insufficient. (2:38)
- Story: When the very early, imperfect version of Claude Code was given to the internal team, engineers began using it for their daily work almost immediately, signaling its inherent value even before it was polished. (3:16)
- Analogy: The AI model is like a horse, and the tool built around it (like Claude Code) is the "harness" or "saddle." This harness—which includes the system prompt, context management, and tools—is crucial for steering the model effectively. (4:02)
- Insight: AI models and the products built on them "co-evolve" organically. Researchers building the models use the tools (like Claude Code) daily, discover their limitations, and incorporate those learnings back into training the next generation of models. (5:05)
- Insight: A key metric of model improvement is the length of time it can operate autonomously on a task before needing human course correction. This duration is steadily increasing with newer models. (6:20)
- Point of View: The best way to evaluate a new coding model isn't through synthetic benchmarks but through a "vibe check"—simply using it for your actual, day-to-day work to see if it feels more capable and helpful. (6:51)
- Insight: Synthetic benchmarks like SWE-bench are becoming less useful as models improve; the most valuable signal for model quality comes from real-world usage and "vibes," even if it's hard to quantify. (8:37)
- Actionable Takeaway: To create a powerful product feedback loop, respond to user feedback and fix bugs as fast as possible, commenting back to users to show their input is valued. This encourages a continuous firehose of high-quality feedback. (9:37)
- Product Philosophy: Claude Code was designed from the start to be as simple and "hackable" as possible, with an increasing number of extension points like CLAUDE.md, hooks, and user-defined slash commands. (10:47)
- Forecast (6-12 months): The software engineer's role will become a mix of hands-on work (using AI to manipulate text) and a more supervisory role, where they review proactive work done by AI and set high-level goals rather than executing individual tasks. (13:27)
- Prediction: AI coding agents are making software development dramatically more accessible, shifting the focus from mastering complex technical stacks to the quality of the idea itself. (16:11)
- Insight: As AI handles more implementation, the code itself becomes less precious and more disposable. It can be rewritten constantly, allowing for faster iteration. (16:34)
- Actionable Takeaway (How to Upskill): Engineers should still learn the fundamental craft of coding (languages, system design), but they must now also focus on being more creative and action-oriented, as they can now build and test ideas almost instantly. (17:00)
- Actionable Takeaway (Tip for New Users): When first using an agent like Claude Code, don't start by asking it to write code. Instead, use it to ask questions about the codebase to build understanding and trust in the agent's capabilities first. (18:04)
- Actionable Takeaway (Mental Model): Categorize your coding tasks to use the agent most effectively: delegate easy tasks completely (e.g., via a GitHub comment), collaborate on a plan for medium tasks, and for hard tasks, you remain in the driver's seat using the AI as a powerful pair-programmer. (18:54)
Why This Matters
Boris's insights suggest that within months, successful engineers will be those who can effectively collaborate with AI agents rather than just prompt them. As Boris put it: "The code itself is no longer precious."
