Intent-Driven Development: Claude Code's Evolving Paradigm and Practical Tips

In today's explosion of AI coding tools, Claude Code has become a focal point for developers worldwide. Anthropic engineer Lydia (known online as Avocoder) and YK Sugi — creator of the CS Dojo channel and the 8,000+ star "Claude Code Tips" repository — joined the Agent Factory podcast for a deep dive into how Claude Code is actually used and where it's headed. The Intent-Driven Development framework they introduced is redefining how software engineers work.

What Is Intent-Driven Development

Intent-Driven Development isn't about crafting the perfect prompt — it's about knowing clearly what you want to build and communicating that accurately. As YK put it: "The point isn't to find some magic prompt for better results. It's about knowing exactly what you want to do and conveying that intent precisely to the model."

One counterintuitive tip: use voice input instead of typing. YK found that when typing, people tend to second-guess themselves — "Should I phrase this differently?" — while speaking out loud for a brainstorm naturally and quickly conveys genuine intent. Even if voice transcription has occasional errors, as long as the intent is clear, Claude is smart enough to understand.

This has a basis in cognitive science. Voice input's efficiency relates to the fluency of "verbal thinking" — humans naturally speak in units of meaning, while typing triggers more metacognitive monitoring, fragmenting the thought process. This aligns with the "problem space mapping" theory from programming psychology researchers: great programmers complete a full mental model of the problem space before writing a single line of code, and speech more naturally externalizes that inner model.

Lydia echoed this from a product design angle: Claude's outputs require user validation, which is itself an extension of "intent." People with software engineering backgrounds tend to write better prompts because they can organize content more clearly and articulate architectural requirements. Plan mode combined with image-based validation are both effective components of expressing intent.

A Full Demo: From Idea to Deployment

During the show, YK live-demonstrated how to build a 3D slingshot game from scratch using an intent-driven approach. He gave only a vague description: "Make a slingshot game — the slingshot is fixed in front of the camera, you drag the ball with the mouse to control X/Y direction, hold duration determines Z-axis force, and the farther you pull before releasing, the faster it flies."

Notable was his collaboration rhythm with the model — starting with broad architectural questions (what library, what framework?) and gradually drilling into details. After Claude suggested Three.js with a physics engine, YK worked through several rounds of conversation before settling on Three.js + Rapier.

Three.js is the most mature WebGL wrapper library available, created by Ricardo Cabello (Mr.doob) in 2010. It abstracts WebGL's low-level API so developers can build 3D scenes declaratively. Rapier is a physics engine written in Rust by the Dimforge team, compiled to WebAssembly. Compared to the older Cannon.js and Ammo.js, Rapier offers significant advantages in performance and precision — particularly for rigid body collision simulation, making it ideal for slingshot trajectory physics. Together, they represent the mainstream tech stack for modern browser-based 3D game development.

The demo was sprinkled with practical workflow habits:

• Set command aliases: Configure an alias for the claude command to reduce repetitive typing
• Centralized project management: Keep projects in the same folder so agents can reference components across projects
• Proficiency with Git and GitHub CLI: Starting from git and gh commands lets you automate a large number of repository operations
• Output validation: Test application behavior directly, or automate validation through tests and GitHub Actions

Throughout the process, YK communicated with the agent entirely by voice. Claude even proactively ran tests via Playwright and used terminal open commands to automatically launch the project — many operations the user never had to trigger manually.

Claude Design and Dynamic Workflows

Lydia then took over the demo, showcasing two new features currently in research preview. She pulled YK's game repo with the goal of transforming it into a "complete game" with multiple levels.

Claude Design: A Visual Planning Tool

Claude Design is a visual planning tool. Users can generate presentation decks, slide designs, and full website wireframes via prompt — currently rendered as HTML, with export support for Figma or Canva. Compared to describing things in plain text, this visual approach let Lydia systematically map out the overall design direction before writing a single line of code — from the main menu and level selection to the full game loop, including haptic feedback, charge-to-aim mechanics, and a scoring system.

Dynamic Workflows: Deterministic Task Orchestration

Even more noteworthy is Dynamic Workflows. Sub-agents themselves aren't new — Claude Code automatically spawns sub-agents with independent contexts to handle specific tasks. But the problem is unpredictability: sometimes it generates 4, sometimes 10, sometimes none at all.

Workflows are Anthropic's deterministic solution, capable of orchestrating hundreds of sub-agents and designed for complex tasks that span hours or even days. The core principle is "Deterministic Orchestration" — by encoding the workflow as a JavaScript file, Anthropic essentially separates the AI's decision layer (what to do) from its execution layer (how to do it). The JS file defines task dependencies as a directed acyclic graph (DAG), and individual sub-agents execute specific tasks at defined nodes. This is conceptually aligned with mature workflow orchestration frameworks like Apache Airflow and Temporal — the key difference is that the execution nodes are AI agents rather than traditional functions.

Claude Code generates a JavaScript file describing the entire process. In the demo, the build phase ran four parallel sub-agents simultaneously — building the engine, UI, audio, and tool levels with haptic integration — followed by a review and validation phase.

This workflow can be saved via the / command and re-run deterministically, calling the exact same sub-agent configuration every time. Because sub-agents run in parallel, they typically accomplish more work in the same time it would take a single sub-agent. Users can also directly edit the generated JS file to assign different models to different sub-agents — Opus for heavy tasks, Sonnet for lighter ones.

Auto Mode: Balancing Safety and Efficiency

Lydia highlighted a feature often overlooked by developers — Auto Mode.

Traditional Ask mode requires user confirmation at nearly every step. Over time, this creates "confirmation fatigue": Claude asks a hundred times, and the user stops reading and just approves everything — which is actually more dangerous. The opposite extreme, "skip all confirmations," can be catastrophic — imagine accidentally deleting root files with no way to recover.

Auto Mode sits between these two extremes. It runs a classifier between tool calls that first assesses the risk level of each operation: routine reads, writes, and edits don't interrupt the user, while dangerous operations prompt a confirmation. This classifier was also designed to address a new category of security threat — Prompt Injection Attacks: malicious content (such as hidden instructions on a webpage or commands embedded in code comments) could be misinterpreted by the AI as user instructions, triggering dangerous actions like deleting files or exfiltrating data. The classifier mitigates this through context-awareness — it doesn't just assess the danger of the operation itself, but traces whether the instruction source is trustworthy.

This delivers three key benefits:

1. Fewer unnecessary interruptions, preserving developer flow
2. Better protection against prompt injection attacks
3. Context-dependent judgment — the same rm -rf command that removes a folder isn't flagged as dangerous if the user explicitly requested it, with no need for repeated confirmation

This mechanism allows Claude Code to run more autonomously over longer sessions. Claude Code can also now be deployed directly on Google Cloud's Agent Platform (Vertex AI), configured through the built-in setup vertex wizard — the path from experimentation to production deployment for enterprise teams is now fully connected.

Rethinking the Software Engineer's Role

Faced with community debate that "AI is inflating code volume while degrading quality," both guests offered a grounded response.

YK stressed that Claude Code is ultimately a tool: "It's like a painter switching to a power brush — the final output is still your responsibility to oversee." He argued that the solution to quality issues is straightforward: don't generate so much code, don't just barrel through. Hobby projects don't need line-by-line review, but production code should have a Draft PR with a serious AI-assisted review. Even if you generate 100,000 lines, you don't have to commit all of them — only submit what you genuinely stand behind.

Lydia observed that software engineers' roles are shifting toward product management: you're not just writing code anymore — you own it, understand the architecture, and think carefully about why you're implementing something. This shift has historical precedent: from assembly to C, from C to Python, each rise in abstraction level redefined the core value engineers bring — from manually managing memory to designing algorithmic logic, from focusing on syntax details to thinking about architectural boundaries. Her analogy was apt: "I write TypeScript but I don't think about the machine code it generates. Now we've gone up another level — from caring about syntax, to understanding architecture, spotting weaknesses, and using hooks, permissions, and Claude.md to build the right working environment for Claude."

She also pointed to a profound cognitive shift: when writing code by hand, 90% of the review happens during the writing process; with Claude Code, writing is only 10% of the effort — the review phase becomes unprecedentedly important. Many developers haven't made this mental adjustment yet. When Claude takes on the task of "translating intent into code," the engineer's irreplaceable value moves up to two functions: "defining the correct intent" and "verifying that the output matches business logic" — which is precisely the core of product thinking.

The Terminal as a Universal Interface

Toward the end of the conversation, both guests noted that Claude Code has long since transcended the boundaries of a "coding tool." YK described it as a universal interface connecting your computer to the entire digital world — using it to edit videos, run data analysis, free up disk space, and organize files. He shared a real example: when buying a house, he had Claude Code compile a list of real estate agents and generate a mailing list, ultimately "saving $10,000."

Lydia added that after Anthropic noticed large numbers of employees using it for non-coding tasks internally, the team shipped cowork within a week — a desktop product built with Claude Code. A division of labor is now emerging: technical tasks go through Claude Code, while everyday tasks like checking email or booking flights go through cowork (which has more complete service connectors).

"The only bottleneck now is your imagination," Lydia concluded. In the rapid-fire closing segment, when asked for the one tool they could never live without, both guests gave the same answer simultaneously — Claude Code. That might be the most honest summary of the entire conversation.

Key Takeaways