Opus 4.5 is not the normal AI agent experience that I have had thus far
TL;DR Highlight
Burke Holland built multiple practical apps (Windows utilities, video editor, social auto-poster) in just a few weeks using Claude Opus 4.5.
Who Should Read
Non-ML developers curious about what's actually buildable with AI assistance, and product builders evaluating vibe coding for real apps.
Core Mechanics
- Holland built several complete, usable apps with Claude Opus 4.5 in a timeframe he describes as 'a few weeks' — not toy demos but tools he actually uses.
- The apps covered diverse use cases: a Windows system utility, a video editor with ffmpeg integration, and a social media auto-posting tool.
- The workflow was primarily high-level description + iteration rather than direct coding — he described what he wanted, Claude generated code, he tested and directed refinements.
- Key finding: Claude Opus 4.5 was particularly good at integrating multiple tools/libraries (ffmpeg, OS APIs, social platform APIs) without extensive hand-holding on the integration details.
- Failure modes were mostly around stateful UI edge cases and platform-specific behaviors — things that require running and testing rather than code generation.
- The productivity multiplier for someone with basic coding knowledge but not deep expertise in a particular stack (ffmpeg, Win32 APIs, etc.) was significant.
Evidence
- Holland shared the actual working applications, not just code snippets — the practical functionality validates the claims.
- HN discussion was split: enthusiasts cited this as evidence vibe coding is production-viable; skeptics noted the apps were personal tools without reliability/maintenance requirements.
- Several readers noted the key variable is the human's ability to test and direct — the AI coding productivity multiplier is much larger for someone who can recognize bad output than for someone who can't.
- Comparison to hiring a contractor: Claude is like a very fast contractor who needs clear specs and will occasionally produce work that requires revision.
How to Apply
- For personal productivity tools: if you've had an app idea that's blocked by not knowing a specific stack (ffmpeg, COM automation, shell scripting), this is now the right time to try AI-assisted implementation.
- Set up a test-first workflow: describe what the app should do in testable terms, and use Claude to generate both the implementation and the tests.
- Expect to spend 20-30% of your time on integration testing and platform-specific edge cases — that's where AI currently needs the most human guidance.
- Start with a minimal feature set and expand iteratively rather than trying to generate a full-featured app in one shot.
Terminology
Vibe codingA development workflow where the human describes intent and the AI generates code, with the human in a supervisory/testing role rather than writing code directly.
ffmpegA widely used open-source tool for video/audio processing — powerful but complex CLI with many options, making it a good AI-assistance use case.