@zsolt.decoding.io And I’m sure there are more examples like this in Liquid Glass.

I miss the times when we had long WWDC sessions talking about pixel level spacing and shaming weird UIs.
www.youtube.com/watch?v=Elr…

@zsolt.decoding.io Photos is the worst offender. The interface flips from black to white and back as you scroll through your library. Meanwhile Safari keeps the old blurred rectangle while Mail gets progressive blur. There’s no consistency in the toolbar background logic.

@zsolt.decoding.io The token savings are significant, but the real benefit is focus.

Each persona does one thing well with the right tools for that job. The system prompts shaped distinct behaviors.

@zsolt.decoding.io Based on subagents I had the following idea: what if I split Claude into specialized personas, each loading only the tools it actually needs? Main persona stays clean and fast.

@zsolt.decoding.io Research persona loads Zen for multi-model reasoning, Context7 for library docs, Octocode for code search. GTD persona connects to OmniFocus, Obsidian, DEVONthink.

The most powerful aspect isn’t just the separation. It’s the ability for personas to delegate to each other.

@zsolt.decoding.io When delegation happens, I must approve it first. The main persona will say something like: “I’d like to ask the research persona to analyze this Rails controller for N+1 queries using its specialized tools.”

@zsolt.decoding.io Everything renders JSON for now, as we haven’t developed a Turbo-based UI yet. We focused on getting the Rails conventions right: nested routes, proper placement of calculations, and correct naming.

@zsolt.decoding.io Still need to test thoroughly and build the UI, but the foundation is solid. The models track progress, adjust for current balances, and integrate with the existing budget structure.

@zsolt.decoding.io Claude searches the codebase and manages task creation while ChatGPT analyzes patterns and suggests Rails-idiomatic approaches. Sometimes I jump in to guide them or ask them to wait for my approval on design decisions.

@zsolt.decoding.io The result?

7 detailed task files with 56 test cases total. Each test case addresses real edge cases like “what happens when a monthly goal already has surplus funds?”

Tomorrow, Claude Code takes these plans and starts implementing.

Full post: decoding.io/2025/07/pez…

@zsolt.decoding.io The technical challenge: handling goal changes over time. I went with snapshots where each month preserves what you were thinking at the time.

January’s allocation stays based on January’s goal settings, even if you change everything in March.