حزمة الوكلاء التي أطلقت هذا الموقع.

People ask what 'AI-augmented studio' actually means. The honest answer: agents do about a third of the keystrokes. They don't make the decisions.

Here's the receipt for this site, broken down by which agent touched what.

Claude — primary engineering surface

Most of the React component scaffolding. The Aceternity / 21st.dev port-and-adapt cycle. Type contracts. Server-action plumbing. The first draft of the long-form notes you're reading right now.

What Claude doesn't do: the design decisions. The brand palette. The display-type choices. The information architecture across sections. The voice.

Gemini — copy + alt-prose

Used as a B-pass for tightening sentences. We draft copy in EN, run a Gemini pass with a 'cut filler, keep voice' prompt, accept the cuts that hold up, reject the ones that flatten the rhythm. Same for SQ and AR localization checks.

What Gemini doesn't do: write copy from a blank page. Voice doesn't survive that.

MCP integrations — small but high-leverage

Vercel MCP for deploy verification. Supabase MCP for migration runs. GitHub MCP for issue triage. Each one a small surface, each one cuts about ten minutes per touch.

The compounding effect is real: ten minutes per deploy × five deploys per day × five days a week = four hours of pure friction recovered per developer per week.

What stays human

Brand decisions. Architectural ADRs. Client communication. The yes/no on whether a feature is in scope. The judgment call on whether a section feels right or feels noisy.

Agents are force multipliers on bandwidth, not on judgment. If you put an agent in charge of judgment, the output reads like an agent in charge of judgment.

"The studios that win this decade are the ones that figure out which decisions agents should accelerate and which ones they shouldn't touch."

Our split: agents handle the boring scaffolding. Humans handle the taste. Both contribute. Neither is decoration.