AI writes the first draft now, so who trains your juniors

Your team's tooling is about to change twice. First the prototyping stack, where AI turns a prompt into a landing page in minutes. Then the handoff, where design context skips Figma and travels straight into code. Both make production cheaper. Neither makes the judgment part cheaper, and that gap is where your next hiring problem lives. Let me catch you up.

The deep cut

• Cheap production makes judgment the scarce thing. Same brief, five tools, five different starting points, and someone still has to pick.
• Better context beats a better prompt. DHM fed structured tokens to the agent and hit 90% visual match with no Figma link.
• Skip the reps and you skip the mastery. The Muzli piece calls it the missing ladder, and it puts your junior pipeline at risk.

Same brief, five different answers

One designer ran the same landing page prompt through Figma Make, Bolt.new, Lovable, v0 by Vercel, and Framer AI. Each read it differently. One treated it as a design problem, one as a dev task, one as a marketing page. No universal winner.

The useful measure was not the first version. It was how easy the result was to change. As the tester put it, a prototype that is 80% right but easy to edit beats one that looks great and fights you on every fix.

So the question for your team is not which tool. It is what you are optimizing each tool for, and whether your people can still tell a good draft from a slick one.

When the design system leaves Figma

The bigger shift is in handoff. In one experiment, an AI agent rebuilt a full app screen in SwiftUI with no Figma link and no MCP connection. The design system shipped as a Swift Package: tokens, components, specs, all structured data the code could consume directly.

The result hit an estimated 90% visual match to the original, done in 30 to 45 minutes, with mostly small spacing fixes left. The lesson the author drew was plain: better context can be more useful than a longer prompt. If the agent knows your spacing scale and semantic colors up front, it stops guessing.

For your roadmap, this means the value moves from the Figma file to the structured system behind it. That is a different investment than a prettier component library.

The staircase you are pulling out

Here is the catch nobody puts in the tool demos. The Muzli essay on apprenticeship argues that judgment, taste, and intuition get built through reps. A thousand bad first drafts, a thousand messy workshops. AI now does the first draft, the synthesis, the clustering. It removes the friction that used to train juniors.

The writer calls this the convenience tax: you gain speed now and lose capability later. Worse, tools now produce the signals of seniority, polished decks, sharp wording, before the substance is there. Credibility outruns competence.

This is your hiring risk. The market is romanticizing the top-floor skills while removing the staircase that built them.

Why the last mile is the whole job

The pattern shows up outside design too. Vinoo Ganesh, who built forward-deployed teams at Palantir and Citadel, argues the low-hanging fruit is gone. What is left is the messy 20% of a workflow no product can anticipate from a discovery call.

His example lands. A data engineer blocked a CSV-to-Parquet migration for a year. Every argument failed. Then an FDE watched her work: she checked data quality by double-clicking CSVs open and eyeballing rows. Parquet had no viewer. The tool would have taken away her only instrument.

You cannot infer that from a spec. And AI cannot generate it. That kind of watching, the thing that separates a real insight from a plausible one, is exactly the reps the convenience tax is quietly cutting.

Three questions for your team

• If AI writes the first draft, where do our juniors now get their reps? Name the specific work you will protect for them.
• Is our design system portable as structured context, or does it still stop inside Figma? Decide what to invest in before the next tooling review.
• When two people ship the same polished output, how do we tell who actually made the judgment call? Build that into how you interview and promote.