See it. Then change it by saying so.

The egg hatches and your site appears — a real, scrollable preview, not a wireframe. This is the moment of truth, and Yoshi's promise here is simple: nothing is locked. You don't like the hero? Say so. The colours feel cold? Say so. You shouldn't need to learn a tool to make it yours.

Steering in plain language

You change the site by talking to it, the same way you built it:

  you:   "Make the hero warmer and lead with the 24/7 part."
  yoshi: ⟳ adjusting…  ✓ updated the hero — take a look.
  you:   "The deck-staining service needs its own page."
  yoshi: ✓ added it + 9 neighbourhood variants. 1,809 pages now.
  you:   "Use my photo of the blue truck up top."
  yoshi: ✓ swapped the hero image.

No CSS, no layout panel, no "edit mode" to toggle. You describe the outcome; Yoshi makes the change and re-renders the affected pages. For people who want knobs, the direct controls are there too — but they're the floor, not the price of entry.

Steer the structure, not just the paint

The deeper move: because your site is generated from a blueprint (a structured description of your business), steering can change structure, not just styling. "Add a service" doesn't edit one page — it generates that service across every area you serve, wired into the nav, the sitemap, the internal links. You're directing the factory, not nudging pixels.

Pause-to-steer, built in. Yoshi can stop mid-build to ask you to choose — "I found two plausible logos, which feels right?" — and continue from your answer. The generation is a conversation that can wait for you, not a batch job you either accept or reject. status: planned (the interactive runner; cycle 30 §D5 + cycle 28 preview+iterate).

What "good" looks like before you ship

Yoshi holds your site to a quality bar so you're not shipping thin or broken pages:

• No empty pages. Every generated page has real, specific content or it doesn't ship.
• Looks designed, not generated. The editorial bar is benchmarked against genuinely good sites, not other site-builders. status: live (the editorial component library + quality floor).
• Fast + findable by default. Performance and SEO are built in, not a checklist you do later (see Grow & Measure).

🌳 Open questions (the inquisitive seam)

• How much should "say it and it changes" actually do? Copy + image + colour swaps are safe. But "redesign the whole thing" or "make it look like Apple" is a different magnitude. Where's the line between steer (bounded edits) and regenerate (start a fresh variant)? Leaning: two gestures — tweak (in place) vs try another (new variant, keep both). Cross-ref cycle 40 (variants), 54 §D (select-or-generate / "looks like X").
• Undo + history. Natural-language edits need a fearless undo ("actually, put it back"). Do we keep a visible version history of the site, like a doc? Leaning: yes — named snapshots, since the blueprint makes diffing cheap. Cross-ref cycle 32 (blueprint as source of truth).
• Preview what, exactly? A 1,800-page site can't be previewed page by page. Do we preview the ~6 page archetypes (home, service, area, service×area, about, contact) and trust the rest? Leaning: archetype preview + "jump to any page" search. Design decision.
• The "make it look like X" door. Letting a user point at a reference site they love is incredibly powerful and incredibly risky (legal, quality, expectation). Is it a first-class feature or a power-user escape hatch? Cross-ref cycle 54 reference-lookup (P5).
• When does steering end? There's no "done" button in a relationship. Is publish the natural terminus, with steering continuing forever via the dashboard? Leaning: yes — preview-steer and dashboard-steer are the same gesture, before vs after launch. Bridges to 04 + 06.