Round 2: notes

Five new strategies, each a different way to design and to review (see strategies/s6-s10). Stock photos allowed (bin/stock.ts), the owner's round 1 verdict passed on in TASK.md, a review-and-fix step required.

| | Strategy | Claude | Cost, time | Images | |---|---|---|---|---| | A | s6 photo art director | 7 / 6 | $5.67, 19 min | 39 | | B | s10 Anthropic frontend-design skill | 7 / 6 | $2.64, 15 min | 26 | | C | s7 blueprint remix | 7 / 7 | $5.30, 18 min | 42 | | D | s8 critic loop | 6 / 5 | $11.07, 49 min | 27 | | E | s9 section posters | 7 / 7 | $3.20, 17 min | 30 |

(Claude's scores aren't fully blind this round: the letter map was printed while checking the builds.)

Owner (Travis): getting quite a bit better. One light-touch steer (don't over-correct): a bit more local business and less SaaS: closer to how a good neighborhood practice or shop looks than a tech product page.

Findings

• Photos fixed most of round 1's problem. All five are photo-led, none is cream/navy/gold, tenure is a detail.
• Directions now differ: refined serif (B), warm photographic (A), bold agency (C), varied posters (E), safe split (D). The bold ones overshoot "premium, traditional"; the refined one runs thin below the fold.
• Free stock is too small a pool: A/D and B/C share hero photos. The generated library (library/) is the fix.
• A fresh critic every round doesn't converge. s8's scores went 6.5, 6, 6, 6: each critic wanted a different hero, and the builder kept redoing the top instead of the weaker bottom half. Review needs continuity (one critic across rounds, or the previous critique + a fixed rubric) and should be pointed at the whole page.
• Cost doesn't buy quality here: the cheapest (B, $2.64) is as good as any.

Next (round 3 ideas)

• Use the generated image library (one house style), stock only for gaps.
• Combine: B's restraint and type with A/E's image density and section variety; a style kept within the owner's look.
• Nudge toward "local business, not SaaS": warmth, place and people over pill buttons, card grids and app polish. Lightly: it's a direction, not a style to swing hard into.
• A review loop with memory: the same critic scores against a fixed rubric and sees its last critique.