SDLC Isolation Lab — flavor1 (vague prompt)

raw PLAYS

One prompt, one shot. No stages, no standards, no vault.
256 cells · reveals on click · 0 tests · 0 errors · 1 file (monolith)
50k in / 31k out · real cost ≈ $0.09

Single inlined index.html. Crude but it runs. · open ↗

tight · iter 4 PLAYS · CORRECT · POLISHED

Lean recipe + contemporary design standard (4/8-pt spacing, hierarchy, whitespace).
81 cells · reveals 40 · 23 numbers · 42 tests · 9 files · centered board · stats panel
194k in / 55k out · real cost ≈ $0.28 — half of iter-3, quality up

Classic 010 MINES 🙂 000 TIME, colored numbers, flood-fill, difficulty presets. The fix-until-loads gate + right-sizing did it. · open ↗

leaky BROKEN

7 stages in one shared dir, everything bleeds into everything.
3 files · 0 cells · dead on load · 0 tests passing
cost not recoverable — the shared dir overwrote its own stage logs (only a $0.00 fragment survived)

Cross-stage contract mismatch: minesweeper.js never exported GAME_STATES. · open ↗

raw · glimmer LOADS · BUGGY

One-shot on a dirt-cheap 30B — no test/verify stage.
renders + reveals · 0 tests · 1 file · real bugs in play
20k in / 5k out · real cost ≈ $0.015

Automated probe says "plays" (loads/reveals/no console errors) — but human play finds real bugs. That gap is exactly what the tight arm's test gate closes. Left as-is. · open ↗

tight · glimmer PLAYS · CORRECT

Full 7-stage recipe on the cheap model.
81 cells · reveals 39 · 26 numbers · 19/19 tests · clean classic UI
397k in / 67k out · real cost ≈ $0.30 ≈ Kimi

Cheap + tool-capable + the recipe = Kimi-level result. (Left some debug scratch files.) · open ↗

leaky · glimmer PLAYS

Shared-dir bleed on the cheap model.
81 cells · reveals 50 · 22 numbers · 0 tests · 3 files
48k in / 2k out · real cost ≈ $0.022

Even the bleed arm "works" here — Minesweeper is small enough that leakage causes no fatal contract break. Exactly why we need a harder project.

SDLC Isolation Lab — the same prompt, built three ways