Skip to content

Drop in a screenshot and convert it to clean code (HTML/Tailwind/React/Vue)

Python

Update QA.md

Abi Raja0d739aba353c
diff --git a/QA.md b/QA.md
index 482143dc..ea130074 100644
--- a/QA.md
+++ b/QA.md
@@ -19,7 +19,7 @@
## Running QA efficiently
- **Trust the prompt reports, not the UI.** With `PROMPT_REPORTS_ENABLED=1` + `LOGS_PATH=…`, every LLM request is logged with its tool calls, results, and final HTML — far more reliable to grep than scraping the page. Browse them at `/evals/prompt-reports`.
-- **One generation at a time.** Concurrent runs contend on the backend and hit rate limits — run scenarios sequentially.
+- **One scenario at a time — mainly for clean report attribution.** Concurrent runs interleave their prompt reports into the same folder, which is a pain to untangle (running serially also avoids piling ~4 variants each onto the providers, though I didn't actually hit a rate limit).
- **Detect "done" by the chat input returning** (the "Tell the AI what to change…" box), not by scanning page text — its placeholder isn't in `innerText`.
- **Use distinctive, deterministic fixtures** — an unmistakable logo and a clearly-structured screenshot — so you can eyeball whether the right asset was picked.
- **Assets are content-addressed** (`asset_<sha256[:24]>.png`). To prove an exact upload was used, hash the file and look for that filename in the served assets and the generated HTML.