Running a pilot campaign at my agency before rolling this out to clients, and the platform is clearly in early alpha. Zero reporting on the conversation side, tiny character limits, and no real documentation. Here is what I have run into, plus a few things people in the thread flagged that matched my own testing.

Conversation visibility. OpenAI exposes none of it, not even a topic level view. How are you validating that a context prompt is matching real search intent instead of scooping up irrelevant traffic? My only lever so far is micro-segmenting ad groups with ultra-specific long-tail context prompts so I can at least bucket performance roughly. Someone in the thread tried to validate prompts directly and got nowhere, which tracks with what I am seeing.
Structure. Context prompts versus creative angles. The argument made in the thread that structuring by creative angle outperforms structuring by prompt matched my results, and I want to be honest that the baseline is still weak compared to Meta, Google, and Bing. Long-tail prompting is probably the only play until topic-level reporting exists.
Ad copy. The limits are punishing. Titles run roughly 16 to 24 characters before truncation, body copy somewhere in the 32 to 48 range. Generic promo copy dies immediately here. The head space is different: the user trusts the model's recommendation, so offering a plain ad gives them no reason to leave. The copy has to read like an extension of the conversation, a bonus resource or a deeper context or a concrete next step, and pulling that off blind is genuinely hard.
Negative context. There is no true negative keyword equivalent. Keywords are not keywords here, hints are hints. The closest control you get is a single negation sentence inside the framing block. If you are treating the framing block as a description field, flip that: it is the actual control layer.
Landing page. Does LP content feed back into context matching at all? I have not been able to isolate it, no answers on this one yet.

One operational tip worth verifying on your own export. Commas inside a context string appear to register as separate contexts, so a multi-clause prompt silently becomes several. Audit before you scale anything.

# exported ad group contexts, comma-split check
raw = "emergency plumber austin, 24 hour water damage"
contexts = [c.strip() for c in raw.split(",")]
print(contexts)
# ['emergency plumber austin', '24 hour water damage']

If anyone has topic-level workarounds, framing block patterns that hold up, or a landing page test that actually moved numbers, post it. We are all guessing in the dark right now and blindfolded dart throwing is not a media buying strategy.