Ask HN: What's wrong with my prompting for Claude 5?

**Claude series 5** is drawing frustration on Hacker News after an Ask HN post titled “What’s wrong with my prompting for Claude 5?” The author reports that they cannot get the Claude series 5 models to reliably do what they ask. As a concrete test, they took a conversation originally produced with **Kimi K3**, fed that context into Claude, and asked Claude to refine Kimi’s output rather than start from a blank slate.

The prompt setup is a transfer-and-improve task, not a cold generation. Claude received the original context, a reference to the Kimi conversation, and an explicit frame that Kimi’s answer was “nearly there” but still “missing a few intellectual tweaks.” The user also told Claude that useful material had already been pre-searched, so the model’s job was to apply better judgment and polish on top of supplied information rather than rediscover facts from scratch.

For engineers and builders, the thread is less about a single bad prompt and more about control failure under multi-model handoff. When you pass another model’s near-complete answer into a newer system and ask for targeted refinement, you are testing instruction adherence, context use, and whether the model will stay inside the requested edit scope. If series 5 still drifts, overwrites, or ignores the “nearly there” constraint, that is a practical reliability problem for anyone chaining tools, agents, or review steps.

The comparison is also a live product contest. **Kimi K3** is treated as the baseline that already did most of the work; Claude series 5 is being judged on whether it can finish the last intellectual gap. Framing the ask as “can you do better than Kimi given pre-searched information” turns a private workflow experiment into a public head-to-head on refinement quality, not raw retrieval.

What to watch next is whether others can reproduce the same failure with similar “refine this almost-done draft” prompts, and which prompt patterns actually force series 5 to honor the existing answer instead of re-solving the whole problem. If refinement tasks keep failing while open-ended generation works, builders should treat Claude series 5 as strong at first drafts and weak at constrained edits until they have a prompt pattern that consistently preserves prior work.