claude-promptinglisted

# Claude Prompt Optimization Guidance for reshaping user requests into effective prompts for the `claude` CLI provider. Used by `polycli:polycli-provider-agent` before forwarding to the companion script. The `claude` provider is the official Claude Code CLI invoked as a subprocess. It runs in a **fresh session** that does not inherit the parent conversation's CLAUDE.md, tool state, or working memory. ## When to reshape - Parent conversation has rules from CLAUDE.md (language preference, formatting conventions, project context) that matter for the answer — these will NOT propagate to the child claude session - Task references "this file" / "the current directory" / "above" — the child does not see that context - User request would benefit from making implicit constraints explicit ## When NOT to reshape - User provides a clear, self-contained prompt - Request is simple (e.g., "What is X?") - User explicitly says "ask claude exactly this" ## Claude-specific notes - **CLAUDE.md does not propagate.** If language preference or behavioral rules from the parent's CLAUDE.md matter, restate them inline in the prompt. Verified empirically — the bench at `docs/benchmarks/` showed the parent answering in Chinese while the polycli child answered in English under identical inputs. - **Same model family.** The child shares Claude's strengths: structured output, code reasoning, markdown formatting. No need to over-specify output shape — bullet lists and severity tagg