Description
Summary We run a production content-generation pipeline on the Anthropic API (Claude). It works well and produces output we're happy with. The problem is cost per run — it's come in far higher than we projected, and we need someone who knows this platform properly to bring it down. The pipeline sends a large, fixed instruction set on every call, with a smaller variable input per run. Multi-stage, several distinct steps. What we need Review how requests are currently structured and identify where cost is going Implement prompt caching correctly, including request ordering so the cache actually holds Assess whether some stages can run on a cheaper model without quality loss Add per-call cost reporting so we can see what's happening going forward Hand over cleanly to our in-house developer, who owns integration and will work alongside you Important This is not a prompt-rewriting job. The instruction set is deliberate, heavily tested, and stays intact. Cost savings need to come from the engineering — caching, request structure, model routing, output length — not from cutting the content rules. If your proposal suggests shortening our prompts, we're not a fit. Output quality is non-negotiable. Any change needs to be verified against real inputs before and after. How we work Fast and low-friction. We make decisions quickly and expect the same. Direct communication, no long status reports. To apply, please answer these three questions: Our fixed instruction set runs to several thousand tokens, and the system currently runs in short bursts rather than continuously. What cache TTL would you use, and what would change your answer? How would you decide which stages of a multi-step pipeline can run on a cheaper model? How would you prove a change hasn't degraded output quality?