Claude SME

Data Drone runs partner implementation practices that ship one outcome in 45 days. The Anthropic practice gets one production AI workflow live in 45 days on Anthropic's Claude — past the pilot graveyard, into real use. Typical outcome: roughly 80% of a manual process automated and running.

We sell outcomes, never licenses. No shelfware, no drift. We prove the work with screenshots, not poetry — and that means workflows that hold up in production, not demos that fall over. The work is the conversation. No surprises.

You're the Anthropic and Claude domain expert. You own the technical design of the production AI workflow: how we prompt, where we use agents and tool-use, how we wire the Claude API, what guardrails we set, and how we prove it with evals. You're the person who knows what Claude does well, where it breaks, and how to design around both.

You translate a locked use case into a design that survives production — accurate, bounded, and measurable. You set the eval bar so QA has something concrete to validate against. When the team hits a hard Claude problem, you're the one who solves it.

You start the day on the active design problems: the prompt that needs tightening, the agent loop that needs bounding, the guardrail that needs to hold. You spend deep blocks designing and testing — prompts, tool-use schemas, retrieval strategy, eval suites — and pairing with the Integration Engineer to get the design into the real stack. You set the accuracy and guardrail bar with QA so "done" is measurable, not a vibe.

When Claude behaves in a way the client didn't expect, you diagnose it and design the fix. You close the day with design decisions documented and the eval picture honest. No drift, no hand-waving.

• Own the technical design of the production Claude workflow end to end.
• Engineer prompts, agent/tool-use patterns, and retrieval for accuracy and reliability.
• Design and configure guardrails so the workflow stays bounded and safe in production.
• Build eval suites that make accuracy and quality measurable, not anecdotal.
• Partner with the Integration Engineer to get the design live in the client's stack.
• Set the acceptance bar with QA; define what "good enough to ship" means concretely.
• Diagnose and solve hard Claude behavior problems; document design decisions.

• Real depth in prompt engineering — getting reliable, bounded behavior out of Claude or comparable LLMs.
• Hands‑on command of agents and tool-use patterns and the Claude API.
• Proven skill designing guardrails and building evals to measure quality.
• Strong sense for where LLMs break and how to design workflows that hold in production.
• The ability to set a concrete acceptance bar others can validate against.

• Direct experience shipping Claude (or comparable LLM) workflows to production, not just prototypes.
• Familiarity with retrieval, context engineering, and structured output patterns.
• Experience defining and running eval frameworks at scale.

You design for production, not for the demo. You assume the model will surprise you and you build guardrails and evals so it can't surprise the client. You set a measurable bar because "it looks good" isn't a standard. You'd rather ship one workflow that holds than ten that impress in a sandbox and break in week two.

• 30 days — You own the technical design of active engagements and have a clear eval and guardrail approach for each.
• 60 days — Your designs are live in real stacks with measurable accuracy, bounded behavior, and evals QA can run.
• 90 days — You've shipped production Claude workflows that hold up, set a repeatable design‑and‑eval pattern for the practice, and…