AI-Native Software Engineer — Claude Code Power User (Agentic Workflows, MCP, Sub-Agents)

# AI-Native Software Engineer — Claude Code Power User (Agentic Workflows, MCP, Sub-Agents) QUALIFIED CANDIDATES ONLY **You don't "use AI to code." Claude Code is where you work.** We're hiring an engineer for whom agentic, terminal-first development is the default — not a tool retrofitted onto an autocomplete habit. If you have to think about whether you're AI-native, this isn't the role for you. If reading that sentence made you nod, keep going. --- ## About the work blueberry is a decision-intelligence platform used by private-equity firms and high-growth companies to evaluate talent and make faster, better-supported decisions. We're in beta and moving fast — speed to market is the governing priority this year. The product runs on Claude as its analytical engine, wrapped in strict methodology: defined reasoning rules, a controlled voice, and hard never-say constraints. Your job is to build and harden the system that enforces all of that in production — faithfully, and without drift. This is genuinely interesting engineering. The methodology is owned by a non-technical authority who defines the rules, the voice, and the judgment calls. You translate that into a system that behaves correctly every time — not one that "usually behaves because the model is pretty good." If "you can't really enforce it, you just have to trust the model" is your instinct, we're not a fit. The entire challenge is *not* trusting the model with high-stakes constraints. ## What you'll actually do - Build and maintain the agentic pipelines that power the product (intake → calibration → extraction → evaluation → synthesis → output). - Enforce methodology, voice rules, and never-say lists at the system level — system prompts, skills, structured-output validation, post-processing checks, eval harnesses, and hooks. Defense in depth, not a single brittle prompt. - Stand up and maintain eval harnesses that catch methodology drift before it ships. - Decide where logic belongs — what goes in prompts vs. skills vs. MCP servers vs. code — and defend the tradeoffs. - Keep a non-technical methodology owner in the loop with tooling that lets them inspect and validate what the model is actually doing. ## What "AI-native" means here (read this before applying) We will be able to tell within a few minutes of talking to you whether this is real. Here's what we mean: - **Claude Code is your environment**, not one tab among many. You operate from the terminal and agentic workflows, not a chat window where you paste one prompt at a time and copy the result somewhere else to "actually run it." - **You live in your CLAUDE.md.** You can describe a real one off the top of your head — its sections, roughly how long it is, how it evolved, and a specific rule you added because you caught Claude getting the same thing wrong repeatedly. - **Sub-agents and multi-step workflows are normal for you.** You've hit the failure modes — context bleed, conflicting instructions, sub-agent loops, permission collisions — and you can tell the story of how you fixed one. - **You know MCP cold**, ideally including building at least one server. And you can cleanly explain the difference between an MCP server (new capabilities/connections) and a skill (procedural knowledge), and when to reach for each. - **You manage context deliberately** across long sessions — plan mode, `/clear` on task switches, sub-agents for isolated work, summaries to disk, hooks. Not "I just start a new chat." - **You catch the model being confidently wrong through structure**, not luck — verification steps, tests you always run, a check against the real codebase, a CLAUDE.md rule you update afterward. ## You're probably a fit if - You're a strong software engineer first — solid fundamentals you built *before* going AI-native — and you've genuinely shifted how you work since. - You have opinions about AI-assisted engineering that some of your peers would argue with, and you can defend them. - You think in layers and tradeoffs when asked where to put logic, and you ask clarifying questions instead of assuming. - You can take rules defined by a non-technical owner and encode them faithfully — you treat the methodology as the source of truth, not something to override on your own authority. ## You're not a fit if - You'd copy code out of Claude Code into Cursor or VS Code to "actually run it," "clean it up," or "make sure it works." - You think all AI tools are basically the same, or you'd rather just use Cursor. - Your workflow is sequential and chat-style: one prompt, one answer, copy, repeat. - You want autonomy over methodology decisions. (The engineer encodes the methodology; they don't redefine it.) - You've "never really run into" the model being wrong. That tells us you're not paying attention or not working on hard enough problems. ## How to apply Skip the generic proposal — we read past those instantly, and a templated pitch is an automatic pass for this role specifically. Instead, include these three things: 1. **A link to a real repo (public or private) where you use Claude Code, including your CLAUDE.md.** Two or three sentences on how that CLAUDE.md got to its current state. 2. **One specific story:** a sub-agent or multi-step workflow you built, what broke first, and how you fixed it. Concrete failure mode, concrete fix. 3. **Your take on one architecture question:** for a product where the model must never use certain words and must always follow specific reasoning rules, where would you enforce that — and why isn't a single layer enough? Strong candidates move to a short voice screen, then a live working session with our team. We make no promises on outcome, rate, or timeline until then. ---