Claude Code orchestration · for engineering teams
I install the harness — rules, skills, hooks, memory, and review loops, versioned in your repos — that turns AI coding from per-engineer improvisation into an engineering system your team trusts.
Prefer async? Email a paragraph about your setup — you’ll get a written assessment back within two days.
01 — The pattern
The expensive model writes boilerplate while design, review, and debugging stay manual. Autocomplete with better marketing.
Every engineer prompts differently. Nothing is shared, versioned, or reviewed — output quality depends entirely on who typed.
Agent output is either rubber-stamped or re-read line by line. Neither scales, and both burn the time the tool was meant to buy back.
None of these are model problems. They’re harness problems — and a harness is infrastructure: something you build, version, and test.
02 — The offer
Two to three weeks, part-time, in your repositories. Not a slide deck — working configuration your engineers use the next morning.
I read your repos, interview 3–5 engineers, and map how AI coding actually happens today. You get a written harness design doc before anything changes.
Rules, project skills, hooks, subagent and review architecture, CI wiring — committed through pull requests your team reviews like any other change.
Two 90-minute workshops — operating the harness, then extending it — plus a written playbook and a defined path for adding new rules as your team learns.
Async questions, review of your first harness changes, and tuning — reasonable use, included. I stand behind the install.
Follow-on advisory at $250/hr. Everything produced is work-for-hire and lands in your private repositories.
03 — Proof of work
measured from my own harness · 2026-07-07
The same pattern I install for you — nine surfaces that decide what the model knows, what it may do, and what it must prove:
the harness — versioned in-repo, reviewed like code
A sanitized skeleton of this harness is on GitHub → josephdecker1/harness-skeleton
04 — Who you’d work with
I’m Joseph Decker — a senior site reliability engineer with years of infrastructure and reliability work behind me. For the past two years my daily driver has been an agent harness I built and operate myself: memory systems, adversarial review agents, deterministic pipelines, evals that gate model upgrades.
I sell the thing I already do. If your team’s AI usage is deep, shared, and trusted — you don’t need me.
05 — Questions, answered straight
Teams of roughly 5–50 engineers with Claude Code (or similar) seats and inconsistent results. If nobody on the team has strong opinions about rules files yet, you’re the exact right stage.
Repository access, 3–5 engineers for one interview each in week one, and one champion who reviews the harness PRs. Total team time is a few hours, not a few days.
Work-for-hire. Everything I produce lands in your private repositories and belongs to you. NDA-friendly — send yours or use mine.
The harness pattern ports — rules-as-code, shared skills, review loops, evals. I build Claude Code-native first because it has the deepest hook and subagent surface, and the principles transfer to whatever you run.
The install is a known scope — I’ve done it for my own systems continuously for two years. You shouldn’t fund a learning curve by the hour.
Two clients this quarter
No deck, no discovery theater. Bring one engineer who’s skeptical — that conversation is usually the most useful one.
Or email a paragraph about your setup for a written assessment — async is how I work anyway.