Problem
Teams over-focus on prompts and under-invest in the harness around the model. The result is agents that demo well but fail under production constraints.
Approach
The skill is an installable Claude Code skill that packages reusable patterns for:
- architecture boundaries and contracts via
AGENTS.md(tool catalog, agent map) andPLANS.md(active implementation plans) - deterministic gate sequences:
smoke(tools, deps, typecheck) →check(lint, format) →test(unit, integration, control audit) - observability defaults wired from day one
- low-entropy runbooks that reduce variance in repeated tasks
The core philosophy is entropy control — every agent action should leave the repo in a more predictable state, not a less predictable one.
Architecture overview
The structure combines docs, runbooks, and guardrails so agent execution is constrained by explicit acceptance criteria. The gate sequence runs automatically via pre-commit and pre-push hooks, catching regressions before they reach review.
Current status
Active and evolving as new production lessons are captured. Deployed across multiple repositories with the .control/ directory pattern.
Why it matters
Harness engineering is often the bottleneck between prototype quality and dependable software delivery.