Shareable skills and agent definitions for reliable work with Opus 5 agents — plus the practices guide they come from.
Making Agent Work Reliable — A Practices Guide → The failure modes, the structural fixes, and every artifact below in copy-paste form.working-agreements.md — the project-instructions block that applies to every agent.claude/agents/code-implementer.md — implements settled designs; verifies the premise first, escalates instead of improvising.claude/agents/evidence-gatherer.md — produces cited, dated evidence tables and is barred from concluding.claude/skills/trace-the-chain/ — for "is X enabled / did we fix Y" questions.claude/skills/adversarial-review/ — the pre-merge attack playbook.claude/skills/root-cause-debugging/ — reproduce, verify the instrument, attribute before fixing.claude/skills/empirical-validation/ — pre-register the verdict rule, then measure the number a decision waits on · newer than the guideCopy the .claude/ tree into your project (or into ~/.claude/ for every project), and paste the working agreements into your CLAUDE.md. Agent definitions load unconditionally; skills load when the task matches their description. Full instructions are in the README.