Make Your Coding Assistant a 10x AI Engineer

A free, open-source plugin called agent-engineer augments coding assistants (Claude Code, Codex CLI, Cursor, OpenCode, Gemini CLI) with senior AI-engineering capabilities: 11 modular skills (checklists and decision rules), 4 isolated subagents for heavy tasks, and 5 hooks (a Python runtime hook that selects skills, blocks accidental API-key writes, and demands evidence before finishing). It keeps a shared repo record for multi-session agent development, loads skills on demand to avoid cost, and installs with a couple of commands or an npx skills option. The design emphasizes layered fixes (address the lowest failing layer), clearer tool interfaces, prompt-context caching, concrete numeric evals with CI gates, per-call observability spans, and architectural guardrails to prevent data leaks. The plugin collects no user data and has no backend.

A blinded A/B test ran six realistic agent-engineering tasks (three runs each) and graded outcomes against prewritten checklists; Claude Code (Opus 5.5) satisfied 53% of requirements with the plugin versus 34% without. Concrete improvements included automatic per-refund and daily caps plus deterministic idempotency keys on refunds, an eval gate that prevented unsubstantiated quality claims in triage work, and fewer unmeasured assertions (unmeasured claims fell from 3/3 runs to 1/3). The workflow enforces clearer tool schemas, mandatory failure-taxonomy-driven evals, and observability that ties LLM and tool calls to traceable spans and costs.