The Dark Arts of Skill Engineering — Paul Bakaus, Renaissance Geek (Impeccable)
Reliable agent skills need more than prompts: use hooks as automatic guardrails, test across models, and compile harness-specific variants when instruction-following differs.
Impeccable grew from a normalization prompt into a design skill with **two-agent critiques**, scripts and design hooks. Its hooks support Claude Code, Cursor, Codex and GitHub Copilot, while model- and harness-specific builds adjust instructions and tool syntax.
Treat skills as executable harness components. Put recurring checks into **pre-tool or post-tool hooks**, personalize them with project rules, and test each target model instead of assuming one prompt behaves consistently everywhere.
Impeccable grew from a normalization prompt into a design skill with **two-agent critiques**, scripts and design hooks. Its hooks support Claude Code, Cursor, Codex and GitHub Copilot, while model- and harness-specific builds adjust instructions and tool syntax. Treat skills as executable harness components. Put recurring checks into **pre-tool or post-tool hooks**, personalize them with project rules, and test each target model instead of assuming one prompt behaves consistently everywhere. Pre-tool blocking may be necessary for weaker instruction followers, but it is more restrictive. Cross-harness packaging also remains fragmented: common installers do not preserve Impeccable's provider-specific directory layout, so it uses a custom CLI.
Impeccable sharpens skills from reusable instructions into provider-specific executable harness packages: critiques, hooks, scripts, and blocking behavior must be adapted and tested per model. This conflicts with assumptions that one portable prompt or generic installer is enough, while reinforcing that stronger enforcement increases restriction and maintenance burden.