SKILL.md is thorough but doesn't consolidate the known limitations an agent will encounter. Add a short section listing them clearly:

• CLI only accepts Int args (CLI argument passing: support strings, floats, and typed dispatch #263)
• vera test cannot generate String/Float64/compound inputs (vera test: Float64 and compound type input generation #169)
• Type inference for bare None/Err in generic calls (Combinators cannot infer type variables from bare None / Err constructors #293)
• Effect row variable unification (Effect row variable unification #294)
• map_new()/set_new() need type context for inference
• Inference: no max_tokens/temperature controls (Inference effect: configurable max_tokens / temperature #370)
• Inference: no user-defined handlers (Inference effect: user-defined handlers (handle[Inference]) #372)

This saves agents from discovering these the hard way and generating broken code.