agent-lang
Installation
SKILL.md
agent-lang
Use agent-lang for instructions followed by an agent, not scripts executed by make. Keep each workflow understandable without loading this skill or its references.
Read
- Identify the requested target. Complete its prerequisites successfully before its recipe; failed prerequisites block dependent recipes. Use standard Make dependency ordering, not invented serial ordering of sibling prerequisites.
- Skip the exact first recipe line, the sentinel. Only subsequent
@lines are shell commands; never submit agent instructions or an entire workflow to a shell. - Interpret plain natural-language recipe actions in order, including conditional stops. All actions remain subject to applicable permissions and confirmations.
- Before phase actions, read all invariants declared anywhere in the phase and applicable skill-wide invariants.
INVARIANT[domain-NN]: expressiongoverns the whole declaring phase, not just following lines, and applicable code changes. Explicitly named skill-wide invariants govern every phase and referenced workflow throughout; they cannot override higher-priority instructions or authorize prohibited actions. Check preconditions before relevant actions, preserve constraints during them, and verify postconditions afterward; do not require postconditions to be true beforehand. Abstract expressions require evidence; declaring a condition is not proof that it holds. - At
HUMAN APPROVAL REQUIRED: <named item>, stop and ask for explicit human approval of that item in this conversation; wait before continuing the recipe. Silence or a related message is not approval. - At
SLOT[name]: <capability and intent>, use an available skill or tool. Resolve provider and format at runtime from the caller's skill description and context; ask if no safe suitable provider exists or choices materially affect the outcome.