sjinks/instructions-authoring
Instruction-authoring workflow skills for AI agents: creating portable instruction packages, writing and auditing controlled-English agent instructions, assessing agent skills for readiness, and diagnosing instruction defects.
Use to assess whether an Agent Skill, custom-agent prompt, or instruction package is ready for its task, target models, and runtime. Covers discovery, architecture, completeness, portability, maintainability, and evaluability; returns ratings, findings, priorities, and a verdict. Do not use for exhaustive line-level diagnosis or rewrite-only requests.
Use when writing, rewriting, or auditing model-facing agent instructions for controlled-English clarity, consistent terminology, explicit modality, scope, conditions, precedence, and failure behavior. Do not use for ordinary prose editing, code review, or model selection without an instruction artifact.
Use when creating, revising, or adapting Agent Skills, custom agent prompts, subagent instructions, or instruction packages that should work across multiple models or runtimes, especially smaller/faster and frontier GPT/Claude models. Produces a model-neutral core, separates runtime-specific adapters, and checks both small-model executability and frontier-model overconstraint. Do not use for a review-only audit when no authored or revised artifact is requested.
Use when diagnosing exact defects in an AI instruction artifact, Agent Skill, custom-agent prompt, or instruction package: contradictions, precedence gaps, ambiguity, terminology drift, authority or side-effect conflicts, incomplete decision rules, missing failure handling, harmful cognitive burden or duplication, output-contract defects, or explicitly requested custom diagnostics. Produces evidence-backed findings and corrections. Do not use for holistic readiness ratings or for rewrite-only requests that do not include an instruction audit.