Create, refactor, and package reusable agent skills

Provides a 9-phase method, plus a compact fast-path, for turning messy workflow material - prompts, transcripts, docs, or notes - into a packaged, reusable agent skill without bloating its entrypoint. Phase 1 first qualifies whether the request should become a skill at all, rejecting one-off explanation, summary, translation, or brainstorming requests with no reuse value, and promoting only work that will be reused, is easy to misroute, benefits from deterministic scripts, or needs governance. Later phases run a short intent dialogue capturing the real job, outputs, exclusions, and constraints; select the lightest of four archetypes (Scaffold for exploratory or personal work, Production for team reuse, Library for broad shared infrastructure, Governed for policy-sensitive or release-critical work); set a four-question capability boundary (recurring job, outputs, near-neighbor exclusions, what belongs outside SKILL.md); scan at most 3-5 reference sources, external benchmarks first, user-supplied second, local files third, for reusable patterns while explicitly recording what not to borrow; and author the frontmatter description before expanding the body, testing route quality via trigger_eval.py, optimize_description.py, and blind and adversarial holdouts before growing the file tree.

Use it for skill creation, refactoring an existing skill, running evals, or packaging a workflow for team-ready distribution - not for a one-off explanation, summary, translation, or answer with no reuse value, which the method explicitly rejects at Phase 1. Optional profiling phases are added only when they earn their cost: an Output Risk profile predicts likely mistakes by output type (tutorials guard against vague examples and missing success checks, reports against weak tables and poor hierarchy, screenshot or visual skills against invented visual evidence, citation-heavy skills against unsupported claims, code or command skills against hidden cwd or side-effect assumptions); an Artifact Design profile defines how a generated report or visual output should read; and a Prompt Quality profile applies when the skill's behavior depends on prompt, role, or dialogue design. For Production, Library, Governed, or team-distributed work specifically, a fixed "Skill OS 2.0" gate sequence runs before release - Skill IR, target compiler, trigger and output eval, Skill Atlas, conformance, trust, registry/package/install, upgrade, drift, and waiver checks, plus a Review Studio pass. A package is only promotable once visible, blind, judge-backed blind, and adversarial holdout all fail to regress, and route confusion stays clean.

Input is the raw workflow material to package - a prompt, a transcript, existing documentation, or informal notes - plus, during intent dialogue, direct answers to a small set of design-changing questions asked in a human, teacher-like framing rather than a cold field list. Output, unless the user asks otherwise, is a SKILL.md, an aligned agents/interface.yaml, any justified supporting assets, and a short summary of the boundary, exclusions, gates applied, and next steps. For file-backed, release-critical, or governed packages specifically, the output must name owner, review cadence, input_files, output contract, and rollback boundary, include a trust report and a reports/output_quality_scorecard.md, and mark any unavailable telemetry, approvals, metrics, or benchmarks explicitly as "missing evidence" rather than fabricating it - these audit labels must be preserved literally when they apply.

Organizes generated content by a strict resource-boundary rule: guidance goes in references/, executable logic in scripts/, and evidence in reports/, keeping SKILL.md itself lean and routable. It cross-references its own family of playbooks - Skill Engineering Method, Intent Dialogue, Skill IR Method, Output Eval Method, and Review Studio Method - plus doctrine documents for artifact design, systems thinking, governance, and SkillOps decision policy, and names specific tooling for trigger-quality testing (trigger_eval.py, optimize_description.py) and a formal Promotion Policy gate.

Teams and individual builders converting an ad hoc workflow, prompt, or set of notes into a properly scoped, evaluable, and, when needed, governance-compliant agent skill, who want a repeatable qualify-scope-author-gate-promote process rather than writing SKILL.md files from intuition alone.