workflow-skill-creator

Workflow Skill Creator is a meta-skill from Google DeepMind's science-skills bundle that distills a workflow or interaction that already happened into a reusable agent skill. It is specifically for packaging patterns from a completed multi-step process — not for building skills from scratch, which the description defers to a generic skill-creator. It solves the problem of capturing ad-hoc, one-off workflows before the reasoning behind them is lost, turning them into structured, repeatable skills.

The skill enforces a disciplined multi-phase process. Phase 1 is a mandatory iterative brainstorming conversation with question banks covering workflow understanding, flexibility and error handling, dependencies and resources, scope and shape, and optional testing — the guidance is emphatic that this must complete before any code or SKILL.md is written. Phase 2 produces a design document (skill name and description within frontmatter length rules, directory structure and install location, referenced existing skills, new scripts, rate-limiting strategy, and per-step error handling) for explicit user approval. Phase 3 implements, with guiding principles such as using `uv run` rather than `python`, preferring stdlib libraries like `urllib`, documenting and respecting API rate limits (favoring file-lock–based limiting so concurrent sub-agents share the limit), and referencing existing bundle skills rather than reimplementing them.

The target users are scientists and researchers working within DeepMind's Science Skills environment who want to convert an interaction into a durable skill. Use cases include packaging a data-fetching-and-analysis workflow, capturing a domain protocol, and standardizing a recurring research process into a shareable skill with a CLI helper script when programmatic work is involved.