What problem does it solve? Turning a vague user idea, a scientific article, or a GitHub issue into a well-formed ML experiment proposal requires reading the source, mapping it to the project, and confirming intent — this Skill structures that entire sourcing conversation so nothing is fabricated or misunderstood. ## Core Features & Use Cases - Three sourcing branches: Handles article links (fetched via WebFetch), resource links (GitHub issues via gh, spec files, reference repos), and free-text ideas, each mapped to three shaping questions. - Mandatory confirmation gate: Restates the synthesized proposal in plain text and waits for explicit user approval before returning it to the parent iterate-ml-experiment skill. - Risk surfacing: Flags transfer risks, new dependencies, and unverified domain assertions as open gaps instead of silently baking them into the plan. - Use Case: A user pastes a link to a paper on a new regularization technique; the agent reads it, proposes which file in src/<pkg>/ to change, lists dependency concerns, and returns a user-confirmed Proposal block for the experiment journal. ## Quick Start Ask the agent to iterate from a user source, for example: "I want to try the approach from this article — help me turn it into our next experiment proposal."
