What problem does it solve? Turning a measured, multi-step prompt composition into a shippable skill is error-prone: internal skill ids, variant labels, and wiring code leak into prose, per-step model assignments get lost, and quality evidence is not stamped into the artifact. This Skill automates that packaging so the shipped SKILL.md equals the measured one. ## Core Features & Use Cases - Portable SKILL.md generation: Runs the packager CLI (npx tsx src/packager/cli.ts <name>) to emit packaged/<name>/SKILL.md containing the topology, frozen winning prompts, embedded evals, and a stamped quality bar. - Provisional and contract modes: Supports --allow-provisional for uncalibrated steps, --contract <id> to embed a rubric and enable the LLM judge, and --margin <n> for the quality-bar threshold. - Clean output contract: Hides internal ids in HTML comments, renders dependency lines from wiring, stamps per-step models from winner.json, and adds an Execution section explaining dispatch vs inline execution. - Use Case: After sweeping variants of a research composition, run the packager with the research contract to produce a shippable skill whose self-test bar records the measured quality. ## Quick Start Package the composition named research into a portable skill by running the packager with the research contract and show me the stamped quality bar.
