What problem does it solve? Agents working in multi-run experiments often leave behind notes that are walls of empty headings or final-design pitches with no record of rejected alternatives, forcing future agents to repeat failed work. This Skill standardizes how experiment, infra, focus, and synthesis notes are written so every note carries concrete numbers, mechanisms, and next steps that teammates can act on. ## Core Features & Use Cases - Four note variants: Experiment (per-eval reflection), Infra (grader/build/runtime issues), Focus (direction declaration with abandon-if gates), and Synthesis/Connections/Open-questions (consolidation outputs), each with its own frontmatter schema and section template. - Structured-trace frontmatter: Fields like type, claim, status, confidence, evidence, supersedes, and refutes populate a dashboard knowledge graph with typed edges and confidence-sized nodes. - Bundled helper scripts: stamp.py generates frontmatter-populated skeletons per variant, lint.py mechanizes the self-audit checklist (vocabulary, filename conventions, index cross-references), and unattributed.py finds notes missing the team-critical creator: field. - Use Case: After running a coral eval, an agent stamps an experiment skeleton, fills in results with baselines and deltas, lists rejected approaches, lints the note, and updates notes/index.md so the next agent can build on verified findings instead of rediscovering them. ## Quick Start Use the create-notes skill to write an experiment note for my latest eval by stamping a skeleton with scripts/stamp.py, filling in the result and mechanism sections, and validating it with scripts/lint.py.
