Date: 2026-03-20 Decision: Use the proven 7-phase research-first methodology from Blender, ERPNext, and Draw.io packages. Rationale: This approach produces higher-quality, deterministic skills because research precedes creation. Integration skills require even MORE research than single-technology skills — both sides of every boundary must be deeply understood.
Date: 2026-03-20 Decision: This package should be the LAST AEC skill package to be fully developed. It depends on all other AEC packages being stable. Rationale: Cross-technology integration skills can only be accurate when the individual technology skills they bridge are complete. Building integration skills against incomplete or changing single-technology packages leads to rework.
Date: 2026-03-20 Decision: Target ~15 skills across 4 categories (core: 2, impl: 10, errors: 2, agents: 1). Rationale: Based on initial technology boundary analysis. Covers the most critical integration points between 8 AEC technologies. Will be refined in Phase B3 after deep research.
Date: 2026-03-20 Decision: Focus on the Claude ecosystem. No explicit agentskills.io compliance. Rationale: Claude-specific features (allowed-tools, progressive disclosure) can be used freely without cross-platform constraints. Same decision as all other skill packages in the ecosystem.
Date: 2026-03-20 Decision: Every implementation skill covers exactly one technology boundary (A↔B). Multi-hop scenarios (A→B→C) are handled by the orchestrator agent, not by individual skills. Rationale: Single-boundary skills are more maintainable, testable, and composable. The orchestrator agent chains them for complex pipelines. This prevents combinatorial explosion of skills.
Date: 2026-03-20 Decision: Split skill creation into Wave A (9 skills buildable now) and Wave B (6 skills deferred until dependent packages are complete). Wave A covers: core (2), automation impl (4), errors (2), agent (1). Wave B covers: web visualization impl (3), transport impl (3). Rationale: 4 of 8 dependent packages (Speckle, QGIS, Three.js, ThatOpen) are empty skeleton repos. Building integration skills against incomplete packages leads to rework (per D-002). The 9 Wave A skills have all dependencies satisfied and can be written to full quality now.
Date: 2026-03-20 Decision: crosstech-impl-freecad-ifc-bridge will be more self-contained than other impl skills because there is no dedicated FreeCAD skill package. It must document both sides of the FreeCAD↔IFC boundary without relying on a companion package. Rationale: No FreeCAD-Claude-Skill-Package exists. The IfcOpenShell side is well-covered by the Blender-Bonsai package, but the FreeCAD side must be fully documented within this cross-tech skill.