Release v2.0.0: Leaner Skills, Sharper Routing, Stronger Discipline

Today we are shipping Leopoldo v2.0.0, the largest skill corpus overhaul since the system launched. The changes are invisible to day-to-day users but measurable everywhere: smaller context footprint, faster agent selection, stronger discipline on quality-critical skills.

What Changed

Four phases, sixteen waves, 635 skills touched.

Phase 1: Slim

Engine skills are loaded on every session. Pack skills load when a domain is active. Heavier means slower, so we cut aggressively while preserving rules, disclaimers, and anti-patterns.

Result: engine words down 58% (24K -> 10K), top 50 pack skills down 62% (145K -> 55K). Median skill is now under the 1000-word target.

Phase 2: Split

Thirteen oversize madre skills were split. Smaller ones went into two flat sub-skills (B-split). Larger ones became pure routers with detailed daughter skills (C-split). Madre names stayed the same so agents, mapping.yaml, and the backend pack_map continue to route correctly.

Example: compliance-engine became a router dispatching to mifid-compliance, aml-kyc-compliance, gdpr-financial, and aifmd-ucits-compliance.

Phase 3: Descriptions

Every description now starts with a trigger verb: Use when, Use for, Use at, Use before. No more capability lists. No more workflow summaries. The orchestrator reads descriptions to select skills, so trigger-only descriptions mean sharper routing with less noise.

Before: Comprehensive toolkit for drafting contracts with 12 templates. After: Use when drafting commercial contracts from scratch (sale of goods, services, distribution, supply).

Phase 4: Formalization

This is the structural one.

A type field now lives on every skill: technique, pattern, reference, or discipline. Discipline skills (gates, verification, postmortem) require a Rationalizations STOP table, a five-to-eight row list of excuses an agent might use to skip the rule, paired with the corrective response.

Seven engine discipline skills got these tables: phase-gate, doc-gate, workflow-discipline, skill-postmortem, research-before-scaffold, dependency-checker, coverage-analyzer. The pattern is borrowed from Anthropic's Superpowers verification-gate methodology.

We also removed the legacy skillos YAML block from 388 files (breaking change), fixed 162 em dashes in descriptions, added six more engine skills to the Cowork ENGINE_SKILL_SUMMARIES for packaging compatibility, and created a path-scoped authoring rule at .claude/rules/skill-authoring.md that enforces the token budget going forward.

The Numbers

What Breaks

One thing: the skillos YAML block is gone. If you wrote tooling that reads skillos.requires.hard, skillos.provides, or skillos.triggers, it needs updating. The canonical metadata is now name, description, and type. Everything else (version, author, license) lives under metadata.

For end users on Claude Code or Cowork, nothing breaks. Run /leopoldo update and the new versions arrive cleanly.

What Comes Next

Phase 5 is the long tail: 319 pack skills still exceed the 1000-word budget. They are not broken, just verbose. We will slim them in small batches over the coming weeks, maintaining strict rule and disclaimer preservation.

There is also a new studio skill, version-coherence, that audits the three version sources (LEOPOLDO_VERSION, git tag, state.json) and produces coherent release plans. It is how we validated this release and how future releases will be validated.

Why This Matters

Skills are the substance of Leopoldo. The orchestrator is the conductor, the agents are the performers, but the skills are the score. A leaner, more disciplined score means the conductor reads faster, the performers stay in tune, and the audience gets a cleaner performance.

Full changelog on GitHub at leopoldo-ai/leopoldo.ai, tagged v2.0.0.

Want to see these plugins in action? Browse the catalog.