What problem does it solve? Mechanical one-to-one conversion of functions or endpoints into LLM tools is a documented anti-pattern that pollutes model context and degrades tool selection accuracy. This Skill provides the design rules for turning analyzed source code into a small, curated set of outcome-named tools instead of a per-function dump. ## Core Features & Use Cases - Promotion gate: A checklist that filters discovered functions to only reachable, user-meaningful, effect-classifiable candidates, with default-deny for unclassifiable ones. - Consolidation guidance: Rules for collapsing call-graph clusters and framework-generated CRUD into single outcome-named tools, targeting roughly 20-25 tools from a 300-endpoint app. - Effects metadata mandate: Requires read_only, egress, idempotent, destructive, and auth_scope fields per tool, kept in a typed internal IR since MCP schemas cannot represent them. - Use Case: When reviewing a v2 design that proposes emitting one tool per OpenAPI endpoint, apply this Skill to gate candidates, consolidate clusters, and enforce effect classification before generation. ## Quick Start Ask the agent to review the proposed tool generation design against the tool-synthesis-from-code promotion gate and consolidation rules.
