llm-judge

Designed for teams packaging ERP, CRM, HIS, and similar workflows, the tool probes each step for data sources, decision logic, and failure handling before generating the skill. External interfaces are wrapped in Python scripts with mock support, followed by evaluation and iteration. It also supports editing existing skills, changing deployment parameters, and importing reusable domain knowledge.

sub-agents-skills proposes a Markdown format for describing a sub-agent's role and skills so the same task-specific definition can be reused with Codex, Cursor, Gemini CLI, Claude Code, and related tools. It is useful for keeping repeated agent roles in readable files. Actual loading and execution still depend on how each target assistant supports Markdown or skill files.

Skill Master searches local, GitHub, and web sources for a skill that matches a task, and can research and create one when no suitable option exists. It executes the selected skill, reviews what actually happened against the instructions, and updates the skill only when divergences appear. Later feedback can repair the current output and refine the originating skill together.

Token-Diet lays out an ordered optimization workflow for LLM applications: prune retrieved context first, reuse cached material, route work by complexity, keep structured state, and compress prose only at the final step. It also defines metrics and guardrails for token use, latency, cost, and quality, while excluding code, schemas, legal text, URLs, and identifiers from linguistic compression.