Agentica — надёжные промпты для мультиагентных систем

ai-tooling

agentica-prompts is a Claude Code skill that provides prompt engineering patterns for Agentica and REPL agents, addressing the ~35% failure rate caused by LLM instruction ambiguity in standard natural-language prompts. It defines a six-phase orchestration workflow — RESEARCH → PLAN → VALIDATE → IMPLEMENT (TDD) → REVIEW → DEBUG — where agents hand off context via `.claude/cache/agents/` directories rather than TaskOutput, keeping the context clean. The skill ships with system prompt templates that inject agent identity, upstream/downstream agent info, and a RepoPrompt code map, plus explicit verb mappings (RETRIEVE, WRITE, EDIT) that replace ambiguous phrasing. Pattern-specific prompts cover Swarm, Hierarchical, Generator/Critic, and Jury coordination styles. According to the source, adopting RETRIEVE with explicit return formats raises success rates from ~60% to ~95%, and structured directory handoffs push that to ~98% — making it essential for teams building reliable multi-agent pipelines on Agentica.