Advanced meta-agent that analyzes full conversation context, maps working patterns, recommends multi-skill workflows (not just single skills), and maintains a learning log of what works.
Contents
- references/skill-inventory.md - how to scan the skills directory and the category snapshot
- references/chains.md - chain design principles, notation, library, and custom-chain builder
- references/patterns-and-logging.md - pattern recognition, usage log schema, learning, proactive tips
- references/response-format.md - required response structure, context templates, edge cases, carryover
Workflow
- Deep context scan. Read the full conversation from start to current message. Identify the primary goal, sub-goals, dependencies, blockers, and past attempts. Check session history in
~/.claude/. Use the context template in references/response-format.md. - Inventory skills. Scan the installed skills directories (
~/.claude/skills/for personal skills,.claude/skills/in the current project, and any plugin skill folders) and read each SKILL.md frontmatter. Never recommend a skill without verifying it exists. See references/skill-inventory.md. - Design chains. Where the task has multiple steps, design a multi-skill workflow so each step feeds the next. Use the notation, library, and builder protocol in references/chains.md.
- Recognize patterns. Read
~/.claude/rules/session-context.mdand~/.claude/projects/memory. Detect recurring tasks, workflow gaps, and underutilized skills. See references/patterns-and-logging.md. - Log usage. Read the existing log at
~/.claude/scout-pro-usage-log.json, factor past outcomes into current recommendations, then append the new recommendation. Update entries when the user reports an outcome. See references/patterns-and-logging.md. - Recommend proactively. Surface valuable unsolicited suggestions grounded in observed patterns.
- Respond. Emit the analysis using the structure in references/response-format.md.
Rules
- Never recommend a skill without verifying it exists; scan the directory first, every run.
- Always explain the "why" behind a recommendation. Do not just list skills.
- Prefer chains over individual skills when the task has multiple steps.
- Respect the user's time. If a one-skill solution works, do not recommend a five-skill chain.
- Be honest about limitations. If no skill is a great fit, say so.
- Update the usage log on every recommendation.
- Do not hallucinate skills. Only recommend skills that exist in the directory or as known slash commands.