Agent Pattern Mining#

Mine reusable agent patterns from real codebases and turn them into concrete upgrades instead of vague inspiration.

Workflow#

1. Scope the source repo and the upgrade target.
   • Clarify whether the goal is analysis only, a new skill, workflow changes, implementation work, or all of them.
   • Prefer targeted reading over exhaustive reading.
2. Map the architecture quickly.
   • Inspect entrypoints, command/tool registries, memory/context code, planning/tasking, permissions, plugin/extensibility, remote/background execution, and observability.
   • Ignore vendor-specific infrastructure unless it carries a transferable pattern.
3. Build a transfer matrix.
   • For each promising pattern, capture:
     • source feature
     • user benefit
     • portability/risk
     • nearest OpenClaw equivalent
     • concrete local action
4. Classify findings.
   • Adopt now: can be captured as a skill, workspace rule, reference note, or small implementation.
   • Prototype later: good idea, but needs tool/runtime changes.
   • Do not copy: tightly vendor-locked, high-complexity, or low-value for the local setup.
5. Apply improvements in this order.
   • Create or improve a skill when the pattern is reusable.
   • Update AGENTS.md for stable workflow changes.
   • Update TOOLS.md for environment-specific operational notes.
   • Update daily memory or MEMORY.md for durable lessons.
   • Propose config/runtime changes only when the user explicitly asks for them.
6. Report the result concretely.
   • Summarize what the source system does well.
   • List what was adopted locally.
   • List what still requires deeper implementation work.
   • Name the exact files created or changed.

What to Look For#

Prioritize patterns that reduce operator burden or improve agent reliability:

• explicit planning modes
• visible task tracking
• context budget awareness
• selective memory recall
• permission boundaries and review gates
• multi-agent orchestration
• deferred tool discovery
• plugin/skill hot reload
• change/diff observability
• session recovery and history search

Deprioritize patterns that are mostly infrastructure-specific:

• first-party SaaS integrations
• internal telemetry wiring
• organization-specific policy systems
• vendor-only remote services without a local analogue

Output Shape#

Use this structure when delivering results:

1. Architecture sweep — the subsystems reviewed
2. Best transferable ideas — ranked by impact
3. Concrete local upgrades — skill/doc/workflow changes already made
4. Gaps left open — good ideas not yet implemented
5. Files changed — exact paths

References#

• Read references/claude-code-patterns.md when the source is Claude Code or when the user wants proven patterns from agentic coding CLIs.

Guardrails#

• Do not claim an upgrade happened unless a file, workflow, or configuration actually changed.
• Prefer lean skills: keep SKILL.md procedural and move detailed analysis to references/.
• When the repo is large, read representative files first, then deepen only where the pattern is promising.
• If the user asks for self-upgrade, translate findings into persistent artifacts, not just commentary.