LLM Wiki

Appearance

Built 26/04/17 07:54commit f2be6c7

Affaan Mustafa Claude Code Longform Guide ​

中文 | English

Summary ​

This source extends the shorthand setup guide into an advanced operating playbook for Claude Code: session logs, strategic compaction, dynamic system-prompt injection, hook-based memory persistence, eval loops, token-aware delegation, bounded parallelization, and reusable patterns that transfer across agent tools.

Source ​

Key Contributions ​

  • Treats session state as a durable artifact through .tmp logs, /learn, and hook-driven persistence instead of relying on one endless thread.
  • Distinguishes strategic compaction, context resets, and system-prompt injection as different tools for controlling what authority and memory each session carries.
  • Makes token optimization operational through model-tiering, subagent benchmarking, CLI-over-MCP substitutions, and explicit cost-awareness.
  • Connects verification to staged eval design: cheap graders, transcript reading, pass-rate metrics, and human review all have distinct roles.
  • Frames parallelization as a bounded technique with explicit kickoff patterns, worktree isolation, and low-overlap task splits.

Strongest Claims ​

  • Long-running agent work stays coherent when memory, plans, and learnings are externalized into files rather than left implicit in chat history.
  • Verification should be designed as a loop with explicit graders and escalation paths, not as a final self-confidence check.
  • Reusable patterns compound across model generations more reliably than tool-specific tricks.
  • Lazy-loaded MCPs reduce startup context pressure, but CLI wrappers and skills can still win on token cost.