How I Use Every Claude Code Feature

claude-codeagentic-codingdeveloper-workflowsubagentshookscontext-management

Abstraction: Practitioner's guide to Claude Code features and workflows

Key points:

Shrivu Shankar (Abnormal AI, 2025-11-02) documents professional use consuming several billion tokens/month for codegen; philosophy is "shoot and forget"—delegate and judge the final PR, not the process.
CLAUDE.md is the single most important file (their monorepo's is 13KB); treat it as curated guardrails, not a manual—document what Claude gets wrong, "pitch" the agent on when to read linked docs, always give alternatives instead of "never."
Avoids /compact (opaque); prefers /clear + custom /catchup, and a "Document & Clear" pattern for durable external memory.
Prefers built-in Task(...) "Master-Clone" delegation over custom subagents, which gatekeep context and force rigid human workflows.
Uses PreToolUse "block-at-submit" hooks (gate git commit on passing tests) but avoids block-at-write hooks; endorses Agent Skills over MCP (per Simon Willison), keeping MCP only as a secure data gateway (e.g. Playwright).
Claude Code SDK used as default agent framework over LangChain/CrewAI; the GitHub Action enables auditable "PR-from-anywhere" with 2 human approvals for AI-initiated PRs.

Connections: Claude Code · Anthropic · Simon Willison · Agentic Coding · Model Context Protocol · AI Agents · Context Engineering