Claude: Settings & Config

18 sources · Updated March 27, 2026
Claude Code configuration operates at three layers: environment (bypass permissions via --dangerously-skip-permissions, Zed autosave at 500ms, Ghostty + Zed split-screen, game sound hooks for task notifications), context architecture (CLAUDE.md under 200 lines with one-liner declarative behaviors, .claude/rules/ with path-scoped YAML, _MANIFEST.md with Tier 1/2/3 priority loading), and skill design (progressive disclosure cutting 89% context bloat, state-machine skills with data dependency optimization). The platform has six distinct extension mechanisms: Plugins (packaged workflows), Skills (on-demand knowledge), MCPs (external tool connections), Commands (prompt shortcuts), Subagents (delegated tasks), and Hooks (pre/post action scripts). Web fetch uses a subagent architecture that summarizes pages to prevent context overflow but can lose detail. The biggest insight is that system engineering beats prompt engineering — investing 2 hours in context architecture lets subsequent prompts be 10 words long. An 8-phase bootstrap prompt can build an entire workspace from scratch. Effort levels (low/medium/high) are configurable and sticky; conflicting CLAUDE.md instructions are the most common cause of degraded output at high effort.

Insights

  • Exact bypass permissions config: "defaultMode": "bypassPermissions" + "skipDangerousModePermissionPrompt": true in ~/.claude/settings.json. Without skipDangerousModePermissionPrompt, Claude asks for confirmation every session. Shift+Tab toggles bypass mode. (from every claude code hack mvanhorn)

  • Stop hook for audio notification: Add "hooks": {"Stop": [{"hooks": [{"type": "command", "command": "afplay /System/Library/Sounds/Blow.aiff"}]}]} to settings.json. Essential when running 4-6 parallel sessions — know which one just finished. (from every claude code hack mvanhorn)

  • Zed autosave at 500ms creates a Google Docs-like collab experience: {"autosave": {"after_delay": {"milliseconds": 500}}} in Zed settings. Claude edits appear in Zed instantly, and your typing is visible to Claude within a second. Ghostty + Zed split-screen. (from every claude code hack mvanhorn)

  • Claude Code desktop app supports direct DOM element selection — click instead of describe; sends tag name, CSS classes, styles, surrounding HTML context, cropped screenshot, and for React apps also source file path, component name, and current props (from claude code dom element selection)

  • Structure .claude with 35 agent markdown files organized into 7 departments — engineering agents (frontend, backend, mobile, AI, devops, prototyper), testing department (tool evaluation, API testing, workflow optimization, performance, test analysis); each agent file has instructions, personality, scope sections (from one person startup claude agents)

  • Keep CLAUDE.md under 200 lines — longer files eat context and instruction adherence drops (from claude folder anatomy)

  • Two .claude directories: project-level (git) and ~/.claude/ (personal); .claude/rules/ with YAML frontmatter paths field scopes rules to directories (from claude folder anatomy)

  • Commands are single markdown files that become slash commands; Skills are subdirectories with SKILL.md; define subagent personas in .claude/agents/ with restricted tools and cheaper model (from claude folder anatomy)

  • settings.json allow/deny lists control tool permissions; CLAUDE.local.md and settings.local.json for personal overrides (auto-gitignored) (from claude folder anatomy)

  • A three-tier _MANIFEST.md in each project folder (Tier 1: source-of-truth docs, Tier 2: domain folders, Tier 3: archive) gives Claude a prioritized loading order — read the most important context first (from claude cowork context architecture checklist)

  • The minimum viable context setup is three markdown files: about-me.md (role + priorities + work samples), brand-voice.md (tone + writing samples + anti-patterns), and working-style.md (guardrails) (from claude cowork context architecture checklist)

  • Global instructions should direct Claude to read the manifest first, load only Tier 1 files, and ask clarifying questions before starting — this prevents context overload (from claude cowork context architecture checklist)

  • Cowork rewards system engineering over prompt engineering — investing ~2 hours in context architecture means subsequent prompts can be as short as 10 words and still produce high-quality output (from claude cowork context architecture checklist)

  • A folder structure with custom skill files per folder acts as Claude's long-term memory, letting it know exactly how each type of content should be written (from claude cowork workspace setup system)

  • An 8-phase bootstrap prompt can systematically build an entire workspace: Plugins → About Me → Brand Voice → Working Preferences → Content Strategy → Team → Projects → Memory System → Skill Files (from claude cowork workspace setup system)

  • The Memory plugin provides a two-tier context system persisting across sessions, and the Productivity plugin adds task tracking and daily updates — together they make sessions feel continuous (from claude cowork workspace setup system)

  • Content drafts that previously required 3 revision rounds land on the first try once Claude has pre-loaded context about identity, voice, projects, and team (from claude cowork workspace setup system)

  • Progressive disclosure in skill files means restructuring a monolithic instruction file into a slim main file (table of contents) plus separate reference files loaded on demand — 89% context reduction with no loss of functionality (from progressive disclosure claude skills)

  • The instruction file for a Claude skill should function like a table of contents: rules and decision logic up front, reference material (examples, templates, edge cases) in separate lazy-loaded files (from progressive disclosure claude skills)

  • Claude Code changed its default effort level from high to medium to balance intelligence with speed; effort is configurable via /model selector (low/medium/high), and the setting is sticky across sessions (from claude code effort levels)

  • Confusing or conflicting instructions in CLAUDE.md files are the most common cause of unexpected Claude Code behavior at high effort, per the Claude Code team (from claude code effort levels)

  • Skills, MCP servers, and hooks serve distinct roles: skills for reusable prompts, MCP for tool integrations, hooks for lifecycle automation (from claude code daily workflow vibePM)

  • Power users build skills as complex state machines — Brad Feld's /start command is 1,400 lines with 15 steps; mapping data dependencies between steps revealed 12-22 seconds of removable overhead (from claude skill state machine optimization)

  • Skills accumulate "defensive machinery" over time — checks added during development that become redundant but remain as performance drag, suggesting periodic dependency audits for complex skills (from claude skill state machine optimization)

  • Treating a skill as a "state machine with data dependencies" — each step as a node with explicit input/output contracts — is an advanced mental model for systematic skill optimization (from claude skill state machine optimization)

  • Claude Code Desktop now supports --dangerously-skip-permissions for fully autonomous operation; the deliberately scary flag name is a UX pattern that makes the footgun obvious (from claude code desktop skip permissions)

  • A one-line CLAUDE.md instruction ("When you complete work, log it to memory/weekly-recaps/current-week.md") turns Claude Code into a persistent work logger, demonstrating declarative agent configuration with two-minute setup (from weekly recap agent memory)

  • Claude hooks with game sounds (Starcraft, Warcraft, Mario) for task completion/permission notifications solve a key autonomous workflow UX problem; 5.2K likes signals strong appetite for playful DX customization (from claude hooks sound alerts)

  • Claude Code has six distinct extension mechanisms: Plugins (packaged workflows), Skills (on-demand knowledge), MCPs (external tool connections), Commands (prompt shortcuts), Subagents (delegated focused tasks), and Hooks (pre/post action scripts) (from claude code extensions crash course)

  • Claude Code's web fetch uses a subagent architecture: a smaller model summarizes full page content to prevent context window overflow, but critical details can be lost; workaround is saving the full page to a temp folder and analyzing in sections (from claude code web fetch summarizes)

  • Hyper-personalization system built entirely with plain text Markdown files (USER.md, MEMORY.md, family/ directory) — no database, no vector store needed (from shpigford hyper personalization ai)

  • "Daily drip" pattern: cron job asks one thoughtful personal question per day, processes answer, files to right place — adds more useful context after 6 weeks than initial onboarding interview (from shpigford hyper personalization ai)

  • AI onboarding interview (10-15 min structured dialogue, not a form) gets ~60% coverage of personal context; remaining depth comes from incremental daily questions (from shpigford hyper personalization ai)

  • Effective AI memory should be curated and distilled (decisions, lessons, opinions), not raw conversation log — AI reads MEMORY.md every session and updates when something worth remembering happens (from shpigford hyper personalization ai)

  • After 6 weeks of accumulated personal knowledge, AI interactions shift from transactional to collaborative — stops asking clarifying questions because answers are in files it read seconds ago (from shpigford hyper personalization ai)