ultra-instinct-claude-code

# Ultra Instinct Claude Code
# 176 tips from 17 repos, sorted by impact
# Full guide: https://ultra-instinct-claude-code.vercel.app
# GitHub: https://github.com/infiniV/ultra-instinct-claude-code
#
# This file is a condensed reference. Each section has a link to the full
# tip file on GitHub if you need more detail. Read the sections relevant
# to your current task -- don't load everything at once.

---

## 01 SETUP
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/01-setup.md

Install: `curl -fsSL https://claude.ai/install.sh | sh` (npm is deprecated)
Bootstrap: `/init` does 8-phase onboarding, imports from .cursor/rules, AGENTS.md, etc.

Aliases:
  alias c='claude'
  alias cc='claude -c'    # continue last session
  alias cr='claude -r'    # resume session picker

Settings hierarchy (later overrides earlier):
  ~/.claude/settings.json          # global
  .claude/settings.json            # project (commit this)
  .claude/settings.local.json      # personal (gitignore this)

Sandbox mode (restricts file + network without Docker):
  { "sandbox": { "enabled": true, "network": { "allowedDomains": ["api.github.com"] } } }

Custom agents (.claude/agents/reviewer.md):
  ---
  name: reviewer
  model: sonnet
  tools: Read, Grep, Glob
  ---
  Review git diff for bugs and security issues.

Headless: `claude --bare -p "..."` skips all loading, ~14% faster for scripts/CI.

---

## 02 CLAUDE.md
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/02-claude-md.md

Hard limits: ~4,000 chars per file, ~12,000 total. Check: `wc -c CLAUDE.md`
If too long, split into .claude/rules/testing.md, .claude/rules/style.md, etc.
Each rules file supports `paths:` frontmatter to scope when it loads.

@import syntax for large references:
  For API patterns, see @docs/api-conventions.md

CLAUDE.md loads AFTER system defaults, so explicit overrides work:
  "When implementing features, create separate files for each concern."

Add compact instructions to control what survives compaction:
  ## Compact Instructions
  When summarizing, always preserve: file paths, errors, test results, decisions, TODOs.

<important> tags for critical rules:
  <important if="working with database">NEVER run DROP without WHERE clause.</important>

Update CLAUDE.md after every correction. Keep under 30 lines for cache efficiency.
After 2 failed corrections, /clear and write a better prompt.

---

## 03 CONTEXT MANAGEMENT
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/03-context-management.md

Sweet spot: 10-30 min sessions. Quality drops past ~40% context usage.
/compact at ~50% usage. /clear between unrelated tasks.

What survives compaction: CLAUDE.md, tasks, memory files, git state, disk files.
What's lost: intermediate reasoning, file contents, conversation nuance.

The 3-file pattern for complex tasks:
  task_plan.md   -- phases, progress
  findings.md    -- research, decisions, evidence
  progress.md    -- session log, test results, blockers
  Rule: save findings to disk every 2 research operations.

Handoff before context runs out:
  "Write HANDOFF.md with what we built, what works, what's broken, next steps."

Token hogs ranked: MCP tools (~500 tokens each, always loaded) > long CLAUDE.md > agent descriptions.
Keep under 10 MCP servers, 80 total tools. Use ENABLE_TOOL_SEARCH=true for large sets.
/context shows what's eating your window.
--add-dir ../backend for multi-repo access.

---

## 04 COMMANDS & SHORTCUTS
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/04-commands-and-shortcuts.md

Shift+Tab (x2)    Plan Mode
/compact           Compress context
/clear             Fresh start
/branch            Fork conversation for alternate approach
/teleport          Move session to another device
/rename            Name session (helps with multiple instances)
Esc                Cancel response
Ctrl+B             Background a running command
Ctrl+G             Open prompt in external editor
Ctrl+S             Stash current prompt input
claude -c          Continue last session
claude -r          Resume session picker
claude -p "..."    Headless mode (for scripts)
claude -p --output-format json   Machine-readable output

---

## 05 GIT & GITHUB
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/05-git-and-github.md

Worktrees (biggest productivity unlock):
  git worktree add ../feature-a -b feature-a
  # Run separate Claude sessions in each

/commit-push-pr    One-step: create branch, commit, push, open PR
--from-pr 42       Resume session linked to a specific PR
Draft PRs:         gh pr create --draft (Claude creates, you review before merge)

Cross-model review: Claude plans -> Codex reviews -> Claude implements -> Codex verifies
Fresh context review: claude -p "Review this branch against main for bugs." (avoids anchoring)

---

## 06 PROMPTING
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/06-prompting.md

Say what, not how: "Fix the failing CI tests" not "open file X, change line Y"
Paste bug, say "fix": zero micromanagement, Claude debugs well alone.
After mediocre fix: "Knowing everything, scrap this and implement the elegant solution."
After 2 failed corrections: /clear and better prompt. Don't correct a 3rd time.
Force depth: "ultrathink", or "give me 2 plans with tradeoffs before coding."
Anti-sycophancy: "Never say 'interesting approach' -- take a position instead."

Instruction budget: ~150-200 rules max before adherence drops.
If over 150 total (CLAUDE.md + skills + system + tools), split into focused stages.

Assumptions mode (faster than interview):
  "Read the codebase first. Present assumptions with confidence levels.
   I'll only correct what's wrong." Reduces interactions from ~15 to ~3.

---

## 07 PLANNING & SPECS
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/07-planning-and-specs.md

Shift+Tab twice for plan mode. The plan is the highest-leverage part.

Interview pattern: "Interview me about this spec. Ask 40+ questions. Don't implement until done."
Then execute in a FRESH session with /clear.

Phase-gated: Phase 1 -> tests pass -> Phase 2 -> tests pass -> Phase 3
Vertical slices over horizontal: "User auth end-to-end" not "all models then all APIs then all UI"

Confidence check before coding:
  "Rate your confidence 1-10. If below 7, stop and ask questions."

Design doc before implementation:
  "Write a ~200-line design doc: what you found, what you'll build, risks, interfaces."

---

## 08 TESTING & VERIFICATION
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/08-testing-and-verification.md

Feedback loops (2-3x quality): put test/lint/build commands in CLAUDE.md.
TDD: write failing test first, then implement until it passes.
Iron law: "Confidence is not evidence." Always run the actual test.
Fix-first review: auto-fix mechanical issues, only ASK about design decisions.
Santa method: two separate agents review same code, compare findings.
Adversarial verification: "Your job is to BREAK it, not confirm it."
Cross-phase regression: after each phase, run ALL prior phases' tests.
Cross-model QA: use a different model (Gemini, Codex) to audit Claude's output.

---

## 09 AGENTS & ORCHESTRATION
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/09-multi-agent.md

Subagents for exploration: offload file-reading, only the summary returns.
One input, one output per agent. /clear between sequential agents.

Model routing:
  Haiku:  explore, lookups, file reading (cheap)
  Sonnet: implementation, debugging (workhorse)
  Opus:   architecture, security review, planning (expensive but fewer corrections)

Agent frontmatter (.claude/agents/explorer.md):
  model: haiku | sonnet | opus | inherit
  tools: Read, Grep, Glob (scope to minimum needed)
  isolation: worktree (prevents file conflicts)

Agent teams (experimental): CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
Ralph loop: /ralph-loop "fix all tests" --max-iterations 5 (iterates until passing)

---

## 10 HOOKS & AUTOMATION
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/10-hooks-and-automation.md

Exit codes: 0=allow, 2=block tool call, other=warn.
Hook denials are user feedback -- include alternatives, not just "blocked."

22+ hook events: PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart, SessionEnd,
  UserPromptSubmit, PreCompact, PostCompact, TaskCreated, TaskCompleted, CwdChanged,
  FileChanged, ConfigChange, PermissionDenied, WorktreeCreate, WorktreeRemove, etc.

Prompt-based hooks (officially recommended over bash):
  { "hooks": { "Stop": [{ "type": "prompt", "prompt": "Did the agent run tests?" }] } }

HTTP hooks: { "type": "http", "url": "https://your-server.com/webhook" }

Conditional firing: { "if": "Bash(git *)", "command": "./hooks/check-git.sh" }
  Saves thousands of unnecessary process spawns per session.

PreToolUse hooks MUST be fast (<50ms). They fire 10,000+ times per session.

---

## 11 SKILLS & MARKETPLACE
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/11-skills-and-marketplace.md

Skills auto-trigger on context. Commands require manual /invoke. Skills are preferred.

Plugin marketplace: /plugin to browse, install, validate.
Plugin structure: .claude-plugin/plugin.json + commands/ + agents/ + skills/ + hooks/
Sources: git repos, npm packages, local directories.

Useful plugins: code-review (5 parallel agents), feature-dev (7-phase workflow),
  hookify (create hooks from conversation), commit-commands (/commit-push-pr).

Context7 MCP for live library docs (avoids outdated training data).
Skills can store persistent state: ${CLAUDE_PLUGIN_DATA}/data.json
disable-model-invocation: true prevents Claude from auto-triggering dangerous skills.

Search order before building: repo -> npm -> marketplace -> MCP -> community -> build.

---

## 12 MCP & TOOLS
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/12-mcp-and-tools.md

MCP servers in .mcp.json (project) or ~/.claude.json (global).
Each tool schema costs ~500 tokens, loaded every message. Budget: <10 servers, <80 tools.
Replace MCPs with CLI skills for same functionality, zero idle cost.
ENABLE_TOOL_SEARCH=true lazy-loads tool schemas on demand.
Fewer expressive tools beat many narrow ones (20% benchmark improvement).
MCP scopes: Subagent > Project > User.
Bash wrappers eliminate schema overhead: up to 90% token reduction.

---

## 13 PERFORMANCE & COST
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/13-performance-and-cost.md

Cost drivers ranked:
  1. Long sessions (1hr = 10-50x five 12-min sessions)
  2. Opus usage (but fewer corrections can make it cheaper overall)
  3. Parallel subagents (each bills independently)
  4. Large file reads
  5. Unfocused tool calls

Cache: 95% hit rate on sessions >30 min. Three killers: editing CLAUDE.md mid-session,
  long CLAUDE.md files, switching models (destroys entire cache namespace).

PTC (Programmatic Tool Calling): ~37% token reduction for multi-step operations.
CLAUDE.md is in the dynamic (uncached) zone -- always fresh but costs tokens every turn.
Fast mode (/fast): same Opus model (4.8/4.7/4.6), ~2.5x lower latency at higher rate.
  Toggle at session start, not mid-conversation.

Cost benchmarks: bug fix $0.30-0.80, feature $1-3, refactor $5-15, architecture $5-20.

---

## 14 SECURITY & PERMISSIONS
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/14-security.md

Permission deny rules:
  { "deny": ["Read(~/.ssh/**)", "Read(~/.aws/**)", "Read(**/.env*)", "Bash(curl*|bash)"] }

Sandbox mode: { "sandbox": { "enabled": true } } -- restricts filesystem + network.
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 strips API keys from subprocesses.

Scan skills/hooks for prompt injection: 36% of public skills contain injection vectors.
  rg -nP '[\x{200B}\x{200C}\x{200D}\x{2060}\x{FEFF}]' skills/ hooks/

Separate agent identity from personal accounts. Dedicated bot users, scoped API keys.
Never use --dangerously-skip-permissions on autonomous loops or public repos.

Enterprise: managed-settings.json with allowManagedPermissionRulesOnly, allowManagedHooksOnly.

---

## 15 ADVANCED PATTERNS
# Full: https://github.com/infiniV/ultra-instinct-claude-code/blob/main/tips/15-advanced-patterns.md

Sequential pipeline with fresh context:
  claude -p "Analyze..." > analysis.txt
  claude -p "Fix: $(cat analysis.txt)"
  claude -p "Run tests, fix failures"

Model routing: opus for analysis/review, sonnet for implementation.
Tool restrictions: --allowedTools "Read,Grep,Glob" for read-only analysis steps.
SHARED_TASK_NOTES.md for cross-iteration context between claude -p calls.

Docker sandbox: networks: { internal: { internal: true } } blocks all outbound.
Kill process groups, not parents: process.kill(-child.pid, 'SIGKILL')

Self-improvement: save learnings to .claude/learnings.md at session end.
Sprint workflow: think -> plan -> build -> review -> test -> ship -> reflect.