claude-code-kit

Formerly ai-prompts. Old links and clones redirect to this repo.

Skills, mods, subagents, hooks, slash commands and guides for Claude Code — installable by your agent (INSTALL.md).

Project-agnostic guides and executable agents for setting up Claude Code with 70-90% token reduction.

This library contains reusable prompts for implementing global Claude Code optimization across any project. Share with other developers or use to recreate your setup.

---

🤖 Install with an agent

Point your coding agent at INSTALL.md:

Read https://github.com/escapeboy/claude-code-kit/blob/master/INSTALL.md and install what fits my setup.

The agent inspects your environment, proposes a selection (starter, code intelligence, delivery, multi-agent, mods, security, unattended, stack-specific), installs it from the latest release tag after your yes, without overwriting your files, and verifies the result. llms.txt is the index agents use to find their way around.

---

📁 Contents

01-global-optimization

Set up global optimization (ONE-TIME, applies to ALL projects)

• guide.md - Step-by-step installation guide
• setup-agent.md - Executable agent for automated setup
• checklist.md - Verification checklist
• skills/ - Complete SKILL.md files for all 18 global skills (installed to ~/.claude/skills/)
  • optimize/ - /optimize — max token efficiency mode (multi-file: thin core + references/)
  • context/ - /ctx — memory management (multi-file: thin core + references/)
  • cache-inspector/ - /cache-inspector — cache monitoring (multi-file: thin core + references/)
  • update-docs/ - /update-docs — documentation refresh (multi-file: thin core + references/)
  • init-project/ - /init-project — new project setup (multi-file: thin core + references/)
  • agent-ready/ - /agent-ready — AI-agent-readiness audit + selective remediation (multi-file: scanner script + ROI/implementation references)
  • continuity/ - /continuity — repo-local resume→work→finalize lifecycle + evidence-weighted .continuity/STATE.md (multi-file: lint/scaffold script + format reference)
  • self-improve/ - /self-improve — closing-the-loop for the skill library: mine recurring feedback → bounded edit → tiered eval gate (deterministic skill-lint.py + trigger accuracy + LLM judge + with/without-skill behavioral eval via claude plugin eval) → converge (multi-file: linter + behavior-stats.py + deepeval_tier3.py scripts, rubric/behavior-eval/integration/deepeval-setup references)
  • video-digest/ - /video-digest — video or recording → short digest, related notes, fact-checked memory candidates
  • code-research/ - /code-research — multi-agent grounded audit of a local or git codebase into a cross-linked knowledge base (multi-file: per-phase references/)
  • agent-team/ - /agent-team — Agent Teams presets: pr-review, debug, feature, custom
  • codebase-memory/ - codebase-memory-mcp knowledge-graph queries: callers, call chains, dead code, Cypher
  • confidence-check/ - /confidence-check — ≥90% readiness gate before implementation (+ confidence.ts reference implementation)
  • decision-classify/ - /decision-classify — Mechanical / Taste / User Challenge classification to cut interruptions
  • sprint-orchestrate/ - /sprint-orchestrate — Think → Plan → Build → Review → Test → Ship → Reflect with decision gates (--from-design, --no-merge for callers like /company)
  • company/ - /company — an IT company for one task: clarify → research → split into parts → staff teams with fitting models → deliver through sprint-orchestrate; two stops for the user (multi-file: org-design, roster, briefs, workflows, memory references; back office in the company-hq mod)
  • sync-features/ - /sync-features — sync a project's feature inventory into Serena memories and auto-memory
  • ui-ux-review/ - /ui-ux-review — audit UI code for design consistency and accessibility (multi-file: examples + sample report)
• system-prompts/ - Global system prompt files
  • global-optimization.md - Core optimization rules
  • symbol-first-protocol.md - Symbol-first exploration protocol

Time: 2-3 hours (one-time) Benefit: 70-90% token reduction on ALL future projects ROI: Pays for itself in 2-3 sessions

02-project-activation

Activate optimization for a specific project (10-15 min per project)

• guide.md - Project activation walkthrough
• activation-agent.md - Executable agent for Serena activation
• memory-templates/ - Sample memory files
  • architecture-template.md - Project structure template
  • conventions-template.md - Coding patterns template

Time: 10-15 minutes per project Benefit: Project-specific memories for 60-70% session savings Required: After global setup, before starting work

03-custom-skills

Create custom slash commands (skills)

• guide.md - How to write skills
• skill-template.md - Blank template to copy
• examples/ - Working examples
  • example-simple-skill.md - Simple single-action skill
  • example-complex-skill.md - Multi-action skill with integration
  • pwa.md - PWA features skill (service worker, manifest, offline, push notifications)
  • module-mcp/ - /module:mcp — Add a Laravel MCP server to any project (dual transport, domain tools, auth)
  • module-assistant/ - /module:assistant — Add an AI assistant chat panel (Livewire, PrismPHP tools, local agent support)

Use when: You want custom commands like /deploy or /migrate, or full modules like /module:mcp and /module:assistant Benefit: Encapsulate common workflows, reduce repetition; module skills bootstrap entire features

04-research-integration

Research and integrate new Claude API features

• guide.md - How to research and update docs
• research-agent.md - Web research automation
• sources.md - Curated source list

Use when: New Claude API features released, quarterly reviews Benefit: Keep documentation current, adopt new optimizations

05-token-optimization

Deep dive into token reduction techniques

• guide.md - Complete optimization guide
• optimization-agent.md - Token usage analyzer
• metrics.md - Measurement and tracking

Use when: Analyzing token usage, optimizing specific workflows Benefit: Understand and maximize savings, track ROI

06-advanced-patterns

Advanced techniques for complex scenarios

• parallel-agents-guide.md - Multi-agent coordination via Task tool + git isolation with Worktrunk
• agent-teams-guide.md - Native Agent Teams (v2.1.32+, Opus 4.6+) — autonomous peer-to-peer coordination on shared codebases; includes /agent-team skill with pr-review, debug, feature, custom presets
• checkpoint-system-guide.md - Save/resume workflows
• constitution-guide.md - Architectural decision frameworks
• observability-guide.md - Session monitoring, cost tracking, usage pattern analysis
• sprint-orchestration-guide.md - Chain skills into a Think → Plan → Build → Review → Test → Ship → Reflect lifecycle with artifact flow between phases
• decision-classification-guide.md - Mechanical / Taste / User Challenge decision framework, anti-sycophancy rules, dual effort scales

Use when: Complex projects, team coordination, critical decisions, cost visibility Benefit: Handle advanced scenarios with proven patterns; understand where tokens go

07-custom-commands

Custom slash commands for specialized workflows

• debug.md - Debug Agent for systematic debugging
• i18n.md - Internationalization management
• qa.md - QA automation with browser testing
• content-review.md - Content audit for accuracy, consistency, grammar, and translations
• retro.md - Sprint retrospective with git analytics, shipping metrics, per-author breakdowns, and actionable insights

Use when: Specialized workflows, testing, debugging, sprint retrospectives, content quality assurance Benefit: Encapsulate complex workflows into simple commands

08-ui-ux-development

Production-ready UI/UX implementation with Claude Code

• ui-ux-pro-skill.md - Complete UI/UX Pro Max skill documentation
  • 50+ UI styles (Glassmorphism, Minimalism, Brutalism, etc.)
  • 21 color palettes with accessibility guidance
  • 50 font pairings
  • shadcn/ui MCP integration
• dashboard-workflow-guide.md - Step-by-step dashboard implementation
  • Real API data integration
  • Empty states and loading patterns
  • Security best practices (XSS prevention)
  • Dark mode and multilingual support
• browser-testing-guide.md - Systematic browser testing
  • Chrome DevTools via Claude in Chrome MCP
  • Network, console, visual verification
  • Responsive and accessibility testing

Use when: Building dashboards, admin panels, landing pages, SaaS interfaces Benefit: 40-60% token savings, production-ready code with security and accessibility Time: 60-90 minutes per dashboard (vs 3-4 hours manual)

09-laravel-mcp-integration

Connect Claude Code with Laravel's MCP ecosystem

• README.md - Laravel MCP ecosystem overview
• setup-guide.md - Installation and configuration
• laravel-claude-md-template.md - CLAUDE.md template for Laravel projects
• laravel-subagent.md - Custom subagent for Laravel development

Use when: Working on Laravel 11.x/12.x projects with Claude Code Benefit: Project-aware assistance via Laravel Boost + package discovery via LaraPlugins.io Setup: 5 minutes per project (composer install + MCP registration)

10-subagents

Create custom AI agents with specialized knowledge and tools

• README.md - Subagent system overview
• guide.md - Complete subagent creation guide + all v2.1.83 frontmatter fields + production agents
  • plan-challenger (Opus) — adversarial plan review across 5 dimensions with refutation check
  • output-evaluator (Haiku) — LLM-as-Judge: APPROVE/NEEDS_REVIEW/REJECT before commit
  • loop-monitor (Haiku) — watchdog for autonomous sessions: stall/runaway/loop detection
• examples/ - Working subagent examples
  • code-reviewer.md - Read-only code review agent
  • laravel-specialist.md - Laravel development agent
  • debugger.md - Debugging specialist
  • test-generator.md - Test generation agent

Use when: Repetitive specialized tasks, team standardization, cost optimization Benefit: Reusable agents with controlled tool access and model selection Setup: 5-10 minutes per agent definition

11-mobile-development

Mobile development with Claude Code across all major platforms

• README.md - Platform comparison and quick start
• setup-guide.md - MCP server installation for all platforms
• iOS: guide | subagent | CLAUDE.md template
• Android: guide | subagent | CLAUDE.md template
• React Native: guide | subagent | CLAUDE.md template
• Flutter: guide | subagent | CLAUDE.md template
• Skills: /mobile-build | /mobile-test

Use when: Developing iOS (Swift/SwiftUI), Android (Kotlin/Compose), React Native, or Flutter apps Benefit: Platform-specific MCP integration, build/test automation, device control Setup: 5-10 minutes per platform (MCP installation + CLAUDE.md template)

12-desktop-development

Desktop development with Claude Code for macOS, Tauri, and Electron

• README.md - Platform comparison and quick start
• setup-guide.md - MCP server installation for all platforms
• macOS: guide | subagent | CLAUDE.md template
• Tauri: guide | subagent | CLAUDE.md template
• Electron: guide | subagent | CLAUDE.md template
• Skills: /desktop-build | /desktop-test

Use when: Building macOS native (SwiftUI/AppKit), Tauri (Rust + Web), or Electron (Node.js + Chromium) desktop apps Benefit: Platform-specific MCP integration, build/package automation, code signing and notarization workflows Setup: 5-10 minutes per platform (MCP installation + CLAUDE.md template)

14-webmcp

WebMCP — structured browser tools for AI agents (W3C Draft, Chrome 146 Canary)

• guide.md - WebMCP integration guide
  • What WebMCP solves (89% token savings vs screenshot-based approaches)
  • WebMCP vs MCP comparison (frontend vs backend)
  • navigator.modelContext API reference with code examples
  • Implementation patterns (read-only, form actions, declarative HTML)
  • Integration with Chrome MCP and Playwright MCP
  • CLAUDE.md template for WebMCP-enabled projects
  • Current limitations and browser support matrix

Use when: Building web applications that AI agents will interact with Benefit: Structured tool access instead of DOM scraping; 89% token reduction Status: Early Preview — Chrome 146 Canary only, spec actively changing

---

13-security-hardening

Protect Claude Code workflows from MCP attacks, prompt injection, and accidental data loss

• guide.md - Complete security hardening guide
  • MCP vetting checklist + community-vetted safe list
  • Known CVEs (2025-2026) with versions and mitigations
  • Prompt injection defense hooks (PreToolUse + PostToolUse)
  • 6 production safety rules with settings.json + hook implementations
  • permissions.deny hardening templates (global + project-level)
  • Agent Skills supply chain risks and scanning
• hooks/ - Production hook library (actual scripts, copy-paste ready)
  • dangerous-actions-blocker.py — blocks rm -rf /-class commands in any spelling, force-push to main, DROP TABLE, over-broad pkill/killall, edits to key files
  • pre-commit-secrets.py — scans staged content for API keys / private keys / DB URLs before every git commit
  • block-interactive-sudo.py — refuses sudo that would hang on a password prompt
  • tests/ — two-sided matrices (must-block AND must-pass) · WHY.md — the failure behind each hook
  • README with settings.json wiring and the PreToolUse hook contract
  • Productivity hooks (package-version checker, session-start memory loader, shell-habits) live in 01-global-optimization/hooks/

Use when: Team environments, production codebases, regulated industries, before adding new MCP servers Benefit: Prevent data exfiltration, block destructive operations, audit MCP supply chain Time: 15-30 minutes for initial hardening; 5 minutes per new MCP added

16-autonomous-agents

Run Claude Code unattended — cron agents, heartbeat watchdogs, and session journaling

• guide.md - Autonomous & scheduled agents guide
  • The three primitives: headless cron runs, /loop, and hooks — and when each applies
  • Daily journaling agent recipe (~120 production runs/month): idempotent in-place note editing, replace-vs-append sections, self-contained cron briefs
  • Heartbeat watchdog protocol: exact HEARTBEAT_OK token, 200-token failure budget, probe allowlist — born from sessions killed for drifting into investigations
  • Watchdog patterns: stall / token-runaway / repeated-action-loop detection
  • Anti-pattern table — each entry cost a real incident
• heartbeat-template.md - Copy-paste HEARTBEAT protocol file
• session-summary-hook.py - Stop hook: Haiku-summarized session entries appended to a daily note (~$0.001/session)

Use when: Scheduled health checks, automated journaling, any unattended Claude Code run Benefit: Agents that complete within budget instead of drifting; a daily work journal nobody has to write Time: 30-60 minutes for the first cron agent

17-mods

Claude Code mods (v2.1.287+) — TypeScript hooks inside the engine

• guide.md - What mods can do that settings hooks cannot, anatomy, events and engine API, patterns that held up, testing
• marketplace/ - Example marketplace ai-prompts-mods: 13 mods with tests — context-meter, cache-guard, spend-ledger, subagent-models, secret-redactor, ssh-guard, deploy-verify, ci-watch, cleanup-tracker, aside, fleet-status, lang-guard, company-hq

Use when: A policy must hold on every tool call or subagent spawn, output must be rewritten before the model sees it, or you want live UI (meters, panes, status) Benefit: Secrets out of transcripts, model routing that plugin updates cannot undo, cache and spend visible while you work Time: 10 minutes to install the set; 1-2 hours to write your first mod

---

🚀 Quick Start

First Time Setup (2-3 hours)

Option 1: Automated (Recommended)

# Navigate to Claude Code
cd ~/.claude

# Use the setup agent
# Copy contents of 01-global-optimization/setup-agent.md
# Paste into Claude Code conversation
# Agent will create all files automatically

Option 2: Manual

# Follow the step-by-step guide
# Read: 01-global-optimization/guide.md
# Create files as instructed
# Verify with: 01-global-optimization/checklist.md

Activate for Your Project (10-15 min)

# Navigate to your project
cd ~/projects/your-project

# Use the activation agent
# Copy contents of 02-project-activation/activation-agent.md
# Paste into Claude Code conversation
# Agent will activate Serena and create memories

Start Optimized Work

# In your project directory
/optimize "Your task here"

# Or use /init-project for new projects
/init-project --full

---

📊 Expected Outcomes

Token Reduction Targets

Scenario	Baseline	Optimized	Savings
Simple task (bug fix)	22,000 tokens	1,600 tokens	93%
Medium task (new feature)	31,000 tokens	8,500 tokens	73%
Complex task (module creation)	85,000 tokens	18,000 tokens	79%

Conservative target: 30-50% overall reduction Aggressive target: 50-70% with full optimization Maximum achieved: 80-90% with prompt caching on large contexts

Cost Savings

Per session (average medium task):

• Baseline: $0.93 (31,000 tokens @ $3/M)
• Optimized: $0.26 (8,500 tokens @ $3/M)
• Savings: $0.67 per session (72%)

Monthly (30 sessions):

• Baseline: $27.90
• Optimized: $7.80
• Savings: $20.10 per month

Annual (360 sessions):

• Baseline: $334.80
• Optimized: $93.60
• Savings: $241.20 per year

Multiple projects (3 projects, 60 sessions/month):

• Annual savings: $723.60

---

🛠️ What Gets Created

Global Files (in ~/.claude/)

Agents (orchestration):

• agents/pm-orchestrator.md - Central coordinator
• agents/plan-challenger.md - Adversarial plan review (Opus)
• agents/output-evaluator.md - Code quality judge before commit (Haiku)
• agents/loop-monitor.md - Autonomous session watchdog (Haiku)

Hooks (automation):

• hooks/dangerous-actions-blocker.py - Blocks destructive commands and protected files
• hooks/pre-commit-secrets.py - Scans staged files for API keys before commit
• hooks/block-interactive-sudo.py - Refuses sudo that would wait for a password
• hooks/smart-suggest.sh - Suggests the right tool/agent per prompt intent

Settings (configuration):

• settings/prompt-caching.json - 90% savings on cached content
• settings/beta-features.json - Token-efficient tools (14-70% savings)
• settings/model-strategy.json - Haiku/Sonnet/Opus selection
• settings/token-optimization.json - All optimization strategies

Skills (slash commands):

• skills/optimize/SKILL.md - /optimize command
• skills/ctx/SKILL.md - /ctx command
• skills/cache-inspector/SKILL.md - /cache-inspector command
• skills/update-docs/SKILL.md - /update-docs command
• skills/init-project/SKILL.md - /init-project command
• skills/agent-ready/SKILL.md - /agent-ready command (+ scripts/ + references/)
• skills/continuity/SKILL.md - /continuity command (+ scripts/ + references/)
• skills/self-improve/SKILL.md - /self-improve command (+ scripts/ + references/)
• skills/agent-team/SKILL.md - /agent-team command (pr-review, debug, feature, custom)
• skills/video-digest/SKILL.md - /video-digest command (+ scripts/ + references/)
• skills/code-research/SKILL.md - /code-research command (+ references/)
• skills/codebase-memory/SKILL.md - knowledge-graph query reference (auto-triggered)
• skills/confidence-check/SKILL.md - /confidence-check command (+ confidence.ts)
• skills/decision-classify/SKILL.md - /decision-classify command
• skills/sprint-orchestrate/SKILL.md - /sprint-orchestrate command
• skills/sync-features/SKILL.md - /sync-features command
• skills/ui-ux-review/SKILL.md - /ui-ux-review command (+ references/)

System Prompts (behavior):

• system-prompts/global-optimization.md - Applies to all projects/agents
• system-prompts/symbol-first-protocol.md - 65-75% savings via Serena

Documentation:

• README.md - Complete guide
• INSTALLATION-COMPLETE.md - Setup summary
• QUICK-REFERENCE.md - Command cheat sheet

Project-Specific Files (in .claude/)

Per project (10-15 min setup):

• settings/constitution.json - Architectural rules (optional)
• .serena/memories/architecture.md - Project structure
• .serena/memories/codebase-conventions.md - Coding patterns
• .serena/memories/module-structure.md - Navigation (if applicable)
• .serena/memories/testing-strategy.md - Test patterns
• .serena/memories/docker-workflow.md - Docker commands (if applicable)

---

📖 How to Use This Library

Sharing with Others

Send the entire directory:

# Zip the library
tar -czf claude-optimization.tar.gz ~/claude-code-kit

# Share with team
# Recipients extract and follow guides

Or share specific guides:

• New to Claude Code? Start with 01-global-optimization/guide.md
• Need project setup? Use 02-project-activation/guide.md
• Want custom commands? Read 03-custom-skills/guide.md

Rebuilding After Data Loss

If you lose your ~/.claude/ directory:

1. Follow global setup (2-3 hours):

   • Read 01-global-optimization/guide.md
   • Or run 01-global-optimization/setup-agent.md
   • Verify with 01-global-optimization/checklist.md

2. Reactivate each project (10-15 min per project):

   • Navigate to project
   • Run 02-project-activation/activation-agent.md
   • Or manually create memories using templates

3. Resume work with full optimization:

   • All token savings restored
   • Slash commands available
   • PM Orchestrator coordinating

Adapting for Your Team

Customize guides:

• Add team-specific conventions
• Include company coding standards
• Reference your internal documentation

Create team templates:

• Shared constitution for all projects
• Standard memory templates
• Custom skills for team workflows

Version control:

cd ~/claude-code-kit
git init
git add .
git commit -m "Initial Claude Code optimization library"
git remote add origin <your-repo>
git push -u origin main

---

🔧 Maintenance

Weekly

• Review optimization performance: /cache-inspector analyze
• Check token usage trends
• Update memories if code patterns change

Monthly

• Run /update-docs validate to check doc accuracy
• Review and update constitution if needed
• Check for new Claude API features

Quarterly

• Full documentation review: /update-docs research all
• Update memory templates with learned patterns
• Analyze cumulative ROI

---

🆘 Troubleshooting

"Serena not working"

Solution: See 02-project-activation/guide.md → Troubleshooting section

"Low token savings"

Solution: See 05-token-optimization/guide.md → Measurement section

"Skill not found"

Solution: Verify skill installed in ~/.claude/skills/[skill-name]/SKILL.md

"Cache not working"

Solution: Run /cache-inspector status to diagnose

"Constitution not enforced"

Solution: Verify PM Orchestrator loaded, check settings

---

📚 Additional Resources

Official Documentation

• Anthropic Prompt Caching
• Token-Saving Updates
• Extended Thinking
• Claude Pricing

Community Resources

• FastMCP Skills
• SuperClaude Framework
• Serena MCP
• Awesome Claude Code Subagents
• Claude Code Cheat Sheet
• Awesome LLM Apps - 95K+ star collection of LLM agent implementations; see awesome_agent_skills/ for 18 SKILL.md-format skills and multi-agent team patterns
• Claude Code Ultimate Guide - Comprehensive reference covering methodologies, security CVEs, observability, and production patterns

Laravel MCP Ecosystem

• Laravel Boost - Official Laravel MCP server
• LaraPlugins.io MCP - Package discovery MCP

Claude Code Documentation

• Claude Code Subagents - Official subagent docs
• Claude Code MCP - MCP integration guide
• Claude Code Skills - Custom skills guide

Related Tools

• context-mode - MCP plugin that sandboxes tool output to protect context window (98% savings on bulk commands)
• Worktrunk - Git worktree management for parallel AI agents (Apache 2.0)
• Codex-CLI - Code refactoring
• Chrome-DevTools MCP - Browser automation
• Sequential-Thinking - Reasoning support

---

🎯 Success Stories

Project 1: Laravel Multi-Tenant SaaS

• Baseline: 85,000 tokens/session
• Optimized: 18,000 tokens/session
• Savings: 79% reduction, $2.01/session
• Annual: $723.60 saved (360 sessions)

Project 2: React Dashboard

• Baseline: 45,000 tokens/session
• Optimized: 12,000 tokens/session
• Savings: 73% reduction, $0.99/session

Project 3: Python/Django API

• Baseline: 67,000 tokens/session
• Optimized: 14,000 tokens/session
• Savings: 79% reduction, $1.59/session

Combined across 3 projects (60 sessions/month):

• Total savings: $2,170.80/year
• Setup time: 5 hours total (global + 3 projects)
• ROI: Pays for setup in first week

---

🤝 Contributing

If you create custom skills, guides, or improvements:

1. Test thoroughly with your projects
2. Document clearly with examples
3. Share back with the community
4. Consider adding to this library

Format:

• Markdown guides: Clear, step-by-step, with examples
• Agent files: YAML frontmatter + detailed instructions
• Include token savings estimates
• Add troubleshooting sections

---

📄 License

Released under the MIT License — free to use, copy, modify, and share, including commercially. The only condition is keeping the copyright notice.

These prompts are project-agnostic and can be freely adapted for your team's needs.

---

🚀 Next Steps

1. If this is your first time:

   • Read 01-global-optimization/guide.md
   • Allocate 2-3 hours for global setup
   • Use setup-agent.md for automation

2. If global setup is complete:

   • Navigate to your project
   • Read 02-project-activation/guide.md
   • Allocate 10-15 minutes for activation

3. If everything is set up:

   • Start using /optimize for all tasks
   • Monitor savings with /cache-inspector
   • Share this library with your team

---

Questions or issues? Check the troubleshooting sections in each guide or review the official Claude documentation.

Remember: This is a ONE-TIME global setup (2-3 hours) + 10-15 min per project. All future projects benefit automatically with minimal activation effort.

Time investment: 3-5 hours total (global + 3 projects) Time savings: Hundreds of hours in reduced token usage and faster execution Cost savings: $700-$2,000+ per year (depending on usage) ROI: Immediate (pays for itself in 2-3 sessions)

---

Created: 2026-01-04 Last Updated: 2026-10-09 Version: 1.35.1 Compatibility: Claude Code v2.1.32+, Claude API (Fable 5: claude-fable-5, Opus 4.8: claude-opus-4-8, Opus 4.7: claude-opus-4-7, Sonnet 5: claude-sonnet-5, Haiku 4.5: claude-haiku-4-5)

---

📝 Version History

v1.35.1 (2026-10-09)

Changed: /company has no budget cap by default; capUsd on open_project is optional (company-hq mod 0.4.1). Set a cap at stop 2 when you want one.

v1.35.0 (2026-10-08)

Added: optional live dashboard for /company (company-hq mod 0.4.0: one state file per company, local read-only Python server, optional FleetQ ingest, questions and decisions during unattended work); sprint-orchestrate keeps a ## Tasks checklist (Build ticks, Review verifies on a stronger model, Test writes Result:).

v1.34.1 (2026-10-08)

Fixed: removed a stale .pyc file that was committed by accident and copied into every self-improve install; .gitignore now ignores __pycache__/ and *.pyc.

v1.34.0 (2026-10-08)

Changed: repository renamed from ai-prompts to claude-code-kit; links, clone path and merge markers updated. Old URLs redirect; the mods marketplace keeps the name ai-prompts-mods.

v1.33.0 (2026-10-08)

Added: INSTALL.md, a runbook an AI agent follows to install the parts of this library that fit a user's setup (bundles, skill dependencies, no-overwrite rules, verification, updates), and llms.txt as the agent index. Updated: 17-mods notes that mod tests need Claude Code ≥ 2.1.294.

v1.32.0 (2026-10-08)

Added: /company skill (clarify → research → teams → delivery through sprint-orchestrate, two user stops, budget cap) and section 17-mods — guide plus an example marketplace of 13 tested Claude Code mods, including company-hq. Updated: sprint-orchestrate gains --from-design and --no-merge; agent-team links to company.

v1.31.4 (2026-10-07)

Fixed: git-sync-branches no longer rebases (a rebased pushed branch would need a force-push) — Rebase first became Hold, kept and reported; conflicts are staged by file name instead of git add ., and an unresolvable conflict is aborted and held.

v1.31.3 (2026-10-07)

Fixed: git-sync-branches deleted every feat/ branch with -D, including unmerged ones it had not closed. It now deletes only merged branches and those triaged Close; other unmerged branches are kept and reported.

v1.31.2 (2026-10-07)

Fixed: the Tier-3 judge no longer marks a skill down for doing what it is invoked for (e.g. git-sync-branches deleting merged branches). Rubric and 03-custom-skills/guide.md now state the convention: an explicitly invoked skill authorizes its own purpose; gate only what goes beyond it.

v1.31.1 (2026-10-07)

Added: deepeval_tier3.py ~/.claude/skills/ grades every skill in a directory (3 at a time, one line per skill). Fixed: the GPTModel deprecation warning on DeepEval 4.2.

v1.31.0 (2026-10-07)

Added: TIER3_JUDGE=claude-cli — the DeepEval Tier-3 judge runs through claude -p with your Claude Code login (no API key), stripped of tools, MCP, hooks and CLAUDE.md. Changed: fixed evaluation steps + a 1–5 rubric per dimension make scores stable (two Opus runs over 17 skills: mean Δ 0.02, verdicts 17/17 identical; a 7B local judge agreed on 10/17); threshold 0.5 = 3/5. Fixed: confidence-check gains its missing when NOT to table — the first real catch of the scripted judge.

v1.30.0 (2026-10-07)

Added: self-improve/scripts/deepeval_tier3.py — the Tier-3 rubric as DeepEval G-Eval metrics with a swappable OpenAI-compatible judge (LM Studio, Ollama, OpenAI, OpenRouter), plus references/deepeval-setup.md with install steps and the judge-quality caveat: small local judges are smoke-only.

v1.29.0 (2026-10-07)

Added: eight skills from daily local use — code-research (multi-agent grounded codebase audit into a cross-linked knowledge base), agent-team (Agent Teams presets), codebase-memory (knowledge-graph query reference), confidence-check (≥90% readiness gate), decision-classify (Mechanical / Taste / User Challenge), sprint-orchestrate (Think → … → Reflect with gates and an independent verifier before the PR), sync-features and ui-ux-review. Each has When-to-Use and Boundaries sections and links to its sibling skills and guides, with links back from 06-advanced-patterns and 08-ui-ux-development. Skill count 9 → 17.

v1.28.0 (2026-10-02)

Added: the video-digest skill — a video link or local recording becomes a short digest in your language, so you know what it says without watching. Subtitles first, local Parakeet/Whisper (MLX, Apple Silicon) otherwise; keeps description and chapters; links related notes; a separate subagent fact-checks each memory candidate against a page it opened; saves only after asking where. Facts enter Svod as provisional, the raw transcript stays local, and the transcript is treated as untrusted input. Skill count 8 → 9.

v1.27.0 (2026-09-27)

Changed: the guard hooks in 13-security-hardening/hooks/ are now Python — dangerous-actions-blocker.py and pre-commit-secrets.py replace the .sh versions, which matched substrings and were wrong both ways (only one spelling of rm -rf / blocked, rm -rf /tmp/scratch refused, private keys never matched, cmd/ and admin/ never scanned). Update the settings.json paths. The blocker also refuses over-broad pkill -f / killall -m / kill -1 (after a pkill -f "cat" -U … -x terminated ~30 macOS apps), and ignores heredoc bodies that are data rather than shell input. Added: block-interactive-sudo.py; 01-global-optimization/hooks/shell-habits.py (refuses cd X && and standalone grep, naming the fix); two-sided test matrices — tests/test_guards.py (76 cases: 41 must-block, 35 must-pass) and tests/test_kill_rules.py (22); hooks/WHY.md, one line per hook with the failure behind it and when it can go; a note on shell aliases leaking into the Bash tool (unalias under $CLAUDECODE). Ideas from marmelab — The State Of AI Harness Engineering 2026.

v1.26.1 (2026-09-24)

Changed: authoring-convention pass surfaced by a Tier-3 re-judge of the library. agent-ready gains a When to Use (and When NOT to) table and a Boundaries section (its Phase-3 checkpoint and guardrails folded in); init-project's description gains a "Use when" trigger clause. The five 03-custom-skills/examples/ skills (compliance-audit, fix-bug, git-sync-branches, image-optimize, onepassword-integrate) now model the conventions they teach: When-to-Use table + Always / Ask first / Never. image-optimize also copies the originals before its lossy in-place loops. No workflow steps changed.

v1.26.0 (2026-09-24)

Changed: the /context skill is renamed /ctx (skills/context/ → skills/ctx/) — Claude Code ships a built-in /context (context-window usage) that the skill shadowed. All /context load|save|… references now read /ctx; re-install by removing ~/.claude/skills/context and copying skills/ctx. Fixed: skill-lint.py no longer fails on ~/.claude/skills/synced/ (skills Claude Code downloads from claude.ai); "See also" links that left the skill folder are now absolute GitHub URLs, so they resolve from the installed copy too (three were broken even in the repo).

v1.25.0 (2026-09-23)

Added: self-improve Tier 4 — behavioral eval. Tiers 1–3 judge the SKILL.md text; Tier 4 checks that an agent with the skill does the task better than one without it, using Claude Code's built-in claude plugin eval (with/without-skill ablation on a clean baseline). New scripts/behavior-stats.py adds what the runner does not report — normalized gain (p_with − p_without)/(1 − p_without), unbiased pass^k for critical cases, and a discrimination check (a case both arms pass proves nothing) — and fails the cycle on any regression; mean gain ≈ 0 flags the skill as dead weight. New references/behavior-eval.md (case layout, grader types, grader validation, cost, verified runner behavior on Claude Code 2.1.280). skill-lint.py now skips evals/. Ideas borrowed from mgechev/skillgrade, evaluated and not adopted as a tool for a Claude-only library.

v1.24.2 (2026-08-06)

Docs: model-lineup refresh for the Claude Opus 5 release (claude-opus-5, $5/$25, current Opus). Added Opus 5 to the live lineup surfaces — 04-research-integration/sources.md (Models Overview line + pricing table), CLAUDE.md compatibility line. Corrected the Fast mode facts in 01-global-optimization/system-prompts/global-optimization.md and 05-token-optimization/guide.md: speed:"fast" is Opus 5 / Opus 4.8 only at $10/$50 (was wrongly "Opus 4.6 only, $30/$150"; 4.7 fast mode removed; first-party Claude API only). Historical version-history/CHANGELOG entries left frozen.

v1.24.1 (2026-07-18)

Docs: doc-sync after the v1.22–v1.24 skill additions/restructures — updated stale skill counts (6/5 → 8) and multi-file structure notes across README, 01-global-optimization/{guide,setup-agent,checklist}.md, and 03-custom-skills/guide.md; added the 3 missing skills to the Bundled-Files list. Historical entries untouched; no behavior change.

v1.24.0 (2026-07-18)

Changed: 5 global skills (optimize, context, cache-inspector, update-docs, init-project) restructured to the v1.21 authoring conventions — leading When to Use (and When NOT to) table, closing Boundaries, and progressive-disclosure split into references/ — without changing functionality (250–333-line monoliths → thin 87–180-line cores + linked references; content moved, not deleted). Surfaced by dogfooding the self-improve Tier-3 subagent judge over the whole library (these scored 0.2–0.4) and confirmed fixed by re-judging (0.8–1.0, all pass). Tier-2 skill-lint clean.

v1.23.1 (2026-07-18)

Added: .githooks/pre-commit — wires skill-lint.py (Tier-2) as a version-controlled pre-commit gate over 01-global-optimization/skills/. Runs only when a staged change touches the skills tree (docs-only commits pass through); errors block, warnings are advisory, --no-verify bypasses. Enable per clone: git config core.hooksPath .githooks (see .githooks/README.md).

v1.23.0 (2026-07-18)

Added: /self-improve global skill (8th global skill, 01-global-optimization/skills/self-improve/) — a converging self-improvement loop for the skill library: mine recurring review feedback (frequency threshold → "an undocumented requirement") → apply a bounded edit (≤5 changes / ≤100 lines, regression-aborts) → gate through three-tier evaluation → promote the rule → measure convergence. The three tiers: Tier 2 deterministic scripts/skill-lint.py (stdlib linter — frontmatter, kebab-name, description bounds, unreferenced files, duplicate names, trigger-collision; --json emits a sealable evidence record; ready as a pre-commit hook), Tier 1 trigger-accuracy subagent (10 should-trigger / 10 should-not in the real crowded skill env), Tier 3 rubric LLM-as-judge. Multi-file: linter + references/rubric.md (signal-mining + eval protocol) + references/integration-seams.md (optional deterministic-execution / decision-memory / governance / code-intelligence seams). Idea from Salesforce Engineering, "Closing the Loop: How to Build Self-Improving AI Systems with Automated Feedback Loops" (2026-07-17, forcedotcom/sf-skills). Updated: 01-global-optimization/system-prompts/global-optimization.md (the global ~/.claude/CLAUDE.md rules) — new Convention Promotion (recurring feedback → rule) section: a correction seen ~3+ times is promoted into the generator surface + durable policy instead of accumulating another near-duplicate memory; disputed signal calibrates the rule down.

v1.22.0 (2026-07-18)

Added: /continuity global skill (7th global skill, 01-global-optimization/skills/continuity/) — a repo-local resume → work → finalize lifecycle that preserves the operational thread across coding sessions (what's mid-flight, what already failed, what's next), the gap that bigger context/vector/chat memory doesn't close. State lives in an inspectable, git-diffable .continuity/STATE.md where every fact carries an evidence tag ([observed]/[validated]/[user] = trust; [claimed]/[contradicted]/[unknown] = low-trust) so uncertainty stays visible. Multi-file: scripts/continuity.py (stdlib-only scaffold + lint of structure/tags/provenance) and references/format.md (section spec + worked example + local-vs-portable split). Explicitly a thin layer over the existing memory stack (Serena/Svod/codebase-memory/auto-memory), not a new store — the skill forbids creating .continuity/ for single-shot work to avoid a sixth-memory-store. Idea adapted from Santi Santamaria Medel, "Maybe Coding Agents Don't Need a Bigger Memory. Maybe They Need Continuity." (Level Up Coding, 2026-06-18). Updated: 01-global-optimization/system-prompts/global-optimization.md (the global ~/.claude/CLAUDE.md rules) — new Session Continuity (resume → work → finalize) section wiring the lifecycle into the mandatory session-start/session-end rituals.

v1.21.0 (2026-07-14)

Added: Five skill-authoring patterns adapted from garrettw/php-arch-skills, folded into 03-custom-skills/ and the root CLAUDE.md: (1) a "When to Use (and When NOT to)" decision table as a standard leading section — the operational form of the don't-gold-plate rule; (2) a "Boundaries" rubric (Always Do / Ask First / Never Do), with Ask First aligned to Action Safety; (3) references/ progressive disclosure as the default — the template now ships a thin-core + scripts/ + references/ multi-file structure; (4) explicit bidirectional cross-linking between skills (new guide section + strengthened When Editing rule); (5) a reference-split (matrix) pattern (guide Pattern 5) for one concern across many variants, with 11-mobile-development / 12-desktop-development as worked examples. Updated: 03-custom-skills/skill-template.md, 03-custom-skills/guide.md, and root CLAUDE.md.

v1.20.0 (2026-07-14)

Added: /agent-ready global skill (6th global skill) — audits a project's public site for AI-agent readiness via Cloudflare's isitagentready.com scanner (robots.txt, sitemap, llms.txt, Link headers, Content Signals, Markdown-for-Agents, MCP / OAuth / A2A / Agent-Skills discovery, agentic commerce), then triages each failing check by ROI for the site type (content vs. API vs. agentic-commerce) and implements only what pays back — treating "implement every check" as the failure mode. First multi-file skill in the repo: a stdlib-only scanner script (scripts/scan.py, API-first with a direct-probe fallback) plus references/applicability.md (ROI decision table + AI-crawler policy) and references/implementations.md (per-stack recipes for Laravel / Next.js / Node / static). Cross-referenced: linked from 14-webmcp/guide.md (the webMcp discovery check) and 07-custom-commands/seo.md (shared robots/sitemap/AI-crawler/GEO surface).

v1.19.0 (2026-07-02)

Added: Default Tooling Stack in the global template — Lattice (browser + web search), Svod (persistent memory), Svod Foundry (reusable tools), Harbormaster (cross-project agent comms), each with a fallback Updated: /update-docs skill hardened (→ v1.1.0) from two real runs — prefer Lattice / ctx_fetch_and_index over the context-mode-blocked WebFetch, lead with authoritative platform.claude.com + claude-api sources, analyze as the default action, an immutable-historical-records rule + Releasing-convention pointer, de-staled examples, and clearer scope vs /content-review

v1.18.0 (2026-07-02)

Updated: Sonnet 5 model refresh (from a /update-docs research pass against the live Claude docs)

• Claude Sonnet 5 (claude-sonnet-5) is now the current Sonnet — a drop-in upgrade for Sonnet 4.6 with adaptive thinking on by default and manual extended thinking / non-default sampling params returning 400
• Bumped live claude-sonnet-4-6 example/frontmatter/config references to claude-sonnet-5 across 01–05 and 10; updated compatibility lines (README footer, CLAUDE.md, 09/11/12) and pricing tables (04/05/06) with Sonnet 5's intro $2/$10 (through 2026-08-31) → $3/$15 standard
• Fable 5 / Opus 4.8 / Haiku 4.5 lineup and pricing re-verified against the live docs — unchanged
• Content hygiene: translated untranslated Bulgarian prose/comments to English (03, 06, 10); kept deliberate Bulgarian example content (07 content-review language rules, 08 native-script language-switcher sample)

v1.17.1 (2026-06-15)

Added: LICENSE file — the repository is now released under the MIT License

• Previously the README declared a "freely shared" intent but had no LICENSE file, so GitHub reported "No license" (all-rights-reserved by default)
• MIT covers both the prose guides and the executable hook/skill files; README License section updated to match

v1.17.0 (2026-06-10)

Added: Production patterns distilled from 3 months / ~450 sessions of real Claude Code work (all content sanitized — no personal projects, hosts, or infrastructure)

• 16-autonomous-agents/ — new section: cron journaling agent recipe, heartbeat watchdog protocol (HEARTBEAT_OK token + budget design), watchdog detection patterns, /loop usage, copy-paste heartbeat template + session-summary Stop hook
• Hooks as actual files — 13-security-hardening/hooks/ (dangerous-actions-blocker, pre-commit-secrets) and 01-global-optimization/hooks/ (check-package-latest, session-start-memory-load), each with settings.json wiring and design notes
• 06-advanced-patterns/headless-review-fanout-guide.md — diff-scoped parallel reviewers in worktrees with structured output + adversarial verify pass (caught a real cross-tenant IDOR in production use)
• 06-advanced-patterns/cross-project-delegation-guide.md — concept guide: MCP message bus between project-scoped agents (delegation, fan-out Q&A, clarification loop)
• 03-custom-skills/skill-taxonomy-guide.md — designing families of 40+ skills: methodology → dispatcher → leaf skills → validation gates, dated-backup versioning, token economics
• 5 new skill examples in 03: fix-bug (three-phase), git-sync-branches (submodule-aware), image-optimize, compliance-audit, onepassword-integrate
• 4 new subagent examples in 10: loop-monitor (watchdog), output-evaluator (LLM-as-Judge gate), plan-challenger (adversarial + refutation), self-review

Updated:

• 01 global template — ported 11 battle-tested sections: Code Discipline, Faithful Reporting, rename/removal discipline, bug-fix completeness, background delegation, communication protocol, action safety, worktree isolation, auto-memory, recurring tasks, effort levels
• 02-project-activation — new Memory Hygiene & GC section (append-don't-create, durable-facts-only, periodic GC, feature-inventory sync)
• 04-research-integration — new section: license-compatibility gate (AGPL/SSPL blockers) + evaluating unsolicited vendor audit PRs

v1.16.0 (2026-06-10)

Updated: Model lineup refresh for the Fable 5 / Opus 4.8 era

• New model tier documented: Fable 5 (claude-fable-5, $10/$50 per MTok); Opus 4.8/4.7 (claude-opus-4-8/-4-7, $5/$25, 1M context) replace Opus 4.6 as the recommended Opus
• Fixed wrong pricing in observability guide (Opus listed at $15/$75 → actual $5/$25; Haiku $0.80/$4 → $1/$5)
• Opus 4.7+ breaking changes documented: budget_tokens and temperature/top_p/top_k removed (400), thinking text omitted by default (display: "summarized" to opt in), effort: xhigh added
• New token levers in 05: Task Budgets beta (task-budgets-2026-03-13), server-side compaction header (compact-2026-01-12), mid-conversation system messages (mid-conversation-system-2026-04-07)
• Fast mode caveat: speed: "fast" API param is Opus 4.6-only; no fast variant on 4.7/4.8
• All claude-opus-4-6 example/frontmatter references bumped to claude-opus-4-8; stray claude-sonnet-4-5 default bumped to claude-sonnet-4-6

v1.15.0 (2026-03-31)

Added: Source-informed improvements from analysis of Claude Code's actual system prompt architecture

• 4 new MANDATORY sections for the global CLAUDE.md template: Code Discipline, Faithful Reporting, Communication Protocol, Action Safety
• Memory type system (user/feedback/project/reference) with "what NOT to save" exclusion list in 02-project-activation
• Cache-friendly prompt ordering (static/dynamic boundary pattern) in 05-token-optimization
• Subagent advanced patterns: fork context protection, read-only agent template, anti-delegation rule
• Output style skill examples (explanatory, audit, teaching) in 03-custom-skills

Added: Sprint lifecycle patterns inspired by gstack analysis

• /retro skill — sprint retrospective with git analytics, per-author breakdowns, shipping streaks
• Sprint orchestration guide — Think → Plan → Build → Review → Test → Ship → Reflect pipeline
• Decision classification guide — Mechanical / Taste / User Challenge framework with anti-sycophancy rules
• AI slop detection in UI/UX Pro skill — 0-10 scoring rubric, 10 common patterns, fix checklist
• Anti-sycophancy and decision authority sections in skill template

v1.13.0 (2026-03-28)

Added: context-mode Guide (15-context-mode/)

• Installation via Claude Code plugin (/plugin marketplace add mksglu/context-mode) and MCP-only mode
• Tool reference: ctx_batch_execute, ctx_search, ctx_execute, ctx_execute_file, ctx_fetch_and_index, ctx_index
• Mandatory tool selection hierarchy (Tier 1 → Tier 2 → Tier 3) and forbidden patterns
• Session continuity: SQLite FTS5 event capture, --continue restoration across compaction
• Integration matrix with Serena, memory system, prompt caching, and Agent Teams
• Savings benchmarks: 94-99.9% context reduction vs raw tool output

Updated: README.md — added context-mode to Related Tools, bumped to v1.13.0

v1.12.0 (2026-03-27)

Added: Hooks library, production agents & agent-team skill

• "Production Hook Library" section in 13-security-hardening — dangerous-actions-blocker.sh, pre-commit-secrets.sh, smart-suggest.sh with matcher-based settings.json registration and exit 2 vs exit 0 semantics
• "/agent-team Skill" section in agent-teams guide — 4 preset modes (pr-review, debug, feature, custom), feature flag via settings.json env block, navigation shortcuts
• "Production Agents (Real Examples)" section in 10-subagents — plan-challenger (adversarial plan review), output-evaluator (LLM-as-Judge verdict gate), loop-monitor (autonomous session watchdog)

v1.11.0 (2026-03-26)

Updated: New subagent and skill frontmatter fields from Claude Code v2.0.43–v2.1.83

• Subagents: disallowedTools, permissionMode, background, isolation: "worktree", effort, maxTurns, hooks, memory, initialPrompt, skills, full model IDs in model:
• Skills: context: fork, agent, user-invocable: false, effort, allowed-tools wildcard syntax, $0/$1 argument shorthand, ${CLAUDE_SESSION_ID} and ${CLAUDE_SKILL_DIR} variables

v1.10.0 (2026-03-21)

Added: /content-review skill (07-custom-commands/content-review.md)

• Content audit for documentation, UI copy, error messages, and translation files
• Scope filtering (docs/ui/errors/api), multilingual checks (Bulgarian/English), structured report output
• Based on production audits of Zonex/Signalio (23+ issues found and fixed)

v1.9.0 (2026-03-20)

Added: WebMCP Integration Guide (14-webmcp/)

• W3C Draft Community Group Report (Google + Microsoft) — navigator.modelContext API
• WebMCP vs MCP comparison (frontend/ephemeral vs backend/persistent)
• API reference: registerTool(), unregisterTool(), requestUserInteraction()
• Implementation patterns: read-only data, form actions with confirmation, declarative HTML
• Integration guide for Chrome MCP and Playwright MCP workflows
• CLAUDE.md template for WebMCP-enabled projects
• Browser support matrix and current limitations
• 89% token savings vs screenshot-based approaches

Updated: README.md — added 14-webmcp section, bumped to v1.9.0

v1.8.0 (2026-03-20)

Fixed: Critical audit of ~/.claude/ setup — identified files Claude Code never reads

• ~/.claude/system-prompts/ is NOT loaded by Claude Code — migrated content to ~/.claude/CLAUDE.md (which IS loaded globally)
• ~/.claude/settings/*.json files (prompt-caching, beta-features, model-strategy, token-optimization) are reference docs only, not active config
• Removed 13 duplicate skill copies from ~/.claude/settings/ (already in ~/.claude/skills/)
• Removed 15 misplaced .md files from ~/.claude/settings/ (agent/command copies)
• Removed dead docs from ~/.claude/ root (README.md, QUICK-REFERENCE.md, INSTALLATION-COMPLETE.md)
• Updated all model IDs: claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5-20251001

Added: Important compatibility notice in 01-global-optimization/guide.md explaining what Claude Code actually reads from ~/.claude/

Updated: CLAUDE.md — corrected architecture description and model compatibility

v1.7.1 (2026-03-20)

Added: Git isolation with Worktrunk section in 06-advanced-patterns/parallel-agents-guide.md

• Worktrunk CLI integration for running parallel Claude agents in isolated git worktrees
• Installation guide, core commands, Claude Code parallel workflow example
• Hooks integration and decision table (when to use worktrees vs in-process parallelism)
• Added Worktrunk to Related Tools in README

v1.7.0 (2026-03-18)

Fixed: Critical content gaps in 01-global-optimization/ (reported in issue #1)

Added: Complete SKILL.md files for all 5 global skills (previously placeholder text):

• /optimize — task complexity analysis, planning strategy selection, model routing, metrics reporting
• /ctx — full memory management (load, save, refresh, list, inspect, clear) with token savings docs
• /cache-inspector — cache hit rate analysis, cost breakdown, optimization recommendations
• /update-docs — web research + doc comparison + targeted update workflow
• /init-project — full stack detection, constitution generation, memory creation, per-framework support

Added: Complete system prompt files (previously placeholder text):

• global-optimization.md — all mandatory/automatic optimization rules ready to copy to ~/.claude/
• symbol-first-protocol.md — complete symbol-first protocol with patterns, examples, token savings table

Fixed: setup-agent.md no longer has circular references — now points to actual files in repo Fixed: 03-custom-skills/guide.md now references pwa.md example Fixed: 04-research-integration/guide.md now links to sources.md

v1.6.0 (2026-03-09)

Added: Security Hardening Section (13-security-hardening/)

• MCP vetting workflow with 5-minute audit checklist
• Known CVEs (2025-2026) table with affected versions and patches
• Prompt injection defense hooks (PreToolUse + PostToolUse)
• 6 production safety rules with settings.json deny rules and hook implementations
• permissions.deny hardening templates (global ~/.claude/ + per-project)
• Agent Skills supply chain risk data (36.8% of public skills have security flaws)
• Incident response playbook

Added: Agent Teams Guide (06-advanced-patterns/agent-teams-guide.md)

• Native experimental feature (v2.1.32+, Opus 4.6+, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
• Architecture: peer-to-peer messaging, git-based locking, isolated 1M-token contexts per agent
• Decision matrix: Agent Teams vs Parallel Agents vs Dual-Instance vs Multi-Instance
• 4 copy-paste patterns: pre-release review, security PR review, multi-file doc update, parallel refactor
• Limitations, red flags, and anti-patterns

Added: Observability Guide (06-advanced-patterns/observability-guide.md)

• Session search script (cs) — zero-dependency bash, ~15ms for recent sessions
• Community tools comparison (session-search.sh vs claude-conversation-extractor vs ran CLI)
• Cost tracking scripts: per-session and weekly reports
• Usage pattern analysis: most-read files, tool distribution
• Team visibility: aggregate usage across developers
• Cross-folder session migration guide

Updated: README.md community resources, compatibility to Claude Code v2.1.32+

v1.5.0 (2026-02-15)

Added: Module Skills — reusable skills that bootstrap entire features on any Laravel project

• /module:mcp (03-custom-skills/examples/module-mcp/) — Full MCP server implementation
  • Domain analysis → tool scaffolding → implementation workflow
  • Read/Write/Destructive tool patterns with proper annotations
  • Dual transport (HTTP/SSE + stdio), auth bootstrap trait, global scope fix
  • analyze, add <domain>, sync subcommands
  • Based on Agent Fleet reference implementation (61 tools, 14 domains)
• /module:assistant (03-custom-skills/examples/module-assistant/) — AI assistant chat panel
  • Livewire panel with conversation history, context awareness, resizable sidebar
  • PrismPHP tool registry with role-based access (read/write/destructive tiers)
  • Three provider strategies: cloud (native tools), Claude Code (<tool_call> loop), Codex (MCP native)
  • System prompt architecture with domain description, user context, and tool schemas
  • Based on Agent Fleet reference implementation (streaming, optimistic UI, local agents)

v1.4.0 (2026-02-14)

Added: Desktop Development Section (12-desktop-development/)

• macOS native development guide with XcodeBuildMCP (59 tools) and xclaude-plugin (87% token savings)
• Tauri v2 development guide with tauri-plugin-mcp and Rust + web frontend workflows
• Electron development guide with electron-mcp-server and main/renderer process patterns
• Platform-specific subagent definitions for macOS, Tauri, and Electron
• CLAUDE.md templates for all three platforms with customization variants
• Cross-platform /desktop-build and /desktop-test skills with auto-detection
• macOS-specific coverage: sandboxing, entitlements, notarization, AppKit bridging, window management
• Tauri-specific coverage: IPC commands, capability-based security, plugin system, CSP
• Electron-specific coverage: process isolation, contextBridge, Forge vs Builder, Playwright testing

v1.3.0 (2026-02-14)

Added: Mobile Development Section (11-mobile-development/)

• iOS development guide with XcodeBuildMCP (59 tools) and xclaude-plugin (87% token savings)
• Android development guide with JetBrains MCP and android-mcp-server (ADB)
• React Native development guide with Expo MCP (EAS Build/Update)
• Flutter development guide with Dart/Flutter MCP, Flutter MCP (50k+ packages), DCM MCP (450+ rules)
• Platform-specific subagent definitions for iOS, Android, React Native, and Flutter
• CLAUDE.md templates for all four platforms with customization variants
• Cross-platform /mobile-build and /mobile-test skills with auto-detection
• Shared mobile-mcp integration for simulator/emulator/device control

v1.2.0 (2026-02-09)

Added: Laravel MCP Integration (09-laravel-mcp-integration/)

• Laravel Boost MCP setup guide (official Laravel MCP server)
• LaraPlugins.io MCP configuration for package discovery
• CLAUDE.md template for Laravel projects (Livewire, Inertia, Filament variants)
• Laravel-specialized subagent definition with Boost MCP integration

Added: Subagents Section (10-subagents/)

• Complete subagent creation guide with YAML frontmatter format
• Tool access patterns (read-only, research, code writer, full access)
• Model selection guide (haiku/sonnet/opus by task type)
• Four example subagents: code-reviewer, laravel-specialist, debugger, test-generator

Added: Selective Deep Plan Analysis (06-advanced-patterns/)

• Tiered agent selection system (Tier 1/2/3) to prevent context overflow
• Context budget calculator formula for safe agent count
• Agent priority matrix by project type (Laravel, React, API, Data)
• Compact-between-agents and background agent patterns
• Deep plan anti-patterns section

Updated: Model compatibility to Claude 4.6 (Opus 4.6, Sonnet 4.5, Haiku 4.5) Updated: Resources with Laravel MCP, subagent docs, and community links

v1.1.0 (2026-01-22)

Added: UI/UX Development Section (08-ui-ux-development/)

• Complete UI/UX Pro Max skill documentation with 50+ styles, 21 palettes, 50 font pairings
• Dashboard workflow guide with real API integration patterns
• Browser testing guide with Chrome DevTools automation
• Security best practices (XSS prevention, CSRF protection)
• Empty state patterns and dark mode implementation
• Based on production Zonex dashboard implementation
• Token savings: 40-60% for UI/UX tasks

v1.0.0 (2026-01-04)

Initial release:

• Global optimization setup (01-06)
• Project activation workflows
• Custom skills and commands
• Research and token optimization guides
• Advanced patterns for complex scenarios