Optimize Commands

optimize-commands — Always-On Default Behavior v3.0

No command needed. Active every session from first prompt.

MULTI-LLM BURST — DEFAULT ON (every task, every session)

llm-burst = 8 models fire in parallel → judge picks best → Claude synthesizes final.

# Auto-invoked on sub-tasks. Manual use:
~/.claude/bin/llm-burst "your prompt"
~/.claude/bin/llm-burst --models groq,gemini,deepseek,g0dm0d3 "focused prompt"
~/.claude/bin/llm-burst --json "prompt" | jq '.all[] | {model,score}'  # see all scores

Model roster (all fire simultaneously — 15 models total):

DeepSeek-R1-7B / 1.5B: disabled — pre-tokenizer unsupported by gpt4all 2.8.2 llama.cpp (tokenizer type deepseek-r1-qwen unknown)

Judge scoring: completeness(25) + structure(10) + actionability(10) + length_balance(30) + G0DM0D3_bonus(+20) Winner = highest score. Top 2-3 synthesized if complementary signals found.

CODE AGENT SYSTEM — ALWAYS-ON (auto-routes every coding task, no command needed)

Activated automatically on any coding-related prompt. No /codegen or /launch-optimized required.

Auto-routing by intent:

Hard laws (enforced on every coding task):

• PLAN before BUILD — Sonnet PlanMode first, always. Skip = reject
• Opus only for critical/architectural. Never for docs, unit tests, simple bugs.
• Haiku only for extraction/parsing. Never for generation.
• Max 3 exchanges per chat — open fresh chat for next sub-task
• Cite all library/API references: [SOURCE: name|version|verified:yes/no]
• Flag unverified claims: [UNVERIFIED: claim] or [ASSUMPTION: reason]

Full framework: ~/Downloads/Claude-Code-Automation-System.md

REPORT CREATOR — ALWAYS-ON (auto-fires on any report/document task, no command needed)

Professional institutional-grade reports across PDF, DOCX, XLSX, LaTeX, Excel, Dashboards, PPT. Zero tolerance for overlap, spacing issues, or inconsistent formatting.

Auto-routing by trigger — Template Library:

Template library paths:

~/.claude/templates/
├── latex/
│   ├── nasa-latex-docs/       ← institutional/technical (★★★★★)
│   ├── corporate-latex/       ← branded corporate (★★★★★)
│   ├── heitzmann-latex/       ← academic/research (★★★★½)
│   ├── thomasbenas-report/    ← clean minimal (★★★★)
│   └── chrrel-report/         ← quick start (★★★★)
├── excel/
│   ├── sales-dashboard/       ← sales KPI (★★★★★)
│   ├── bizfin-templates/      ← business finance (★★★★)
│   ├── finance-kpi/           ← finance ops (★★★)
│   └── financial-macros/      ← VBA macros (★★★)
├── dashboards/
│   ├── aduet-dashboards/      ← multi-type dashboards (★★★★)
│   └── dynamic-kpi/           ← interactive KPIs (★★★)
└── ppt/
    └── gbif-ppt/              ← branded presentation (★★★½)

Non-negotiable quality laws (every report):

• Zero overlapping elements — text never obscures charts/images
• Pre-delivery checklist: spacing ✓ typography ✓ charts labeled ✓ colors accessible ✓ data accurate ✓
• Output always → ~/Downloads/ (never Desktop)
• Brand palette from client URL when available
• ReportLab for PDF (direct — never HTML-to-PDF) | LaTeX for institutional/technical
• Select best-fit template automatically based on report type

Full skill: ~/.claude/skills/report-creator/SKILL.md

AGENCY-AGENTS + N8N AUTO-TRIGGER — ALWAYS-ON

Auto-fires on EVERY task. No command needed. Detects intent → activates right agent + suggests n8n workflow.

Installed locations:

• Agents: ~/.claude/agents/ (210 specialists, all active)
• Workflows: ~/installed-repos/n8nworkflows.xyz/workflows/ (8,159 JSONs)
• Index: ~/installed-repos/n8nworkflows.xyz/workflow_index.txt

Keyword → Agent + n8n Auto-Routing:

How to use:

• Say "Activate [Division] mode" to load a specific agent personality
• /agency-run → orchestrator activates ALL relevant agents per division
• /all-agents → all 210 agents fire simultaneously
• grep -i '[topic]' ~/installed-repos/n8nworkflows.xyz/workflow_index.txt → find n8n workflows

Agent Divisions (210 total): Engineering(29) · Marketing(30) · Specialized(41) · Paid Media(7) · Game Dev(10) · Testing(8) · Design(8) · Sales(8) · Project Mgmt(6) · Strategy(6) · Spatial Computing(6) · Support(6) · Finance(5) · Product(5) · Academic(5)

FREE CODING MODELS — AUTO MODEL DETECTOR (always-on)

Tool: free-coding-models (installed globally at ~/.nvm/versions/node/v24.14.1/bin/free-coding-models) Purpose: Pings ~170 free AI models across 16 providers in real-time → shows live latency + Stability Score → auto-writes winner into coding tool config (OpenClaw, Aider, Goose, etc.)

Auto-detect triggers (no command needed — fires when I detect any of these needs):

16 providers tracked (170 models):

NVIDIA NIM (42)  · Groq (8)  · Cerebras (4)  · Google AI Studio (6)
GitHub Models (15) · Mistral (7) · Cloudflare Workers AI (15) · OpenRouter (31)
SambaNova (6) · OVHcloud (10) · Codestral (1) · ZAI (2)
Scaleway (10) · Alibaba DashScope (9) · Gemini CLI (6) · OpenCode Zen (8)

Keys already configured (in ~/.zshrc): GROQ_API_KEY ✓ · OPENROUTER_API_KEY ✓ · GOOGLE_API_KEY ✓ · DASHSCOPE_API_KEY ✓

Run anytime in terminal:

free-coding-models                          # interactive TUI — pick fastest model
free-coding-models --help                   # all options

Stability Score formula: p95 latency(30%) + jitter/variance(30%) + spike rate(20%) + uptime(20%) Use Stability Score, NOT raw avg latency — a model averaging 1s with 6s spikes is worse than 1.5s stable.

Auto-configure rule: When a new fastest model is found, free-coding-models writes it directly into:

• ~/.openclaw/ (CoWork config)
• ~/.config/opencode/ (OpenCode)
• ~/.aider.conf.yml (Aider)
• Any other connected coding tool

Integration with MULTI-LLM BURST: After running free-coding-models, update llm-burst default model roster with the new fastest provider for that session.

ADS CREATIVE AI STACK — AUTO-ROUTING (always-on)

Installed at: ~/installed-repos/ads-creative/

5 repos — complete AI ad creative production pipeline:

Ratings:

uni1-image-ad quick start (LUMA_API_KEY already configured ✓):

cd ~/installed-repos/ads-creative/uni1-image-ad
# Read skills/uni1-image-ad/ — Claude Code skill, just tell it:
# "Generate a Meta image ad for [brand] — use uni-1, clone structure from ad ID [X]"

meta-ads-spy quick start:

cd ~/installed-repos/ads-creative/meta-ads-spy
pip install -r requirements.txt
# Set META_AD_LIBRARY_TOKEN + AIRTABLE_API_KEY + AIRTABLE_BASE_ID
python3 discover_competitors.py  # find competitor pages
python3 pull_ads.py              # scrape their ads into Airtable

codex-plugin-cc quick start:

cd ~/installed-repos/ads-creative/codex-plugin-cc
npm install
# Copy plugins/ dir to ~/.claude/plugins/
# Then use: /codex:review  /codex:adversarial-review  /codex:rescue

MICROSOFT REPO STACK — AUTO-ROUTING (always-on)

Installed at: ~/installed-repos/microsoft/

Auto-routing table (no command needed — fires on keyword detection):

Repo ratings & use cases:

Key use cases this unlocks:

• Playwright: Browser scraping, automated testing, Playwright MCP server for Claude Code
• Playwright MCP: Wire Claude directly to a real browser — navigate, click, extract data, fill forms
• Semantic Kernel: Build production AI apps with plugins, memory, multi-step planners in Python/C#
• GenAI for Beginners: 21 lessons on prompting, RAG, fine-tuning, agents — full code included
• AI Agents for Beginners: ReAct, tool-use, multi-agent, AutoGen patterns — hands-on notebooks

playwright-mcp quick start:

cd ~/installed-repos/microsoft/playwright-mcp
npm install
# Add to ~/.claude/settings.json MCP servers section

semantic-kernel quick start:

cd ~/installed-repos/microsoft/semantic-kernel/python
pip install semantic-kernel
# Use SK for agent orchestration, memory, plugin systems

L99 + OODA — PERMANENT (every session, every task, cannot disable)

L99: Max performance mode. Full capability. No hedging. No half-measures. Best output always.

PREMIUM WEB DESIGN STACK — AUTO-ROUTING (always-on)

Skill: ~/.claude/skills/premium-web-design/SKILL.md

Auto-routing triggers:

4-tool stack:

1. Google Stitch MCP → pixel-perfect mockup from prompt
2. Nano Banana/2 → shadows, textures, depth refinement
3. GitHub UX/UI skill pack → spacing + hierarchy rules auto-applied
4. 21st.dev asset library → 3D widgets + reactive components

Rate unlock: Generic AI site = $200–500 → This workflow = $1,500–4,000

UGC AGENCY STACK — AUTO-ROUTING (always-on)

Skill: ~/.claude/skills/ugc-agency/SKILL.md Arcads API key: $ARCADS_API_KEY (set in ~/.zshrc when obtained)

Auto-routing triggers:

Pipeline (one prompt → 20 finished MP4s):

1. Claude writes 20 scripts (testimonial / DR / pattern interrupt / problem-solution / social proof)
2. Actor matching (age/vibe/energy per script)
3. Parallel render to Arcads API (all 20 simultaneously)
4. MP4s land in ~/ugc-output/ named by hook type

Agency math: $1,500/mo in · ~$150/mo tools · 90% gross margin

Setup:

# Get key: arcads.ai → API settings
export ARCADS_API_KEY="your-key"
git clone --depth=1 https://github.com/krusemediallc/arcads-claude-code ~/.claude/skills/arcads/

AIRTABLE SDK — AUTO-ROUTING (always-on)

Skill: ~/.claude/skills/airtable-sdk/SKILL.md Repo: ~/installed-repos/airtable.js/ (official Airtable JS SDK) API key: $AIRTABLE_API_KEY ✓ configured Base ID: $AIRTABLE_BASE_ID (needed — get from airtable.com/api)

Triggers: "airtable" / "airtable base" / "write to airtable" / "sync airtable" / "airtable sdk" / "airtable integration" / "meta ads to airtable"

Capabilities:

• Read/create/update/delete records (JS + Python)
• Batch operations (auto-chunks at 10 per Airtable API limit)
• Filter formulas, pagination, upsert
• Meta Ads Spy → Airtable pipeline schema built-in
• pyairtable (Python) + official airtable.js (Node) both documented

KIMI CODE — Moonshot AI CLI (always-on)

Binary: ~/.local/bin/kimi · Version: 1.41.0 Trigger: "kimi" / "kimi code" / "moonshot" / "kimi cli"

kimi           # start interactive session
kimi --help    # all commands

Tier 0 routing: Use Kimi for coding tasks as an alternative to DeepSeek/Groq — free quota, fast.

KIMI — TOKEN-MINIMIZED TIER 0 ROUTING (always-on)

Rule: Kimi K2.6 replaces Claude Opus for ALL long-context + reasoning sub-tasks. Kimi costs ~95% less than Opus. Same quality on most tasks. Always prefer Kimi first.

Auto-routing by task type

Caveman compression — applied to ALL Kimi calls automatically

Every Kimi API call includes this system prompt (hardcoded in llm-burst):

Reply ULTRA SHORT. No filler. No repeat. Bullet points only if needed.
Max 150 words unless code. Never say 'certainly' or 'here is'. Just answer.

This gives 40-60% fewer output tokens on every Kimi call regardless of model.

How to call Kimi from Claude session

# Quick (8K — cheapest)
~/.claude/bin/llm-burst --models kimi "your prompt"

# Long context (128K)
~/.claude/bin/llm-burst --models kimi-long "your prompt"

# Reasoning / Opus replacement (262K)
~/.claude/bin/llm-burst --models kimi-k2 "your prompt"

# Race Kimi K2 against Groq + Gemini — pick best
~/.claude/bin/llm-burst --models kimi-k2,groq,gemini "your prompt"

Direct Python (any script)

import openai, os
client = openai.OpenAI(
    api_key=os.environ["KIMI_API_KEY"],
    base_url="https://api.moonshot.ai/v1"
)

CAVEMAN = ("Reply ULTRA SHORT. No filler. No repeat. Max 150 words unless code. "
           "Bullet points only if needed. Just answer.")

def kimi(prompt, model="moonshot-v1-8k"):
    """Auto-compressed Kimi call — works for any model incl. future ones"""
    r = client.chat.completions.create(
        model=model,
        messages=[
            {"role": "system", "content": CAVEMAN},
            {"role": "user",   "content": prompt}
        ],
        max_tokens=1024,
        temperature=0.3,
    )
    return r.choices[0].message.content

# Usage — same pattern works for ALL current + future Kimi models:
kimi("analyze this ad copy", model="moonshot-v1-8k")      # cheap
kimi("review this codebase", model="moonshot-v1-128k")    # long
kimi("reason through this strategy", model="kimi-k2.5")   # reasoning
kimi("analyze this video ad", model="kimi-k2.6")          # vision+video
kimi("...", model="kimi-k3")                              # future models — just change model name

WEBSITE BUILDER — $10K Site From One Prompt (always-on)

Skill: ~/.claude/skills/website-builder/SKILL.md Command: /website-builder-setup (installed at ~/.claude/commands/website-builder-setup/)

Auto-routing triggers: "build website" / "build a site" / "create website" / "10k website" / "framer motion" / "21st.dev" / "ui.ux pro max" / "deploy vercel" / "next.js site" / "agency website"

5-Actor Pipeline:

1. Brief Extractor (Kimi 8K) → extract business type, audience, vibe, colors, sections
2. Design Director (UI/UX Pro Max) → style system, color palette, font pairing, design spec JSON
3. Component Selector (21st.dev MCP / Kimi K2.5) → Hero, Features, Testimonials, Pricing, CTA, Footer
4. Site Builder (Claude Code) → Next.js + Tailwind + Framer Motion, full file structure
5. QA + Deploy (Kimi K2.6 vision) → Playwright screenshot diff → auto-fix → npx vercel --prod

One-prompt format:

Build me a premium website for [BUSINESS].
Audience: [WHO]. Vibe: [STYLE]. Colors: [HEX or "choose for me"].
Sections: Hero, Features, Testimonials, Pricing, CTA, Footer.
Use Framer Motion animations. Pull components from 21st.dev.
Apply UI/UX Pro Max design system. Output: Next.js + Tailwind. Deploy to Vercel.

Repos:

• ~/installed-repos/website-builder-setup/ — setup skill
• ~/installed-repos/ui-ux-pro-max-skill-main/ — design system
• ~/.claude/commands/website-builder-setup/ — /website-builder-setup command

Agency pricing: Starter $1,500 · Growth $3,500 · Premium $8,000 · Retainer $500/mo

LEAD GEN AI — Automated Lead Extraction (always-on)

Skill: ~/.claude/skills/lead-gen-ai/SKILL.md

Auto-routing triggers: "find leads" / "lead generation" / "extract leads" / "find contacts" / "vibe prospecting" / "apollo leads" / "find [business type] in [city]" / "find emails" / "find phones" / "leads excel"

5-Actor Pipeline:

1. Brief Generator (Kimi 8K) → structured extraction prompt from rough request
2. Vibe Prospecting Extractor (MCP) → fetch-entities + enrich-prospects + export-to-csv
3. Apollo Enricher (MCP) → apollo_mixed_companies_search + apollo_people_match + apollo_contacts_search
4. Data Formatter (Python/openpyxl) → Excel with Name, Address, Phone, Email, Website, LinkedIn, Instagram, Facebook, Rating, Reviews
5. Outreach Sequence (Kimi K2.5) → personalized email + LinkedIn DM + cold DM per lead → Mailchimp CSV

Quick-start prompts:

# Extract:
"Find top 20 [BUSINESS TYPE] in [CITY]. Get phone, email, website, Instagram, Google rating. Export Excel."

# Enrich (add decision makers):
"For each business, find owner/marketing manager name + email via Apollo"

# Generate outreach:
"Write personalized cold email for each lead offering [SERVICE]. Export CSV for Mailchimp."

MCP tools wired:

• mcp__Vibe_Prospecting__fetch-entities — search businesses
• mcp__Vibe_Prospecting__enrich-prospects — add contact data
• mcp__Vibe_Prospecting__export-to-csv — download dataset
• mcp__8b8885d8-497d-4b96-be50-89e1511947c7__apollo_mixed_companies_search — find companies
• mcp__8b8885d8-497d-4b96-be50-89e1511947c7__apollo_people_match — find decision makers
• mcp__8b8885d8-497d-4b96-be50-89e1511947c7__apollo_contacts_search — filter contacts

Output: ~/Downloads/leads.xlsx — 11-column Excel, auto-width, bold headers