# Agents > AI agent design and automation skill for solopreneurs. Use when user needs to audit workflows for automation, design AI agents, or build automation recipes using n8n + Claude API. Triggers on "ai agent", "automation", "workflow automation", "n8n", "agent design", "automate my business", or /agent commands. - Skill: `orbweva/agents` (Agent Skill) - Install (CLI): `npx skillmds@latest add orbweva/agents` - Raw SKILL.md: https://api.skillmd.com/api/skills/orbweva/agents/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: ORBWEVA (https://skillmd.com/u/orbweva) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/orbweva/agents --- # AI Agent Skill Your AI-powered business operating system. Design, build, and deploy AI agents that handle the work you shouldn't be doing manually. ## Commands | Command | Description | |---------|-------------| | `/agent:audit` | Audit your workflow, rank automation opportunities by ROI | | `/agent:design` | Design an AI agent spec (trigger, workflow, tools, output) | | `/agent:build` | Step-by-step build guide (n8n + Claude API) | | `/agent:leads` | Lead qualifier agent recipe | | `/agent:support` | Customer support agent recipe | | `/agent:content` | Content repurposing agent recipe | | `/agent:inbox` | Email triage + draft agent recipe | | `/agent:social` | Social media scheduling agent recipe | | `/agent:onboard` | Client onboarding agent recipe | | `/agent:invoice` | Invoice + payment follow-up agent recipe | | `/agent:research` | Market/competitor research agent recipe | | `/agent:help` | Command reference | --- ## Principles 1. **Automate the repeatable, humanize the valuable.** If it takes judgment, taste, or relationship — keep it human. 2. **Every agent needs a human checkpoint.** No fully autonomous agent should touch clients or money without a review gate. 3. **Start with the highest-ROI automation, not the coolest one.** Boring automations that save 5 hours/week beat flashy ones that save 20 minutes. 4. **An agent that fails silently is worse than no agent.** Every agent must alert on failure. 5. **Cost per automation < cost of your time.** If the API bill exceeds the hourly rate you'd pay someone, rethink the approach. --- ## Stage 1: Automation Audit (`/agent:audit`) Before building anything, figure out what's worth automating. Most solopreneurs automate the wrong things first. ### Step 1: Task Inventory List everything you do in a typical week. Be honest — include the 10-minute tasks you think don't matter. They add up. ``` WEEKLY TASK INVENTORY ====================== | # | Task | Category | Frequency | Time/Occurrence | Total hrs/wk | |---|-----------------------------|---------------|-------------|-----------------|--------------| | 1 | [e.g., Reply to inquiries] | [Sales] | [Daily] | [15 min] | [1.25] | | 2 | [e.g., Post on LinkedIn] | [Marketing] | [3x/week] | [30 min] | [1.5] | | 3 | [e.g., Send invoices] | [Finance] | [Weekly] | [45 min] | [0.75] | | 4 | [e.g., Onboard new client] | [Operations] | [2x/month] | [2 hours] | [1.0] | | 5 | | | | | | |...| | | | | | ``` Categories: Sales, Marketing, Finance, Operations, Support, Admin, Research ### Step 2: Automation Scorecard Score each task on four dimensions. Be ruthless. ``` AUTOMATION SCORECARD ===================== | # | Task | Repetitive | Rule-Based | Error-Prone | Time (hrs/wk) | ROI Score | |---|-----------------------|------------|------------|-------------|----------------|-----------| | | | (1-5) | (1-5) | (1-5) | | | | 1 | Reply to inquiries | 4 | 3 | 2 | 1.25 | 11.25 | | 2 | Post on LinkedIn | 5 | 4 | 1 | 1.5 | 15.0 | | 3 | Send invoices | 5 | 5 | 3 | 0.75 | 9.75 | | 4 | Onboard new client | 4 | 4 | 4 | 1.0 | 12.0 | ``` **ROI Score formula:** ``` ROI Score = (Repetitive + Rule-Based + Error-Prone) x Time (hrs/wk) ``` **Scoring guide:** ``` Repetitive (1-5) 1 = Unique every time, no two instances alike 2 = Some patterns but mostly custom 3 = Half repeatable, half custom 4 = Mostly the same steps, minor variations 5 = Identical every single time Rule-Based (1-5) 1 = Requires deep judgment, creativity, or intuition 2 = Mostly judgment with some clear rules 3 = Mix of rules and judgment calls 4 = Clear rules with occasional edge cases 5 = Fully deterministic — if X then Y, always Error-Prone (1-5) 1 = Almost never make mistakes here 2 = Rare errors, low impact 3 = Occasional errors, moderate impact 4 = Frequent small errors or occasional big ones 5 = Regularly make mistakes, high impact when wrong ``` ### Step 3: Priority Matrix Plot your tasks into four quadrants: ``` HIGH ROI | STRATEGIC | QUICK WINS (plan these) | (do these first) | HARD TO ---------------+--------------- EASY TO AUTOMATE | AUTOMATE | SKIP | NICE-TO-HAVE (don't bother) | (do when bored) | LOW ROI ``` **Quick Wins** (High ROI + Easy to automate): - Tasks scoring ROI > 10 that use standard tools (email, spreadsheets, forms) - Usually: invoice reminders, social posting, email sorting, data entry **Strategic** (High ROI + Hard to automate): - Tasks scoring ROI > 10 that need custom logic or multiple integrations - Usually: lead qualification, client onboarding, content repurposing **Nice-to-Have** (Low ROI + Easy to automate): - Tasks scoring ROI < 10 that are simple to set up - Usually: file organization, notification routing, calendar management **Skip** (Low ROI + Hard to automate): - Not worth the effort. Revisit in 6 months. ### Step 4: Human-Required Flags Some tasks should never be fully automated, no matter the ROI score. Flag these: ``` HUMAN-REQUIRED FLAGS ===================== [!] Creative work — brand voice, design, strategy [!] Relationship moments — first client call, conflict resolution, celebrations [!] High-stakes decisions — pricing, hiring, partnerships, legal [!] Emotional intelligence — reading tone, detecting frustration, showing empathy [!] Novel situations — first-time problems, unusual requests ``` For flagged tasks: automate the preparation, keep the execution human. Example: AI drafts the proposal, you review and send it. ### Audit Output Deliver a ranked list: ``` AUTOMATION PRIORITIES ====================== Priority 1 (This week): [Quick win with highest ROI score] Priority 2 (This week): [Second quick win] Priority 3 (Next week): [Third quick win or first strategic] Priority 4 (This month): [Strategic automation] Priority 5 (This month): [Strategic automation] Estimated time saved: [X] hours/week Estimated build time: [Y] hours total Payback period: [Y / X] weeks ``` --- ## Stage 2: Agent Design (`/agent:design`) Every agent starts with a spec. No spec, no build. This prevents the most common failure mode: building something that solves the wrong problem. ### Agent Spec Template ``` AGENT SPEC =========== Name: [Descriptive name — e.g., "Lead Qualifier Bot"] Owner: [Who maintains this agent] Version: [Start at 1.0] TRIGGER ------- Type: [Schedule / Webhook / Email / Form / Event / Manual] Source: [Where the trigger comes from] Frequency: [How often it fires] Conditions: [Any filters before the agent runs] INPUT ----- Data: [What data the agent receives] Format: [JSON, email body, form fields, etc.] Required fields: [What must be present or agent should abort] Optional fields: [Nice-to-have data that improves output] WORKFLOW -------- Step 1: [Action — be specific] Step 2: [Action] Step 3: [Action] ... Step N: [Final action] TOOLS & INTEGRATIONS --------------------- - [Service/API]: [What it does in this workflow] - [Service/API]: [What it does in this workflow] - [Database]: [What it reads/writes] AI MODEL -------- Model: [Claude Sonnet for speed, Claude Opus for reasoning, GPT-4o for cost] Why: [Justify the choice — complexity, cost, latency requirements] Token budget: [Max input + output tokens per run] Temperature: [0.0-1.0 — lower for classification, higher for creative] OUTPUT ------ Produces: [Email, Slack message, database entry, file, API call, etc.] Destination: [Where output goes] Format: [Structure of the output] HUMAN CHECKPOINT ----------------- Review point: [Where a human reviews before the agent continues] Reviewer: [Who reviews — you, VA, team member] SLA: [How quickly must the human respond] Auto-proceed: [If no review in X hours, what happens — wait, proceed, alert] FALLBACK -------- On error: [Retry X times, then alert via [channel]] On uncertainty: [If AI confidence < threshold, route to human] On missing data: [Request missing info / skip / use defaults] SUCCESS METRICS ---------------- - [Metric 1]: [Target — e.g., "Process 95% of leads within 5 minutes"] - [Metric 2]: [Target — e.g., "Reduce manual lead sorting from 5hrs to 30min/week"] - [Metric 3]: [Target — e.g., "Zero missed high-priority leads"] COST ESTIMATE -------------- API calls/month: [Volume estimate] Cost/call: [Based on model + token usage] Monthly total: [$X] Break-even: [Hours saved × hourly rate > monthly cost] ``` ### Trigger Types Choose the right trigger for the job: ``` TRIGGER TYPE GUIDE =================== | Type | Use When | n8n Node | Example | |----------------|---------------------------------------------|-------------------|----------------------------| | Schedule | Task runs at fixed intervals | Cron / Schedule | Daily social posts | | Webhook | External event sends data to your agent | Webhook | Form submission | | Email | Incoming email triggers workflow | Email Trigger | Support request | | Polling | Check a source periodically for changes | HTTP Request+Cron | New RSS item | | Event | Database or app event fires | App Trigger | New Stripe payment | | Manual | You explicitly kick it off | Manual Trigger | Research request | | Threshold | Metric crosses a boundary | IF + Schedule | Lead score > 80 | | Chain | Another agent's output triggers this one | Execute Workflow | Lead qualified → onboard | ``` ### Tool Selection Guide Common n8n nodes for solopreneur agents: ``` TOOL SELECTION =============== | Need | n8n Node(s) | Notes | |-------------------------|------------------------------------|--------------------------------| | AI reasoning | HTTP Request (Claude API) | System prompt + user message | | Email send | Gmail / SMTP | Use OAuth for Gmail | | Email receive | Email Trigger (IMAP) | Poll interval matters | | Spreadsheet | Google Sheets | Read, append, update | | CRM | HTTP Request to CRM API | Or native node if available | | Slack/Discord | Slack / Discord node | Webhooks for incoming | | Database | Postgres / Supabase | Structured data storage | | Web scraping | HTTP Request + HTML Extract | Respect robots.txt | | File generation | Code node (JS) | PDF, CSV, etc. | | Payment | Stripe node | Invoices, subscriptions | | Calendar | Google Calendar | Create, read, update events | | Form data | Webhook + Typeform/Tally | Webhook is most flexible | | Social media | HTTP Request to platform APIs | LinkedIn, Twitter, etc. | | Project management | ClickUp / Notion / Trello node | Task creation and updates | ``` ### Prompt Engineering for Agents Agent prompts are different from chat prompts. They must be deterministic, structured, and failure-aware. ``` AGENT PROMPT TEMPLATE ====================== You are [agent name], an AI assistant that [one-sentence purpose]. CONTEXT: - You work for [business name], a [business type] - Your role is to [specific function] - You operate as part of an automated workflow INSTRUCTIONS: 1. [First thing to do with the input] 2. [Second thing to do] 3. [Third thing to do] OUTPUT FORMAT: Return your response as JSON with this exact structure: { "classification": "[category]", "confidence": [0.0-1.0], "summary": "[brief summary]", "action": "[recommended action]", "details": "[additional context]" } RULES: - If confidence < 0.7, set action to "HUMAN_REVIEW" - Never fabricate information — if data is missing, say so - Keep summaries under 100 words - [Domain-specific rules] EXAMPLES: Input: [example input] Output: [example output] ``` ### Testing Checklist Before deploying any agent: ``` TESTING CHECKLIST ================== [ ] Happy path — standard input produces correct output [ ] Empty input — agent handles gracefully (doesn't crash or hallucinate) [ ] Malformed input — agent rejects or flags for review [ ] Edge case: very long input — stays within token limits [ ] Edge case: ambiguous input — routes to human review [ ] Edge case: duplicate input — doesn't process twice [ ] Failure mode: API timeout — retries then alerts [ ] Failure mode: API error — logs error, alerts owner [ ] Failure mode: downstream service down — queues for retry [ ] Cost check: run 100 test inputs, verify cost matches estimate [ ] Speed check: end-to-end latency acceptable for use case [ ] Human checkpoint: review gate works, notifications arrive ``` ### Cost Estimation ``` COST CALCULATOR ================ Model: Claude Sonnet 4 Input: $3.00 / 1M tokens Output: $15.00 / 1M tokens Model: Claude Opus 4 Input: $15.00 / 1M tokens Output: $75.00 / 1M tokens Model: GPT-4o Input: $2.50 / 1M tokens Output: $10.00 / 1M tokens Model: GPT-4o-mini Input: $0.15 / 1M tokens Output: $0.60 / 1M tokens Formula: Monthly cost = (avg_input_tokens × input_price + avg_output_tokens × output_price) × runs_per_month Example — Lead Qualifier (Claude Sonnet): ~800 input tokens + ~400 output tokens per lead 100 leads/month = (800 × $0.000003 + 400 × $0.000015) × 100 = ($0.0024 + $0.006) × 100 = $0.84/month Rule of thumb: Most solopreneur agents cost $1-20/month in API fees. If your estimate exceeds $50/month, consider: - Using a smaller model for simple classification - Caching repeated queries - Batching requests - Adding a rule-based pre-filter before the AI step ``` --- ## Stage 3: Build Guide (`/agent:build`) Practical guide to building agents with n8n and the Claude API. ### n8n Setup **Self-hosted (recommended for solopreneurs):** ```yaml # docker-compose.yml version: '3.8' services: n8n: image: n8nio/n8n:latest restart: unless-stopped ports: - "5678:5678" environment: - N8N_HOST=n8n.yourdomain.com - N8N_PORT=5678 - N8N_PROTOCOL=https - WEBHOOK_URL=https://n8n.yourdomain.com/ - N8N_ENCRYPTION_KEY=your-random-encryption-key - N8N_BASIC_AUTH_ACTIVE=true - N8N_BASIC_AUTH_USER=admin - N8N_BASIC_AUTH_PASSWORD=your-secure-password volumes: - n8n_data:/home/node/.n8n volumes: n8n_data: ``` **Cloud (n8n.io):** - Starter plan: $24/month, 5 active workflows - Pro plan: $60/month, 50 active workflows - Good for getting started, migrate to self-hosted when you hit limits ### Claude API Integration in n8n Use an HTTP Request node to call the Claude Messages API: ``` METHOD: POST URL: https://api.anthropic.com/v1/messages HEADERS: x-api-key: {{ $credentials.anthropicApiKey }} anthropic-version: 2023-06-01 content-type: application/json BODY (JSON): { "model": "claude-sonnet-4-20250514", "max_tokens": 1024, "system": "Your system prompt here", "messages": [ { "role": "user", "content": "{{ $json.inputData }}" } ] } ``` **Credential setup:** 1. In n8n, go to Credentials > Add Credential > Header Auth 2. Name: `Anthropic API` 3. Header Name: `x-api-key` 4. Header Value: your Claude API key ### Common n8n Patterns **Pattern 1: Webhook > Process > Respond** Best for: form submissions, external events, API integrations ``` [Webhook] → [Set/Transform] → [HTTP Request (Claude)] → [Parse JSON] → [IF branch] → [Action] | → [Human Review] ``` **Pattern 2: Schedule > Fetch > Transform > Send** Best for: recurring reports, social posting, monitoring ``` [Cron] → [HTTP Request (fetch data)] → [HTTP Request (Claude)] → [Parse JSON] → [Gmail/Slack/Sheets] ``` **Pattern 3: Email Trigger > AI Classify > Route > Act** Best for: inbox triage, support tickets, lead capture ``` [Email Trigger] → [HTTP Request (Claude classify)] → [Switch node] |→ "urgent" → [Slack alert + Draft reply] |→ "client" → [Add to CRM + Draft reply] |→ "spam" → [Archive] |→ "other" → [Label + Skip] ``` ### Error Handling Every production agent needs error handling: ``` ERROR HANDLING PATTERN ======================= 1. Wrap AI calls in a try/catch (n8n Error Trigger node) 2. On API error (429, 500, 503): - Wait 30 seconds - Retry up to 3 times - If still failing, send alert to Slack/email 3. On malformed AI response: - Log the raw response - Route to human review - Don't retry (probably a prompt issue) 4. On downstream service failure: - Queue the action for retry - Alert if queue grows beyond threshold 5. Dead letter queue: - After all retries exhausted, store failed items in a "failed" sheet/table - Review weekly ``` n8n settings for retry: - On the HTTP Request node: Settings > Retry on Fail > Max Retries: 3, Wait Between: 30000ms - Add an Error Trigger workflow for unhandled failures ### Monitoring and Logging ``` MONITORING SETUP ================= 1. Execution log: - n8n stores execution history (Settings > Executions) - Set retention to 30 days minimum - Enable "Save failed executions" always 2. Success/failure tracking: - Add a Google Sheets "Agent Log" with columns: | Timestamp | Agent | Status | Input Summary | Output Summary | Cost | - Append a row at the end of every workflow run 3. Alerting: - Error Trigger workflow → Slack/email notification - Include: agent name, error message, input data, timestamp 4. Weekly review: - Check execution counts vs expected - Review failed executions - Check API costs in provider dashboard - Verify human checkpoints are being actioned ``` ### Cost Management ``` COST MANAGEMENT STRATEGIES ============================ 1. Model routing: - Use GPT-4o-mini or Claude Haiku for simple classification ($0.15-0.80/1M input) - Use Claude Sonnet for medium complexity ($3/1M input) - Reserve Claude Opus for complex reasoning ($15/1M input) 2. Caching: - If the same input produces the same output, cache it - Use a Google Sheet or database as a simple cache - Check cache before calling AI 3. Pre-filtering: - Use rule-based logic (IF nodes) before AI nodes - Only send to AI what actually needs AI - Example: filter spam by sender domain before classifying content 4. Batching: - Collect items over an hour, process in batch - One API call with 10 items < 10 separate API calls - Use n8n's "Wait" node or item batching 5. Token optimization: - Keep system prompts concise (every token costs money) - Limit max_tokens to what you actually need - Use structured output (JSON) to reduce response length Monthly cost monitoring: - Set a budget alert in your API provider dashboard - Track cost per agent per month in your Agent Log sheet - If any agent exceeds $20/month, review and optimize ``` --- ## Stage 4: Lead Qualifier Agent (`/agent:leads`) ``` AGENT: Lead Qualifier ====================== USE CASE: Automatically score and route incoming leads so you respond to hot prospects in minutes, not hours. DIFFICULTY: Medium BUILD TIME: 3-4 hours MONTHLY COST: $1-5 (based on 50-200 leads/month) TRIGGER: New form submission (webhook) or email inquiry SOURCE: Website contact form, Typeform, Tally, or email INPUT: - Name - Email address - Company (if provided) - Message / inquiry text - Source (which form or channel) WORKFLOW: Step 1: Receive form data via webhook Step 2: Extract and normalize fields Step 3: Enrich — look up company website/LinkedIn (optional HTTP request) Step 4: Send to Claude for ICP scoring and classification Step 5: Route based on score: - Hot (80-100): Immediate Slack alert + personalized auto-reply - Warm (50-79): Add to CRM + template reply within 24hrs - Cold (20-49): Add to nurture sequence - Spam (0-19): Archive, no reply Step 6: Log to tracking spreadsheet n8n NODES NEEDED: - Webhook: Receives form submission - Set: Normalize and clean input fields - HTTP Request: (Optional) Enrich from LinkedIn/Clearbit - HTTP Request: Claude API call for scoring - Code: Parse Claude response JSON - Switch: Route by score tier (hot/warm/cold/spam) - Gmail: Send auto-reply (hot leads) - Google Sheets: Log lead + score + classification - Slack: Alert for hot leads CLAUDE PROMPT: You are a lead qualification assistant for [YOUR BUSINESS]. MY IDEAL CUSTOMER PROFILE (ICP): - [Industry/niche] - [Company size or revenue range] - [Role/title of decision maker] - [Pain points we solve] - [Budget range] SCORING CRITERIA: - ICP match (0-30 points): How well does this lead match my ideal customer? - Intent signal (0-30 points): How strong is their buying intent based on their message? - Budget signal (0-20 points): Any indication they can afford our services? - Urgency (0-20 points): How time-sensitive is their need? Analyze this lead and return JSON: { "score": [0-100], "tier": "hot|warm|cold|spam", "confidence": [0.0-1.0], "icp_match": { "score": [0-30], "reasoning": "..." }, "intent": { "score": [0-30], "reasoning": "..." }, "budget": { "score": [0-20], "reasoning": "..." }, "urgency": { "score": [0-20], "reasoning": "..." }, "summary": "[1-2 sentence summary of this lead]", "suggested_reply_tone": "enthusiastic|professional|exploratory|decline", "talking_points": ["point 1", "point 2", "point 3"] } OUTPUT: - Slack notification (hot leads only) - Auto-reply email (hot leads, personalized) - CRM entry or spreadsheet row (all leads) - Weekly lead summary report HUMAN CHECKPOINT: - Hot leads: Slack alert with score + summary. You decide whether to call immediately. - Warm leads: Review batch of draft replies before sending (daily, 5-minute task). - Auto-replies for hot leads can be sent automatically if confidence > 0.9. SUCCESS METRICS: - Response time to hot leads < 10 minutes (vs hours before) - Lead-to-meeting conversion rate improvement - Zero missed hot leads - Time spent on lead sorting: < 30 min/week (vs 3-5 hours before) COMMON PITFALLS: - ICP too vague — be specific. "Small businesses" is useless. "SaaS companies, 10-50 employees, Series A, looking for design services" is actionable. - Auto-reply sounds robotic — write the template in your voice, use merge fields for personalization, keep it short. - Not updating ICP — review and adjust scoring criteria monthly based on which leads actually converted. - Scoring drift — periodically compare AI scores to your gut feeling. Recalibrate the prompt if they diverge. ``` **Auto-Reply Templates by Tier:** ``` HOT LEAD AUTO-REPLY ==================== Subject: Re: {{ lead.subject || "Your inquiry" }} Hi {{ lead.firstName }}, Thanks for reaching out. {{ talking_points[0] }} I'd love to learn more about what you're working on. Are you free for a quick 15-minute call this week? Here's my calendar: [CALENDAR_LINK] {{ your_name }} WARM LEAD REPLY (draft for review) ==================================== Subject: Re: {{ lead.subject || "Your inquiry" }} Hi {{ lead.firstName }}, Thanks for getting in touch. I read through your message and {{ talking_points[0] }}. I've put together some thoughts on how we might help — would you be open to a brief conversation? {{ your_name }} ``` --- ## Stage 5: Customer Support Agent (`/agent:support`) ``` AGENT: Customer Support Triage ================================ USE CASE: Classify support requests, draft responses for common issues, and escalate complex ones — so you're not starting from blank every time. DIFFICULTY: Medium BUILD TIME: 4-5 hours MONTHLY COST: $2-10 (based on 100-500 tickets/month) TRIGGER: New support email received (IMAP trigger) or form submission SOURCE: Support email inbox or help desk form INPUT: - Sender email - Subject line - Email body - Attachments (flagged, not processed) - Previous conversation history (if available) WORKFLOW: Step 1: Receive email via IMAP trigger or webhook Step 2: Check if sender is existing client (lookup in CRM/spreadsheet) Step 3: Send to Claude for classification and draft response Step 4: Route based on classification: - FAQ / Known issue: Auto-draft response → human approval queue - Bug report: Create ticket in project management tool → acknowledge - Feature request: Log in feature request tracker → acknowledge - Billing: Route to billing workflow → flag as priority - Complaint: IMMEDIATE Slack alert → do not auto-respond - Unclear: Route to human with AI summary Step 5: Log interaction in support tracker n8n NODES NEEDED: - Email Trigger (IMAP): Receives support emails - Google Sheets: Client lookup - HTTP Request: Claude API call - Code: Parse response, extract classification - Switch: Route by category - Gmail: Send drafted reply (after approval) - Slack: Alert for complaints and urgent issues - ClickUp/Notion: Create tickets for bugs/features - Google Sheets: Log all interactions CLAUDE PROMPT: You are a customer support assistant for [YOUR BUSINESS]. KNOWLEDGE BASE: - [Common FAQ 1: question and answer] - [Common FAQ 2: question and answer] - [Common FAQ 3: question and answer] - [Known issue 1: description and workaround] - [Known issue 2: description and workaround] CLIENT STATUS: {{ clientStatus || "unknown" }} CLASSIFY this support request into ONE of: - faq (matches a known question/answer) - bug (something is broken) - feature_request (wants something new) - billing (payment, invoice, refund related) - complaint (unhappy, frustrated, threatening) - unclear (can't determine intent) RESPOND with JSON: { "category": "faq|bug|feature_request|billing|complaint|unclear", "confidence": [0.0-1.0], "sentiment": "positive|neutral|negative|angry", "priority": "low|medium|high|urgent", "summary": "[1-sentence summary of the issue]", "draft_response": "[Empathetic, helpful draft reply]", "internal_notes": "[Notes for the team about this ticket]", "needs_human": true|false, "reason_for_human": "[Why this needs human attention, if applicable]" } RULES: - Always be empathetic in draft responses - Never promise timelines you can't keep - If sentiment is "angry", ALWAYS set needs_human to true - If confidence < 0.7, set needs_human to true - Reference knowledge base answers when applicable - Keep draft responses under 150 words OUTPUT: - Draft reply in approval queue (Google Sheet or Slack) - Ticket in project management tool (bugs, features) - Slack alert (complaints, urgent issues) - Logged interaction in support tracker HUMAN CHECKPOINT: - ALL draft replies go through approval queue before sending - Complaints: immediate Slack alert, never auto-respond - Exception: After 2 weeks of reviewing, you can enable auto-send for FAQ responses with confidence > 0.95 ESCALATION RULES: Priority Urgent → Slack alert within 1 minute Priority High → Slack alert within 15 minutes Priority Medium → Batch review twice daily Priority Low → Batch review daily SUCCESS METRICS: - First response time < 30 minutes (business hours) - Correct classification rate > 90% - Draft response acceptance rate > 80% (you send without major edits) - Customer satisfaction maintained or improved - Time spent on support: reduced by 60% COMMON PITFALLS: - Knowledge base too small — start with top 20 questions, expand weekly - Auto-responding to complaints — never. Always human. - Ignoring sentiment — a simple FAQ answer delivered to an angry customer needs a different tone than the same answer to a curious one - Not updating the knowledge base — add new FAQs as you discover them ``` --- ## Stage 6: Content Repurposing Agent (`/agent:content`) ``` AGENT: Content Repurposer =========================== USE CASE: Turn one piece of content into platform-specific posts for LinkedIn, Twitter/X, email newsletter, and Instagram — so one blog post becomes a week of content. DIFFICULTY: Easy BUILD TIME: 2-3 hours MONTHLY COST: $1-3 (based on 4-8 posts/month) TRIGGER: New blog post published (RSS feed or webhook) or manual trigger SOURCE: Blog RSS feed, CMS webhook, or manual URL input INPUT: - Blog post title - Blog post full text - Blog post URL - Author name - Key takeaways (optional, AI will extract if not provided) WORKFLOW: Step 1: Fetch new blog post content (RSS or webhook payload) Step 2: Send full text to Claude for analysis and repurposing Step 3: Generate platform-specific content: - LinkedIn post (professional, insight-led) - Twitter/X thread (punchy, hook-driven) - Email newsletter snippet (personal, value-first) - Instagram caption (visual-friendly, hashtag-ready) Step 4: Add all drafts to review spreadsheet Step 5: Notify via Slack that new content is ready for review Step 6: (Optional) After approval, queue in social scheduler n8n NODES NEEDED: - RSS Feed Read / Webhook: Content trigger - HTTP Request: Fetch full post content (if RSS only gives summary) - HTTP Request: Claude API call - Code: Parse and split platform-specific outputs - Google Sheets: Store drafts for review - Slack: Notify that content is ready - (Optional) HTTP Request: Post to Buffer/Hootsuite/LinkedIn API CLAUDE PROMPT: You are a content repurposing specialist for [YOUR BUSINESS]. BRAND VOICE: [Describe your tone — e.g., "Professional but approachable. Uses concrete examples. Avoids jargon. First-person perspective."] Given this blog post, create platform-specific content: BLOG POST: Title: {{ title }} URL: {{ url }} Content: {{ content }} Return JSON with: { "key_takeaways": ["takeaway 1", "takeaway 2", "takeaway 3"], "linkedin": { "post": "[LinkedIn post, 150-300 words. Start with a hook. Use line breaks for readability. End with a question or CTA. Include the blog URL.]", "hashtags": ["#tag1", "#tag2", "#tag3"] }, "twitter": { "thread": [ "1/ [Hook tweet — the most interesting insight, max 280 chars]", "2/ [Supporting point or example]", "3/ [Another insight]", "4/ [Practical takeaway]", "5/ [CTA + link to full post]" ] }, "email": { "subject_line": "[Compelling email subject, max 60 chars]", "preview_text": "[Preview text, max 90 chars]", "body": "[2-3 paragraph email snippet. Personal tone. Tease the key insight, link to full post.]" }, "instagram": { "caption": "[Instagram caption. Conversational. Include CTA. Max 200 words.]", "hashtags": ["#tag1", "#tag2", "...up to 15 relevant hashtags"] } } RULES: - Each platform version should feel native, not copy-pasted - LinkedIn: professional, thought-leadership tone - Twitter: punchy, quotable, thread-friendly - Email: personal, like writing to a friend who's interested in this topic - Instagram: visual-first language, emoji-light, community-focused - Never start LinkedIn posts with "I'm excited to announce" or "Thrilled to share" - Never use "In today's fast-paced world" or similar cliches OUTPUT: - Google Sheet row per blog post with columns for each platform - Slack notification: "New content batch ready for review: [post title]" HUMAN CHECKPOINT: - Review all drafts before posting (spreadsheet or Slack approval) - Typical review time: 10-15 minutes per blog post batch - After refining the prompt for your voice, approval becomes mostly rubber-stamping SUCCESS METRICS: - One blog post → 4 platform-specific pieces within 10 minutes (vs 2-3 hours manually) - Draft acceptance rate > 75% (posted with minor or no edits) - Consistent posting frequency (no more "I forgot to post this week") - Engagement rate maintained or improved vs manually written posts COMMON PITFALLS: - Generic brand voice description — be specific. Share 3-5 example posts you've written that nail your voice. Put them in the prompt as examples. - Too many hashtags on LinkedIn — 3-5 max. More looks spammy. - Identical content across platforms — the whole point is adaptation, not copy-paste. - Not including the blog link — every piece should drive traffic back to the source. ``` --- ## Stage 7: Email Triage Agent (`/agent:inbox`) ``` AGENT: Email Triage ===================== USE CASE: Classify, prioritize, and draft responses for incoming email so you spend 15 minutes on email instead of 2 hours. DIFFICULTY: Medium BUILD TIME: 3-4 hours MONTHLY COST: $3-15 (based on 200-1000 emails/month) TRIGGER: New email received (IMAP polling, every 5 minutes) SOURCE: Primary business inbox INPUT: - Sender name and email - Subject line - Email body (plain text, stripped of signatures) - Thread history (if reply) - Attachments (flagged, not content-analyzed) WORKFLOW: Step 1: Receive email via IMAP trigger Step 2: Skip if sender is in ignore list (newsletters, notifications) Step 3: Check if sender is known (client, prospect, vendor lookup) Step 4: Send to Claude for classification and priority scoring Step 5: Route based on classification: - Client (urgent): Slack alert + draft reply → top of queue - Client (normal): Draft reply → daily review queue - Prospect: Route to lead qualifier agent - Vendor/partner: Categorize → weekly review batch - Administrative: Auto-label → handle in batch - Spam/irrelevant: Archive Step 6: Apply Gmail labels based on classification Step 7: Log in email tracker spreadsheet n8n NODES NEEDED: - Email Trigger (IMAP): Poll inbox - IF: Filter ignore list - Google Sheets: Contact lookup (client/prospect/vendor) - HTTP Request: Claude API call - Code: Parse classification - Switch: Route by category - Gmail: Apply labels, create drafts - Slack: Alert for urgent items - Google Sheets: Log all classified emails CLAUDE PROMPT: You are an email triage assistant for [YOUR NAME] at [YOUR BUSINESS]. KNOWN CONTACTS: {{ contactContext }} BUSINESS CONTEXT: - I run [business description] - My current priorities are: [list 2-3 current priorities] - Active projects: [list active client projects] CLASSIFY this email: From: {{ senderName }} <{{ senderEmail }}> Subject: {{ subject }} Body: {{ body }} Thread: {{ threadHistory || "New conversation" }} Return JSON: { "category": "client|prospect|vendor|admin|personal|spam", "priority": "urgent|high|medium|low", "confidence": [0.0-1.0], "summary": "[1-sentence summary — what do they want?]", "requires_response": true|false, "response_deadline": "[today|this week|no rush|never]", "suggested_label": "[Gmail label to apply]", "draft_response": "[Draft reply if response needed, null otherwise]", "action_items": ["item 1", "item 2"], "sentiment": "positive|neutral|negative|urgent" } RULES: - Emails from known clients are ALWAYS at least "high" priority - Emails mentioning "invoice", "payment", or "overdue" are ALWAYS "urgent" - Emails from unknown senders with buying signals → category: "prospect" - Newsletter, marketing, notification emails → category: "spam", priority: "low" - If uncertain about category, default to "admin" with needs_human: true - Draft responses should match my communication style: [brief, professional, friendly] - Keep draft responses under 100 words unless the topic requires detail OUTPUT: - Gmail labels applied automatically - Draft replies created in Gmail - Slack alerts for urgent items - Daily summary: "You have X emails to review — Y urgent, Z drafts ready" HUMAN CHECKPOINT: - All draft replies sit in Gmail Drafts until you review and send - Urgent client emails: Slack notification so you can respond personally - Never auto-send on behalf of the user - Daily 15-minute email review: check drafts, send/edit, handle urgent SUCCESS METRICS: - Email processing time: < 15 min/day (vs 1-2 hours) - Zero missed urgent emails - Draft acceptance rate > 70% - All client emails responded to within 4 business hours - Inbox stays under 20 unprocessed emails at end of day COMMON PITFALLS: - Overly aggressive spam filtering — legitimate cold outreach may be valuable - Not updating the ignore list — new newsletter subscriptions slip through - Context gap — the agent doesn't know about a recent phone call or meeting. Add a "recent context" field you update weekly. - Reply-all disasters — never auto-send. Drafts only. ``` --- ## Stage 8: Social Media Agent (`/agent:social`) ``` AGENT: Social Media Scheduler ================================ USE CASE: Generate and schedule social media posts from your content calendar so you never miss a posting day. DIFFICULTY: Easy BUILD TIME: 2-3 hours MONTHLY COST: $1-5 (based on 20-60 posts/month) TRIGGER: Scheduled — runs daily at 7 AM or weekly on Monday SOURCE: Content calendar (Google Sheet) or content repurposer output INPUT: - Content calendar entries for the current period - Past post performance data (optional, for optimization) - Brand voice guidelines - Platform-specific requirements WORKFLOW: Step 1: Cron trigger fires at scheduled time Step 2: Read content calendar from Google Sheet Step 3: Filter for posts due today/this week Step 4: For ea …(truncated)