AI Collaboration Handbook

How to work effectively with clawbot. Task delegation patterns, multi-channel strategies, and expert tips for maximum productivity.

The Foundation: Clear Communication

Unlike cloud AI that only responds, clawbot executes. The difference is profound: vague instructions that generate interesting conversation with ChatGPT become destructive commands when given system access. Precision matters.

Principles of Effective Instructions

1️⃣
Be Specific About Scope

Define exactly what should be affected. AI won't ask for clarification—it will make assumptions based on context.

Good:

"Archive all emails from newsletters@*.com in my inbox from the last 7 days."

Bad:

"Clean up my inbox." (Which emails? What action? How old?)

2️⃣
Specify Triggers and Conditions

For ongoing tasks, define when the action should execute and what conditions must be true.

Good:

"Every weekday at 9 AM, if there are unread emails labeled 'urgent', send me a Telegram summary."

Bad:

"Let me know about urgent emails." (When? How? What defines urgent?)

3️⃣
Define Success Criteria

Tell the AI how to know the task succeeded. This prevents infinite loops and allows meaningful error reporting.

Good:

"Run tests until they all pass or 3 attempts fail, then report results to #dev-channel."

Bad:

"Keep running tests until they work." (How many attempts? What if they never pass?)

4️⃣
Request Confirmation for Destructive Actions

For operations that delete, modify, or publish data, build in approval steps.

Good:

"Draft a response to this customer complaint and send it to me for approval before sending."

Bad:

"Respond to customer complaints automatically." (What if AI misunderstands?)

Six Task Delegation Patterns

Every automation falls into one of six patterns. Understanding these helps you design effective workflows and set appropriate expectations for what AI can handle autonomously.

👁️
Pattern 1: Monitoring

AI watches for changes and notifies you.

The AI continuously observes a data source (logs, APIs, files, dashboards) and alerts when specific conditions occur. No changes are made—purely observational.

Use Case: Server Health Monitoring

"Watch the Prometheus /api/v1/alerts endpoint every 60 seconds. If any alert enters 'firing' state with severity=critical, send details to #ops-emergency Slack channel."

Use Case: Competitor Intelligence

"Check competitor.com/pricing daily at 2 AM. If prices change by more than 10%, screenshot the page and email me the comparison."

Best Practice

Set check intervals based on urgency: critical systems every 30-60s, business metrics hourly, trend monitoring daily.

Pattern 2: Reminders

AI sends time-based or event-triggered notifications.

Unlike simple calendar alerts, AI reminders can include context, preparation steps, or pre-fetched information to make the reminder actionable.

Use Case: Meeting Preparation

"30 minutes before any calendar event tagged 'client-meeting', send me via WhatsApp: (1) attendee names, (2) summary of last conversation from Notion, (3) open action items from Jira."

Use Case: Deadline Management

"7 days before any Jira ticket deadline where I'm the assignee, remind me via Telegram. If still incomplete 1 day before, escalate to urgent."

Best Practice

Include actionable context with reminders. "Meeting in 30 min" is useless; "Client call with Acme Corp in 30 min - they're asking about API limits" is valuable.

Pattern 3: Execution

AI performs actions on your behalf.

The AI actively changes system state: files, databases, APIs, services. This is where self-hosted infrastructure becomes essential—cloud AI can never have this level of access.

Use Case: Deployment Automation

"When a new tag matching v*.*.* is pushed to the main branch, run CI tests. If all pass, deploy to staging. After 10 minutes with no errors in Sentry, promote to production."

Use Case: Data Cleanup

"Every Sunday at 3 AM, archive Slack messages older than 90 days in #random channel. Export to JSON in ~/archives/slack/ before deleting."

Best Practice

Always test execution workflows in non-production environments first. Use --dry-run flags when available. Build in rollback mechanisms.

📊
Pattern 4: Analysis

AI processes data and extracts insights.

The AI reads large volumes of data—logs, metrics, documents, conversations—and summarizes, identifies patterns, or answers questions. No system changes, but generates valuable intelligence.

Use Case: Log Analysis

"Analyze the last 1000 lines of /var/log/nginx/error.log. Identify the 5 most common error patterns, estimate impact (requests affected), and suggest fixes."

Use Case: Meeting Intelligence

"Read transcripts from all Zoom meetings tagged 'product-review' from the past month. What are the top 5 feature requests by frequency? Who mentioned each one?"

Best Practice

Pair analysis with action. Don't just generate reports—configure AI to escalate findings that need immediate attention.

🔗
Pattern 5: Integration

AI connects multiple services and keeps them synchronized.

The AI acts as middleware, moving data between systems that don't natively integrate. This replaces Zapier-style automation but with natural language logic instead of drag-and-drop workflows.

Use Case: CRM ↔ Support Sync

"When a new ticket is created in Zendesk, check if customer exists in HubSpot. If yes, add ticket link to their timeline. If no, create contact record with ticket metadata."

Use Case: Time Tracking

"When I close a Jira ticket, calculate time from 'In Progress' to 'Done', log to Clockify under the project matching the Jira project key."

Best Practice

Design integrations to be idempotent (safe to run multiple times) and handle failures gracefully. Log all sync operations for audit trails.

🤖
Pattern 6: Automation (Multi-Step Workflows)

AI chains multiple actions based on conditional logic.

This combines all previous patterns into sophisticated workflows with branching logic, retry mechanisms, and decision trees. The AI orchestrates complex sequences autonomously.

Use Case: Incident Response

"When Prometheus fires a 'HighMemoryUsage' alert: (1) Check if this service restarted in the last hour. If yes, ignore. (2) If no, capture heap dump. (3) Analyze with gdb. (4) If memory leak detected, restart service and file Jira bug. (5) Post incident report to #ops Slack."

Use Case: Content Publishing Pipeline

"When a new markdown file appears in ~/blog/drafts/: (1) Check spelling and grammar. (2) Generate SEO metadata. (3) Optimize images in the post. (4) Move to ~/blog/ready/. (5) Create PR to website repo. (6) Post preview link to #content Slack."

Best Practice

Break complex automation into stages with checkpoints. Log state at each step so failures can be debugged and workflows can resume mid-process.

Multi-Channel Strategy

clawbot supports 15+ communication channels. The key to effective use isn't connecting them all—it's choosing the right channel for each task type based on context, urgency, and workflow integration.

WhatsApp: Personal, High-Priority

  • Critical alerts (production down, security incidents)
  • Time-sensitive reminders (meeting prep, travel check-in)
  • Personal automation (smart home, family tech support)
  • Voice commands via voice messages

Why: Always with you, native mobile notifications, voice input.

Slack/Discord: Team Collaboration

  • Deployment notifications visible to the team
  • CI/CD status updates in project channels
  • Shared automation (bot commands anyone can trigger)
  • Incident response coordination

Why: Transparency, team context, threaded discussions.

Telegram: Reliable, Fast, Flexible

  • High-frequency updates (monitoring dashboards)
  • Daily digests and reports
  • File sharing (screenshots, logs, exports)
  • Bot API for custom integrations

Why: No rate limits, inline buttons, powerful bot API.

Email: Formal, Archival, External

  • Weekly reports for stakeholders
  • Formal notifications (legal, compliance)
  • Communication with external parties
  • Long-form content with formatting

Why: Universal, formal record, rich formatting.

Pro Tip: Channel Routing Rules

Configure clawbot to automatically route different message types to appropriate channels: errors to Slack #ops, summaries to Telegram, critical alerts to WhatsApp. Define routing logic in your IDENTITY.md or channel-specific configuration.

Skills Ecosystem: Extending Capabilities

clawbot's 565+ community skills (ClawdHub) follow the AgentSkills standard. Skills provide pre-built integrations and domain-specific capabilities you can install, customize, and combine.

Discovering Skills

Skills are distributed via ClawdHub (GitHub-based marketplace). Browse by category or search by keyword:

Via CLI:

clawbot skills search kubernetes

Via Chat:

"Show me skills for monitoring PostgreSQL databases."

Popular categories: DevOps (kubectl, docker, terraform), Cloud (AWS, Azure, GCP), Productivity (calendar, email, note-taking), Development (GitHub, GitLab, CI/CD), Smart Home (Home Assistant, IoT), Data (databases, analytics, ETL).

Installing Skills

Skills install as Node.js packages with declarative capability definitions:

Install from ClawdHub:

clawbot skills install kubectl

Install from Git URL:

clawbot skills install https://github.com/user/custom-skill

Security Note

Skills run with your clawbot's permissions. Review skill source code before installing, especially for skills that execute system commands or access credentials. Prefer skills with high GitHub stars and recent updates.

Customizing Skills

Most skills expose configuration via environment variables or config files. Common customizations:

API Credentials

Store in ~/.clawbot/secrets.env with encryption enabled.

Example:

GITHUB_TOKEN=ghp_xxxxx
SLACK_BOT_TOKEN=xoxb-xxxxx

Skill-Specific Settings

Override defaults in ~/.clawbot/skills/{skill-name}/config.json

Example (kubectl skill):

{"default_namespace": "production", "timeout": 30}

Permission Controls

Restrict which actions the skill can perform via tool allowlists.

Example:

Allow "kubectl get" and "kubectl describe", deny "kubectl delete"

10 Expert-Level Usage Patterns

Advanced techniques from clawbot power users that transform it from "helpful assistant" to "indispensable infrastructure."

1️⃣
Context Awareness via IDENTITY.md

Store your preferences, work context, and domain knowledge in ~/.clawbot/IDENTITY.md. The AI reads this before every interaction, personalizing responses without repeated explanations.

What to include:

• Your role and tech stack
• Project structure and naming conventions
• Preferred communication style
• Time zone and working hours
• Key contacts and their roles
• Decision-making principles

2️⃣
Approval Workflows for Risk Management

For operations with consequences (deployments, data changes, external communications), configure approval gates. AI drafts the action, you approve via chat.

Configuration (in tool policy):

tools.exec.approval: "ask" for commands matching "rm -rf", "kubectl delete", "git push --force"

3️⃣
Feedback Loops for Continuous Improvement

After automation runs, explicitly tell the AI what worked and what didn't. It remembers in session memory and adjusts future behavior.

Feedback examples:

"That email draft was too formal—use a friendlier tone next time."
"The log analysis missed errors in stderr—check both stdout and stderr."

4️⃣
Scheduled Tasks via Gateway Cron

For recurring automation, use clawbot Gateway's cron scheduling instead of system cron. AI-generated cron jobs understand natural language and handle failures intelligently.

Natural language cron:

"Every weekday at 9 AM except holidays" → Gateway translates to cron expression and handles holiday calendar.

5️⃣
Chaining Skills into Workflows

Combine multiple skills into sequences. The AI coordinates data flow between skills without you writing integration code.

Multi-skill workflow:

"Use 'github' skill to fetch latest release notes, then 'slack' skill to post summary to #announcements, then 'twitter' skill to draft a tweet."

6️⃣
Audit Logging for Accountability

Enable comprehensive logging to track every AI action. Essential for debugging, security audits, and understanding what happened during incidents.

Log locations:

~/.clawbot/logs/gateway.log (all messages)
~/.clawbot/logs/exec.log (system commands)
~/.clawbot/logs/skills.log (skill invocations)

7️⃣
Sandbox Testing for High-Risk Automation

Before deploying automation to production systems, test in isolated environments. Use Docker containers, staging servers, or --dry-run modes.

Testing strategy:

1. Test command manually
2. Run AI automation in dry-run mode
3. Deploy to staging with real data
4. Monitor for 24-48 hours
5. Promote to production

8️⃣
Per-Channel Skill Policies

Restrict certain skills to specific channels. Example: allow destructive operations only via WhatsApp (where you're authenticated), never via public Discord.

Channel policy:

WhatsApp: all skills enabled
Slack #ops: DevOps skills only
Discord: read-only skills only
Telegram: monitoring and reporting

9️⃣
Session Continuity for Multi-Turn Tasks

For complex tasks requiring back-and-forth, start a dedicated session. The AI maintains context across messages and remembers intermediate results.

Session example:

"Start a debugging session for the payment service errors."
→ AI creates session, asks clarifying questions
→ You provide logs, AI analyzes
→ AI proposes fixes, you approve
→ AI implements and verifies
"End session and summarize what we fixed."

🔟
Model Selection for Task Type

Different AI models excel at different tasks. Configure clawbot to route intelligently: Claude for complex reasoning, GPT-4 for code, Ollama for offline/private operations.

Model routing rules:

Code reviews: GPT-4 Turbo (best at code)
Data analysis: Claude Sonnet (best at reasoning)
Simple automation: Ollama llama3 (free, local)
Sensitive data: Ollama only (never cloud)

Common Problems and Solutions

Real-world challenges encountered by clawbot users and how to solve them.

Problem: AI Executes Incorrect Commands

Cause: Ambiguous instructions or insufficient context.

Solution: Be more explicit. Use approval workflows for risky commands. Add examples of correct behavior to IDENTITY.md.

Before:

"Delete old logs"

After:

"Delete log files in /var/log/myapp/ older than 30 days with names matching *.log. Preserve the most recent 10 files regardless of age. Ask for confirmation before deleting."

Problem: Automation Stops Working Randomly

Cause: External API rate limits, credential expiration, or transient network failures.

Solution: Build in retry logic with exponential backoff. Monitor ~/.clawbot/logs/ for errors. Set up alerts when automation fails.

Problem: Too Many Notifications

Cause: Monitoring thresholds set too sensitive or no noise filtering.

Solution: Add debouncing ("Only alert if condition persists for 5 minutes") and deduplication ("Don't alert more than once per hour for same issue").

Problem: AI Misunderstands Complex Workflows

Cause: Trying to describe entire workflow in one message.

Solution: Break into stages. Configure one stage, test, then add the next. Use session continuity for incremental refinement.

Problem: Sensitive Data in Logs

Cause: AI logging full command outputs including credentials or private data.

Solution: Configure log sanitization rules to redact patterns (API keys, tokens, passwords). Review logs periodically and add new redaction rules as needed.

Ready to Master clawbot?

Start with one simple automation. As you gain confidence, layer in more patterns. Within weeks, you'll wonder how you ever worked without it.

Get Started See Real Examples