Bernard Bootstrap Kit

Spin up a battle-tested OpenClaw agent in minutes, not months.

This repo contains the distilled operational DNA from a production OpenClaw agent (Mr. Bernard) that's been running since February 2026. It encodes hundreds of hours of hard-won lessons into a reusable bootstrap that any new agent can inherit.

What This Is

A bootstrap kit that takes a fresh OpenClaw install and gives it:

• Workspace structure — organized memory system, heartbeat config, project scaffolding
• Operational playbook — rules for memory discipline, communication, error recovery, security
• Anti-patterns — every mistake already made so your agent doesn't repeat them
• Persona templates — starting points for SOUL.md, IDENTITY.md, AGENTS.md (customize to your taste)
• Skills manifest — which skills to install first and how to configure them

What This Isn't

• Not a copy of Mr. Bernard (you'll build your own persona)
• Not a replacement for OpenClaw itself (you need OpenClaw installed first)
• Not magic — your agent still needs to learn and grow, but it starts from a much better foundation

Quick Start

Option A: Manual Setup (You Have a VPS Already)

git clone https://github.com/<your-username>/bernard-bootstrap.git
cd bernard-bootstrap
bash setup.sh

Then customize:

1. Edit SOUL.md — give your agent a personality
2. Edit USER.md — tell it about yourself
3. Edit IDENTITY.md — name, backstory, operational persona
4. Read the playbook files in playbook/ for best practices

Option B: Full VPS Provisioning (Start from Scratch)

If you need to spin up a fresh VPS and install everything:

# On your local machine or existing server:
bash provision/provision-vps.sh <VPS_IP> <SSH_KEY_PATH>

This will:

• Create the openclaw user with proper permissions
• Install Node.js, OpenClaw, and dependencies
• Run the bootstrap setup
• Configure systemd services
• Set up firewall basics

See provision/README.md for cloud provider-specific instructions (Hetzner, DigitalOcean, etc.).

Structure

bernard-bootstrap/
├── README.md                    ← You are here
├── setup.sh                     ← One-shot workspace setup
├── provision/
│   ├── README.md                ← VPS provisioning guide
│   ├── provision-vps.sh         ← Automated VPS setup script
│   └── setup-telegram-bot.md   ← Telegram bot creation guide
├── templates/
│   ├── AGENTS.md                ← Workspace rules, memory habits, efficiency principles
│   ├── SOUL.md                  ← Personality and core values
│   ├── IDENTITY.md              ← Name, backstory, operational persona
│   ├── USER.md                  ← Template for describing your human
│   ├── HEARTBEAT.md             ← Periodic check configuration
│   └── TOOLS.md                 ← Service and integration notes
├── playbook/
│   ├── memory-discipline.md         ← The #1 thing to get right
│   ├── security.md                  ← Access control, secrets, API keys, Cloudflare, watchdogs
│   ├── communication.md             ← Speaking, routing, voice, social media, email outreach
│   ├── error-recovery.md            ← Gateway, SSH, services, credits, webhooks, remote machines
│   ├── anti-patterns.md             ← Every mistake already made (30+ entries)
│   ├── subagent-orchestration.md    ← Fan-out/synthesis, pipeline state, cost awareness
│   └── web-services.md             ← Dashboards, auth, APIs, systemd, nginx, Playwright
└── skills/
    └── manifest.md                  ← 5-phase skill installation guide

Key Principles

These are the most important lessons from production:

1. Write-through memory — If you learn something, write it to a file immediately. "Mental notes" don't survive restarts.
2. Machine-to-machine first — If a task can be a bash script or cron job, don't route it through an LLM. Reserve tokens for judgment calls.
3. Scripts > LLM for plumbing — Health checks, file syncs, API calls, notifications = pure scripts. Zero LLM involvement.
4. System crontab for deterministic tasks — OpenClaw cron isolated sessions have a known bug where tools may not be provided to models. Use crontab -e for anything that doesn't need LLM judgment.
5. Never auto-harden — Always propose security changes and get approval. Never lock yourself out.
6. Fallback within tier — If your primary API key runs out, fall back to another key on the same provider. Never fall back to free/incapable models.
7. Checkpoint before iterating — Git commit before making changes. Tag important versions. Cheap insurance.
8. Groups see only final output — No progress updates, no system messages, no errors. Deliver the finished product.
9. Build a hallucination watchdog — Automate checks for stale file paths, dead URLs, missing env vars. Run daily. Your docs WILL drift from reality.
10. Pipelines are autonomous after trigger — Multi-step workflows auto-continue. Never make your human manually kick the next stage.

Changelog

v3.1 (2026-02-28)

• NEW: subagent-orchestration.md — fan-out/synthesis, pipeline state, model selection, cost awareness
• NEW: web-services.md — vanilla JS dashboards, URL-param auth, SSE, systemd, nginx, Playwright
• Anti-patterns +8: headless vs Cloudflare, CSS/GSAP conflicts, networkidle hangs, systemd PATH, Bot API webhooks, HEIC speed, C compiler deps, env var typos
• Error recovery +5: port conflicts, credit exhaustion, webhook debugging, remote SSH, subagent timeouts
• Security +3: API key management, Cloudflare tips, hallucination watchdog
• Communication +3: voice messages, social media/Twitter, email outreach
• Skills manifest v2: 5 phases, custom skills section, TTS options, himalaya tips
• Templates updated: AGENTS.md (cost management), TOOLS.md (expanded sections)
• Key principles +2: hallucination watchdog, autonomous pipelines

v3 (2026-02-27)

• Added provision/ directory with automated VPS provisioning script
• Added Telegram bot setup guide
• Added SMS/voice channel handling patterns to AGENTS.md template
• Added emergency contact protocol pattern to AGENTS.md template
• Added heartbeat vs cron decision guide
• Updated anti-patterns with subagent code injection, model size failures
• Updated skills manifest with latest recommendations
• Setup script now supports --provision flag for full VPS setup

v2 (2026-02-26)

• Added machine-to-machine efficiency principle to AGENTS.md
• Added pipeline auto-continuation rules
• Added routing rule (reply goes to sender, not human)
• Added cron isolated session tools bug to anti-patterns
• Added free model fallback chain anti-pattern
• Added subagent code injection anti-pattern
• Added secrets-in-git anti-pattern
• Added field name mismatch anti-pattern
• Added overwriting-without-checkpoint anti-pattern
• Added system vs OpenClaw cron guidance
• Added API credit/auth recovery to error-recovery
• Added cron job debugging to error-recovery
• Added restart context pattern to memory-discipline
• Added MEMORY.md security rule (main session only)
• Added channel parity rule to security
• Added emergency contact protocol to security
• Added routing rule to communication
• Updated heartbeat template with tips and memory maintenance

v1 (2026-02-22)

• Initial release: templates, playbook, setup script, skills manifest