openclaw-config

Production-tested configs, scripts, and workspace templates for OpenClaw.
Not an SDK — just copyable building blocks from a real daily-driver setup.

Quickstart • What's Included • Configs • Skills • Templates • Contributing

---

Quickstart

Pick what you need. Most people start with templates or templates + task runner.

# Clone
git clone https://github.com/unisone/openclaw-config.git
cd openclaw-config

# Copy templates into your workspace
rsync -av templates/ ~/.openclaw/workspace/

# (Optional) Copy the task runner
rsync -av scripts/taskrunner/ ~/.openclaw/workspace/scripts/taskrunner/

# (Optional) Copy skill routing overrides
rsync -av skills/ ~/.openclaw/workspace/skills/

Note: Replace ~/.openclaw/workspace/ with your actual workspace path if different.

---

What's Included

Directory	Description
templates/	Workspace bootstrap files — AGENTS.md, SOUL.md, STATE.md, HEARTBEAT.md, IDENTITY.md, USER.md, artifacts/
skills/	Workspace-level SKILL.md overrides that reduce misfires in overlapping domains
config/	Gateway config snippets (JSON5) — compaction, pruning, model fallbacks, memory, search
scripts/memory-engine/	Shell-based memory lifecycle: capture → recall → decay → learn + self-review
scripts/taskrunner/	Python task automation with retry logic, file locking, structured logging
agents/	Multi-agent persona prompts (researcher, architect, security auditor, etc.)
docs/	Operational guides — context overflow, agent memory research, skill routing
content-pipeline/	Optional cron-driven content workflow: draft → approve → post → metrics

---

Workspace Templates

The templates/ directory provides a complete workspace scaffold:

• AGENTS.md — Session boot sequence, memory hierarchy, group chat etiquette, heartbeat system, artifact conventions (~220 lines of battle-tested instructions)
• SOUL.md — Agent personality and philosophy. Opinionated by design — edit to match your style.
• STATE.md — Live working context (max 3KB). Survives compaction and session resets. Priority markers: 🔴 ACTIVE / 🟡 WAITING / ⚪ DONE
• HEARTBEAT.md — Proactive check template: STATE maintenance, cron health, inbox/calendar rotation, self-improvement
• IDENTITY.md — Agent identity (name, emoji, avatar, history)
• USER.md — About the human (preferences, projects, context)
• artifacts/ — Structured output directory: reports/, code/, data/, images/, temp/

---

Skill Routing Overrides

OpenClaw's bundled skills sometimes collide in overlapping domains. These workspace-level overrides add "Don't use for..." negative examples to improve routing accuracy.

Domain	Skills	Routing Logic
Email	gog, himalaya	gog → Gmail/Google Workspace · himalaya → IMAP/SMTP
Notes	apple-notes, bear-notes, obsidian, notion	Each gets explicit boundaries against the others
Transcription	openai-whisper, openai-whisper-api	Local CLI vs cloud API
Messaging	imsg	Fallback iMessage CLI; prefer bluebubbles if available
Code	coding-agent	Multi-step coding sessions; not for simple shell commands
Summarize	summarize	URLs/podcasts/files; not for simple HTML extraction

Based on OpenAI's Shell + Skills + Compaction guidance — Glean reported 20% misfire reduction after adding negative examples.

Overrides mask bundled updates. When OpenClaw ships improved descriptions upstream, delete the override to fall back. Tracked in openclaw/openclaw#14748.

---

Gateway Config Snippets

Production-tested config examples in config/*.json5. Merge into your config — don't replace.

Essential

Config	What It Does
compaction-and-pruning.json5	Highest-impact stability settings (reserve tokens, soft thresholds)
context-overflow-prevention.json5	Defensive settings for heavy tool use
model-aliases-and-fallbacks.json5	Provider fallbacks (Opus → Sonnet → GPT)
memory-search.json5	Semantic search over workspace memory files

Power User

Config	What It Does
pro-optimization.json5	Maximum capability without sacrificing stability — longer sessions, faster recovery, aggressive caching
feature-unlocks.json5	Enable production-ready features that ship disabled by default — auto-analysis, link fetch, typing indicator, heartbeats
memory-lancedb-setup.json5	LanceDB + OpenAI embeddings for long-term memory
web-search-perplexity.json5	Perplexity search (no rate limits with paid key)

# Apply configs, then restart
openclaw gateway restart

Tip: Apply pro-optimization first, run for a few days, then add feature-unlocks. Changing everything at once makes debugging harder.

---

Scripts

Memory Engine (scripts/memory-engine/)

Shell-based memory lifecycle: capture.sh → recall.sh → decay.sh → learn.sh

• Nightly learning extracts errors, corrections, and insights from daily logs
• Self-review system (MISS/FIX tags) prevents repeated mistakes
• See scripts/memory-engine/README.md for setup

Task Runner (scripts/taskrunner/)

Python-based task automation — replacement for fragile shell scripts:

• Retry logic with exponential backoff
• File locking (no concurrent runs)
• Structured JSON logging (logs/tasks.jsonl)
• Alert system for heartbeat integration (alerts/pending.json)
• Example tasks: memory_capture, memory_consolidate, system_health

python3 ~/.openclaw/workspace/scripts/taskrunner/runner.py memory_capture

---

Agents

Drop-in persona prompts for sessions_spawn in agents/:

• deep-researcher.md — Multi-source research with citation
• architect.md — System design and technical architecture
• security-auditor.md — Security review and hardening
• content-writer.md — Content drafting with voice/tone consistency
• ux-designer.md — UX review and design recommendations
• market-researcher.md — Market analysis and competitive intelligence
• project-planner.md — Project breakdown and timeline estimation
• devils-advocate.md — Structured counterarguments

---

Repo Layout

openclaw-config/
├── agents/               # Multi-agent persona prompts
├── config/               # Gateway config snippets (JSON5)
├── content-pipeline/     # Optional: cron-driven content workflow
├── docs/
│   ├── agent-system-research.md           # Memory systems research (4500+ words)
│   ├── session-context-overflow-fix.md    # Context overflow prevention
│   ├── skill-routing-improvements.md      # Skill routing audit + improvements
│   └── content-strategy/                  # Content framework + specs
├── scripts/
│   ├── memory-engine/    # Shell-based capture/recall/decay/learn
│   └── taskrunner/       # Python task automation
├── skills/               # Workspace-level SKILL.md routing overrides
├── templates/            # Workspace bootstrap files
│   └── artifacts/        # Structured output directory convention
└── workflows/            # Optional Lobster workflows

---

Safety & Privacy

These files are meant to be copied into your workspace. Treat them like production config:

• Never commit secrets — API keys, tokens, cookie databases
• Review before copying — some templates contain placeholder values
• Fork responsibly — no personal data, no logs

See SECURITY.md for reporting vulnerabilities.

---

Contributing

PRs welcome. Requirements:

• Must be tested in a real OpenClaw workspace
• No personal data (keys, usernames, emails, logs)
• Explain why a setting exists, not just what it does

See CONTRIBUTING.md.

---

License

MIT