OpenClaw Agent Templates and Starter Kit: A Practical Walkthrough

Every OpenClaw agent reads a set of markdown files on startup that control its behavior. These files act as system-level prompts: SOUL.md defines personality, AGENTS.md sets operational rules, TOOLS.md maps available integrations, and MEMORY.md provides persistent context across sessions. Writing these from scratch takes time, especially when you need to get the tone, boundaries, and routing logic right for a specific domain.

Templates solve this by providing tested starting points. The awesome-openclaw-agents repository, for example, contains over 200 production-ready SOUL.md configurations across 24 categories. Each template is copy-paste ready, with domain-specific personality, communication rules, and boundary definitions already written.

The openclaw-starter-kit goes further. Rather than just a SOUL.md file, it bundles the full workspace: IDENTITY.md, SOUL.md, USER.md, AGENTS.md, MEMORY.md, TOOLS.md, HEARTBEAT.md, a memory directory structure, and a learnings folder for documented mistakes. This gives you a complete agent scaffold, not just a personality layer.

The distinction matters. A SOUL.md template gives you a personality. A starter kit gives you an operating system for the agent, with session management, memory hygiene, and proactive task scheduling built in.

OpenClaw's template system splits agent configuration across several files, each with a single responsibility. Understanding what belongs where prevents the "ball of mud" anti-pattern where one file tries to do everything.

SOUL.md is the personality layer. It defines who the agent is: communication style, values, tone, and behavioral boundaries. The official SOUL.md reference recommends five sections: Core Truths (foundational principles), Boundaries (what the agent refuses to do), Vibe (personality and tone), Continuity (how the agent maintains context), and Example Responses (concrete tone samples). A well-written SOUL.md runs 30 to 80 lines. If it grows past 100, you're probably mixing in operational rules that belong in AGENTS.md.

AGENTS.md is the operational handbook. While SOUL.md is who the agent is, AGENTS.md is how it operates. This file covers message routing, security policies, scope boundaries, session management, and file organization conventions. The official AGENTS.md reference describes it as a Standard Operating Procedure: what rules to follow, how to handle different session types, and when to escalate.

IDENTITY.md holds the agent's name, emoji identifier, and one-line specialty description. It's the business card.

USER.md stores information about the human operator: name, preferences, active projects, communication style, and timezone. The agent reads this to personalize interactions. In shared sessions, this file gets skipped to protect privacy.

MEMORY.md is the agent's curated long-term context. Session logs capture individual conversations, but MEMORY.md holds only the distilled insights that matter across sessions: learned preferences, recurring patterns, and important decisions.

TOOLS.md contains environment-specific configuration: API endpoints, service URLs, and integration details. This file separates infrastructure from behavior. Operational rules go in AGENTS.md; connection details go in TOOLS.md.

HEARTBEAT.md defines periodic proactive tasks like checking for updates, running maintenance, or sending reminders. It turns a reactive agent into one that takes initiative on a schedule.

The awesome-openclaw-agents repository organizes its templates into categories that span most common use cases. The broadest categories cover development, marketing, and creative work. Newer categories for voice, customer success, and supply chain are smaller but growing.

Here's how the categories map to common use cases:

For developer tooling: The Development and DevOps categories include coding assistants, code reviewers, documentation writers, infrastructure agents, and incident responders. If you're building an agent that interacts with repositories or deployment pipelines, start here.

For client-facing agents: Customer Success and E-Commerce templates provide agents tuned for support conversations, order tracking, and onboarding flows. Legal templates include intake bots and contract reviewers with conservative boundary definitions.

For content teams: Marketing and Content templates cover social media managers, blog writers, SEO assistants, and newsletter curators. Creative templates add design feedback agents and brainstorming partners.

For specialized industries: Healthcare, Finance, Real Estate, and Education categories each provide domain-specific agents with appropriate guardrails. Healthcare templates, for instance, include strict boundaries around medical advice.

For operations: Business, HR, Compliance, and Supply Chain categories cover internal workflow agents for process automation and policy enforcement.

Each template in the repository includes the agent name, a specialty description, and the SOUL.md file. Browse the repository's README for the full category index and deployment options.

Give your OpenClaw agents persistent, searchable storage

Fastio's Business Trial includes 50 GB, included credits, MCP server access, and semantic search. No credit card, no trial expiration.

Start 14-Day Trial

Copying a template verbatim works for experimentation, but production agents need customization. The Stack Junkie guide on OpenClaw system prompt design outlines five principles that hold up well in practice.

Start with the enforceability test. Bootstrap files like SOUL.md and AGENTS.md are advisory. The agent treats them as strong suggestions, not hard constraints. If you need an absolute restriction (never access external APIs, never share user data), enforce it at the configuration level or through tool permissions, not just through markdown instructions. Write "Confirm with the user before any external API call" in AGENTS.md, but also restrict the agent's tool access to match.

Keep files concise. Every line in a bootstrap file consumes tokens on every interaction. A 200-line AGENTS.md means 200 lines of context loaded into every single message. Monitor file sizes to catch silent growth. If a file gets too long, it may be truncated, and rules at the bottom quietly disappear.

Customize SOUL.md by modifying, not appending. When adapting a customer-support template for your brand, replace the example responses with your own voice samples. Don't add a "Brand Override" section at the bottom. Contradicting instructions earlier in the file creates unpredictable behavior. The agent may follow either version depending on context.

Separate what changes together. If your team rotates tool configurations weekly but only updates personality quarterly, those belong in different files. TOOLS.md for the endpoints, SOUL.md for the voice. This separation also makes it practical to version-control templates. You can review a TOOLS.md diff without wading through personality changes.

Set review cadences. The recommended schedule: MEMORY.md weekly (prune stale context), AGENTS.md monthly (update operational rules), TOOLS.md when infrastructure changes, SOUL.md quarterly (evolve personality). Active pruning prevents files from accumulating contradictory instructions over time.

Templates define behavior, but agents also need somewhere to store their output. MEMORY.md and the memory directory handle session context, but files the agent creates for end users (reports, datasets, processed documents) need durable, shareable storage.

Local filesystem storage works for single-machine setups but breaks when agents run in containers, restart between sessions, or need to hand off files to humans. Cloud storage services like S3 or Google Drive solve durability but add authentication complexity and don't provide the semantic search that makes agent-generated files findable.

Fastio fills this gap for OpenClaw agents specifically. The Business Trial includes 50 GB storage, included credits, and 5 workspaces with no credit card required. Agents access storage through the MCP server via Streamable HTTP at /mcp, which means any OpenClaw agent can read, write, search, and share files through standard tool calls.

The storage becomes more useful with Intelligence Mode enabled. Files uploaded to a Fastio workspace are automatically indexed for semantic search and AI chat. An agent that generates weekly reports can store them in Fastio, and a human colleague can later ask "What were the Q1 findings on customer churn?" and get an answer with citations pointing to the specific report, without knowing the filename.

For teams running multiple OpenClaw agents, Fastio's file locks prevent conflicts when two agents try to write to the same file. Webhooks notify agents when files change, enabling reactive workflows without polling. And when an agent finishes building a project, ownership transfer lets it hand the entire workspace to a human while retaining admin access.

Other options worth considering: self-hosted MinIO for teams that need on-premises storage, or Dropbox/Google Drive for simpler setups where semantic search isn't needed. The right choice depends on whether your agents need to find files by meaning or just by path.

Here's how the pieces fit for a concrete scenario: building a content review agent that checks blog drafts for tone, accuracy, and style guide compliance.

Start with a template from the Marketing and Content category in awesome-openclaw-agents. Pick a content editor or copywriting assistant template as your base SOUL.md. The personality is already tuned for editorial feedback: constructive tone, specific suggestions, and awareness of style conventions.

Next, set up the full workspace using the starter kit structure:

• IDENTITY.md: Name the agent (e.g., "ReviewBot"), assign an emoji, and describe its specialty as "Blog draft reviewer for tone and style guide compliance."
• SOUL.md: Customize the template. Replace the example responses with samples that match your publication's voice. Add your style guide's key rules to the Core Truths section. Define boundaries: "Never rewrite drafts. Only provide feedback and suggestions."
• AGENTS.md: Define the operational workflow. When a new draft appears in the workspace, review it against the style guide. Post feedback as comments. Flag factual claims that need citation. Escalate to a human editor when confidence is low.
• USER.md: Add information about the editorial team lead, their preferences for feedback format (bullet points vs. inline comments), and the publication schedule.
• TOOLS.md: Configure the file storage integration. If using Fastio, add the MCP endpoint so the agent can read drafts from a shared workspace and post reviews back.
• MEMORY.md: Start empty. The agent populates this with learned preferences over time, like "The team prefers shorter paragraphs" or "Avoid suggesting header changes, those are locked by the design team."
• HEARTBEAT.md: Set the agent to check for new drafts every morning and send a summary of pending reviews.

This setup takes about 15 minutes with templates. Without them, writing each file from scratch typically takes an hour or more, mostly spent on getting the SOUL.md tone right and the AGENTS.md routing logic correct.

The memory system then improves the agent over time. Each review session adds to daily logs. Periodically, the agent distills patterns into MEMORY.md: which types of feedback the team accepts, which they override, and which style rules generate the most discussion.