Skill is all you need.

A personal AI assistant where capabilities live in Markdown files, not code. Inspired by OpenClaw and nanobot, skillbot pushes the "skill-as-code" philosophy to its logical extreme: 815 lines of core TypeScript replaces what took 300,000+ lines in OpenClaw and 8,500+ in nanobot -- with the same features.

Every new capability is a .md file. No code changes, no redeployment, no dependencies.

The insight is simple: LLMs can read instructions. Instead of writing code that calls APIs, parses responses, and handles errors -- write a Markdown file that tells the LLM how to use existing CLI tools. The LLM becomes the integration layer.

This works because:

• Most tools already have CLIs (gh, curl, docker, git, crontab, ...)
• LLMs are excellent at following structured instructions
• Markdown is the most readable format for both humans and LLMs

Create .env:

Start chatting:

That's it. You have an AI assistant with 33 skills, persistent memory, background agents, and auto-compaction.

The entire core is 9 files. Every feature -- from weather queries to market analysis to smart home control -- is a Markdown skill that teaches the LLM what commands to run.

Most agents hardcode memory into the runtime. skillbot uses a dual-layer Markdown memory that the LLM manages itself:

• MEMORY.md -- long-term facts (user preferences, important context)
• memory/YYYY-MM-DD.md -- daily interaction logs
• Search with grep -ri "keyword" MEMORY.md memory/

The memory skill (always: true) is injected into every prompt, so the LLM proactively saves and retrieves context. Before auto-compaction, it flushes important context to daily logs -- no memory is ever lost.

Only always: true skills (memory, context-manager, persona, security) are fully injected into the system prompt. All other skills show only their name and one-line description in a catalog. When the LLM needs a skill, it reads the file:

This keeps the prompt small (~2K tokens for the catalog vs ~30K if all skills were injected), while giving the LLM access to 33 capabilities.

Two Markdown files define who the bot is and who it's talking to:

Both are injected at the top of every system prompt. Change the bot's personality by editing a file.

Instead of building custom tool implementations for each integration, skillbot has exactly two tools:

• bash -- execute any shell command (with safety guards)
• spawn -- run a background subagent for complex tasks

Everything else (curl, gh, docker, crontab, osascript, ...) is accessed through bash, guided by skills. This means skillbot can do anything your shell can do -- which is effectively everything.

Every skill can be tested without real execution:

The mock system pattern-matches bash commands and returns realistic fake outputs. Complex scenarios test multi-turn reasoning, compaction flow, subagent spawning, and on-demand skill loading.

EOF

skillbot proves that with the right abstraction -- Markdown skills + a minimal LLM loop -- you can match the functionality of projects 100-300x larger.

Skill is all you need.