Descriptions:
OpenClaw, the AI agent framework, ships with a set of plain markdown configuration files that collectively define an agent’s behavior — and one file, soul.md, functions as the injected system prompt that shapes every session. In this walkthrough, creator Sharbel A. explains why most OpenClaw users are running their agents at roughly 30% of their potential by either skipping the file or filling it with generic instructions that the model effectively ignores. He draws a hard line between soul.md (personality and hard limits) and agents.md (procedures and workflows), showing why conflating the two causes agents to behave inconsistently.
The video documents the full file ecosystem: soul.md, agents.md, user.md, memory.md, and heartbeat.md. Sharbel shares his own configuration, including a specific rule banning em dashes to reduce robotic-sounding output, and credits the soul.md published by OpenClaw founder Pete as a turning-point upgrade to his setup. He also describes generating an initial draft by feeding the agent two 15-minute voice notes — one about his business, one about himself personally — to produce a starting soul.md and user.md.
Five actionable principles structure the second half: specificity beats length, hard limits function as a non-negotiable agent contract, user.md is underused but high-leverage, asking the agent to critique its own soul.md after a week of use surfaces real gaps, and the file should stay under 2,000 words. For anyone building personal or business AI agents on OpenClaw, this is the closest thing to a canonical configuration guide currently available.
📺 Source: Sharbel A. · Published April 24, 2026
🏷️ Format: Tutorial Demo
