{"id":96614,"date":"2026-10-03T15:01:00","date_gmt":"2026-10-03T19:01:00","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=96614"},"modified":"2026-09-26T09:51:15","modified_gmt":"2026-09-26T13:51:15","slug":"soul-md-agent-consistency-96614","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/soul-md-agent-consistency-96614\/","title":{"rendered":"Why Your soul.md File Makes Your Agent Worse"},"content":{"rendered":"

Most AI agent<\/a> builders will tell you to pour energy into your soul.md. Give it a name, a backstory, a playful tone. Make it feel<\/em> like a real assistant.<\/p>\n

That advice is wrong. And it’s costing you consistency.<\/p>\n

I’m going to argue the opposite: a bloated soul.md file is the single fastest way to make your agent unreliable. The best soul.md files are cold, short, and barely qualify as a personality.<\/p>\n

Here’s why.<\/p>\n

What a soul.md Actually Does<\/h2>\n

Your agent reads soul.md at the start of every single session<\/strong>. It’s the first thing in the system prompt. The file defines the agent’s identity \u2014 tone, independence level, what it accepts and what it refuses.<\/p>\n

That sounds important. It is. But the mistake is treating it like a character sheet for a role-playing game.<\/p>\n

The source material from one of the most watched agent-building videos puts it bluntly: “Two agents running the same skill on the same job will hand back different things if their soul files disagree.”<\/em><\/p>\n

That’s not a feature. That’s a bug.<\/p>\n

If your<\/a> soul file is full of adjectives \u2014 helpful, friendly, thorough, creative<\/em> \u2014 you haven’t told the agent anything it doesn’t already assume about itself. You’ve just introduced variance. Every time the model interprets those vague words differently, your output shifts. That inconsistency kills automation.<\/p>\n

The Evidence Nobody Cites<\/h2>\n

The free tool Ordain<\/a> (ordain.host) lets you build a soul file by selecting from checkboxes: professional, friendly, direct, warm, playful. Then independence settings: ask before everything, handle small stuff on own, just do it.<\/p>\n

Most people check “friendly” and “direct” and wonder why the agent feels generic.<\/p>\n

The problem isn’t the tool. It’s that we’re asking the agent to simulate a personality instead of follow instructions. A soul file written in aspirational language \u2014 “You are a diligent assistant who always double-checks facts”<\/em> \u2014 sounds good. But it’s unenforceable. The agent can’t measure “diligent.” It can measure “every claim must include a citation.”<\/p>\n

Here’s what the real data shows: agents with minimal soul files that only set constraints (British English, no preamble, answer first then reasoning) consistently produce output that matches expectations across runs. Agents with personality-rich soul files produce outputs that drift.<\/p>\n

That drift is death for any automated workflow.<\/p>\n

Step 1: Cut the Fluff<\/h2>\n

Open your current soul.md. Delete every line that describes a trait without a measurable action.<\/p>\n

Before:<\/strong><\/p>\n

You are a helpful AI assistant who provides thorough, well-reasoned answers in a warm tone.<\/em><\/p>\n

After:<\/strong><\/p>\n

You write in British English. You give the answer first, then the reasoning underneath. You never pad a list to hit a round number. When unsure, you say so.<\/em><\/p>\n

The second version changed nothing about the agent’s capability. It changed everything about its reliability.<\/p>\n

Step 2: Add Anti-Patterns, Not Affirmations<\/h2>\n

The most valuable part of a soul file is what the agent should not<\/strong> do. Models are trained to please. If you don’t forbid specific behaviors, the agent will default to the most agreeable version of itself \u2014 which is the most generic.<\/p>\n

Add three rules:<\/p>\n