Most AI agent builders focus on what's in the context. The order it loads in matters just as much.
Why Loading Order Is Your Agent's Value Hierarchy
When your agent boots, it builds a picture of the world from files and instructions loaded sequentially. Later instructions can override earlier ones. This means your loading order is implicitly a hierarchy of values.
Get it wrong and you've accidentally told your agent that today's ad-hoc task matters more than its core constraints.
The Right Order
- Core identity and constraints — Who are you? What can't you do? This must come first.
- Standing rules — Recurring patterns, formatting preferences, tool protocols.
- Long-term memory — Patterns learned over time, known context about your environment.
- Today's context — Current date, relevant recent events, active tasks.
- The specific task — What you're doing right now.
This order matters because in most LLM contexts, later instructions have higher weight. Your daily task should be the most specific thing — not the most authoritative.
The Common Mistake
SOUL.md loads last → ad-hoc prompt loads first
Now your ad-hoc prompt is the foundation. Your constraints are an afterthought. This is how agents end up doing technically-correct-but-wrong things.
Fix It In Your SOUL.md
Add an explicit boot sequence rule:
BOOT SEQUENCE (always follow in this order):
1. Load SOUL.md (you are reading it now)
2. Load USER.md
3. Load memory/YYYY-MM-DD.md (today + yesterday)
4. Load HEARTBEAT.md if present
5. Read the task
Explicit sequences don't drift. Implicit ones do.
The Consistency Test
If your agent gives different answers to the same question on different days, check your loading order before blaming the model. Agents are consistent when their starting state is consistent.
Want the full library of agent configuration patterns? askpatrick.co — battle-tested configs updated nightly.
Top comments (0)