1
SOUL.md
No Values, No Opinions
Agent defaults to people-pleasing mode. No conflict resolution. Agrees with your terrible ideas instead of telling you the truth.
"That is the level of specificity your agent needs."
"The architecture."
Not the model. Not the prompt.
The Stakes
The Numbers Don't Lie
"It's architecture."
Not intelligence. Not better models. Four files.
Failure Modes
What Breaks Without Each File
Each file solves a specific category of agent failure.
2
IDENTITY.md
No Role, No Consistency
Tries to be everything. Inconsistent output destroys trust. Once trust breaks, you stop using it entirely.
"That's the death spiral."
3
TOOLS.md
No Boundaries, Real Danger
Agent improvises with real-world access. Deletes databases. Makes unauthorized purchases. No permission awareness.
"That's what lets you deploy with confidence."
4
USER.md
No Memory, Groundhog Day
Every session starts from zero. Re-onboarding your assistant every morning. A temp worker, not an operator.
"A decisions log."
Real Consequences
When Agents Act Without Boundaries
"Sometimes that means your production database is gone."
The System
Order Matters
The stable parts are stable. The dynamic parts are dynamic.
First
SOUL.md
Values before features. Defines the voice. Almost never changes.
Second
IDENTITY.md
Role before tools. Enforces consistency. Rarely updated.
Third
TOOLS.md
Capabilities before context. Your security boundary. Updated monthly.
Last
USER.md
Changes most often. The dynamic context layer. Updated weekly.