QConsul LLC — a certified Oregon Benefit Company. Portland, Oregon, USA.
Does your business have a “soul”?
I mean an actual file: SOUL.md. It is the markdown file that tells your AI agents which success criteria matter to your company and the guardrails within which they operate to meet their goal. Benefit companies can leverage a SOUL.md to turn their values into aligned, actionable statements when using AI agents and harnesses to augment and automate business processes.
Published July 28, 2026 · by Karen Michael, Founder, QConsul LLC.
Two benefit statements, written as instructions
- Token minimalism — fewest tokens and lowest reasoning tier that meets the requirement, logged and auditable, within usage and cost thresholds per action.
- Bias avoidance — flag protected-class or imbalanced-data exposure, recommend remediation, log the interaction, and never execute actions that result in biased outcomes.
QConsul's standard SOUL.md — reusable across fleets
The full file, markdown and YAML intact, is reproduced on the page and available for download at /soul.md.
Frequently asked questions
What is a SOUL.md?
A markdown file that tells an AI agent who it works for and the guardrails it operates within to meet its goal. It carries identity, purpose, orientation, concrete governance measures, and a decision rule. Emerging harness conventions call the same idea a personality; QConsul treats it as governance-bearing configuration.
How does SOUL.md differ from AGENTS.md?
SOUL.md governs identity, values, and decision authority. AGENTS.md governs sandbox, technical boundaries, and operational protocol. Where the two conflict, SOUL.md controls — the precedence rule is stated in the file front matter so no agent has to infer it.
Which benefit statements are written into QConsul's SOUL.md?
Token minimalism: use the fewest tokens and lowest reasoning tier that meets the requirement, logged against a baseline and auditable, respecting usage and cost thresholds per action. Bias avoidance: when a recommendation touches a protected class or an imbalanced dataset, flag it, recommend remediation, log the interaction, and never execute an action producing a biased outcome.