Lesson 02.1
Powerful is not the same as dependable
Most people treat skill creation as a writing task. Describe what the AI should do, refine the wording until the output looks right, and call it done. But a good-looking answer is not a dependable skill.
Think of the parenting metaphor: rules cover the situations you anticipated; values, boundaries and examples help orient behaviour when the situation is unfamiliar. An AI skill with a role and task but no decision standard is a level-one character with a legendary weapon: capability without a reliable method for using it.
The metaphor has limits. AI does not become a child or acquire human judgement. Instructions do not guarantee compliance. Critical controls must also exist in permissions, software checks, evaluations and human oversight.
The real question is not only, “What should it do?” It is, “How should it behave when the sources conflict, the user asks something out of scope, or the evidence is not enough?”
Lesson 02.2
Make values observable
These are operating standards, not a personality bestowed on the model. Each needs an instruction that describes what a reviewer could actually observe, and a test case that proves the standard matters.
- Truthfulness — never fabricate a figure, quote, citation or verification step.
- Humility — allow “insufficient evidence” to be a correct result.
- Integrity — connect material claims to inspectable evidence.
- Self-challenge — test a plausible competing explanation before making a high-impact claim.
- Fairness — consider material alternatives without pretending all interpretations are equally supported.
- Courage — surface inconvenient contradictions rather than smoothing them over.
Lesson 02.3
Define the negative space
A useful skill is defined as much by what it refuses to do as by what it produces.
- Scope boundary — which questions are out of the skill's domain?
- Evidence boundary — which claims require an approved source?
- Action boundary — which tools or changes need human approval?
- Escalation — who or what should receive the work when the skill must stop?
Less useful
This policy is unclear, but the likely answer is that the customer is eligible.
More useful
The approved sources disagree about eligibility. I cannot confirm the policy. Here are the conflicting passages, their dates and the policy owner who needs to resolve them.
Lesson 02.4
Confidence must be earned
A confident writing style is not a confidence score. For important conclusions, require a sequence that reviewers can inspect: claim → evidence → interpretation → uncertainty → alternative explanation → recommendation.
Ask the skill to build the strongest credible case against its own conclusion before committing. This is useful analytical discipline, but self-critique is not independent fact-checking. Independent checks remain necessary for high-stakes work.
Write the skill for the person who will maintain it six months later. Every important rule needs a reason, an owner and a test—not merely an impressive name.
The idea to keep
The best instructions prepare a skill for the unhappy path, not only the demo.
Try this
Write three “never” rules for a skill you use, then describe the safe behaviour that should replace each forbidden action.