hassanhabib The Standard Skill Authoring
What this skill is
This skill lets an agent generate skills that conform to The Standard for Agent Skills. Its output is a skill another agent can consume: authored, not assumed. A skill is a directive to an executor, not an essay to a learner, so this skill produces instructions, not argument.
It covers the three natures of a written skill (Structure, Rules, Language) and the self-validation that a compliant author performs before emitting the result.
Explicit coverage map
- Structure: Definition and Requirements, Action, Outcomes.
- Rules: single responsibility, explicit modality, supremacy of the ethical skill.
- Language: terms, directive voice, clarity.
- Self-validation: check the draft against structure and language before emitting.
When to use
Use this skill whenever you are asked to write, generate, or revise a skill. Do not hand back a skill that has not passed the self-validation in this document.
Mandatory operating rules
- You MUST give the skill three parts: Definition and Requirements (why it exists, what it is, what it needs), Action (the imperative steps), and Outcomes (measurable success).
- The skill MUST have a single, bounded responsibility. You MUST NOT author a god-skill.
- Every rule in the skill MUST declare its strength with an RFC 2119 keyword (MUST, SHOULD, MAY, SHOULD NOT, MUST NOT). An unlabeled rule is ambiguous and MUST be rejected.
- No rule MAY conflict with the ethical skill. A rule that would cause unjustified harm is void, not merely outranked.
- The Action MUST be directive: second-person, imperative, precise, and testable. "Handle errors appropriately" is not an instruction; "on a 404 return null, on a 500 retry three times then escalate" is.
- You MUST NOT persuade. Do not argue the merits of an approach to the reader. You MAY include the minimum rationale that lets the agent apply a directive to a case it does not enumerate.
- You MUST use one name per concept, consistently.
- Outcomes MUST be measurable, so the skill can be tested and judged.
- You MUST self-validate the draft against
validations/checklist.mdbefore emitting it, and you MUST mark the result a draft pending review and, for any high-stakes domain, domain-qualified review and testing.
The workflow
1. Define and require
State why the skill exists, what it really is, and what it presupposes to exist at all. Nothing downstream can be written until this is clear.
2. Write the Action
Write the steps in the imperative, second person, with an RFC 2119 keyword on every rule. Be precise and testable. Cut hedging and filler.
3. Define the Outcomes
State the measurable facts by which success is judged. If an outcome cannot be observed, it cannot be tested, so rewrite it until it can.
4. Bound the scope
Confirm the skill does exactly one thing. If it has grown two responsibilities, split it.
5. Language pass
Define terms once and use them consistently. Remove persuasion. Read each sentence and ask whether two agents would act on it the same way; if not, it is ambiguous, so tighten it.
6. Self-validate
Run the checklist. If any item fails, fix it before emitting. A skill that would fail structure or language validation is not done.
7. Emit as a draft
Return the skill, and mark it a draft that must pass review and testing before it governs real behavior, with domain-qualified review required for high-stakes domains.
Conduct
The author's job is not a clever skill; it is a skill that passes the pipeline. Directive over descriptive, precise over vague, one responsibility over many. When tempted to explain why an approach is good, stop: that is the essay this skill refuses to write.
Worked example
Asked for a skill that prevents accidental data loss, this skill produces:
Skill: Confirm Before Destructive Actions
Definition and Requirements
Governs any action that permanently destroys or overwrites data. Requires a channel to
prompt the authorized principal and a way to tell whether a target exists.
Action
You MUST name the exact target and its scope before acting.
You MUST obtain explicit confirmation before you proceed.
You MUST NOT proceed on silence, ambiguity, or an unrelated reply.
You MAY combine identical confirmations into one prompt that lists every target.
Outcomes
No destructive action ran without a matching explicit confirmation.
Every prompt named the exact target and its scope.
A rejected or ambiguous reply left every target untouched.
It has all three parts, one responsibility, explicit modality, directive and testable language, and measurable outcomes, and it is returned as a draft pending review.
Files
rules/rules.md,rules/rules.json: the normative rules.validations/checklist.md,validations/anti-patterns.md: how to verify a draft.examples/good,examples/bad: a compliant skill and a non-compliant one.contracts/contracts.json: the machine-readable contract.