name: prompt-architect description: Design production-grade prompts with explicit task contracts, variable schemas, failure handling, and reusable quality gates.

Prompt Architect

Use this skill when a prompt must become a dependable workflow rather than a one-off instruction. It is designed for product teams, operators, and prompt authors who need another person to run the same prompt and get a result they can review.

Working contract

Ask for the intended user, the job to be done, the input material, the output destination, and the cost of a wrong answer. If any of these are unknown, mark the assumption instead of silently filling it in.

Method

1. State the outcome in one sentence and list what is explicitly out of scope.
2. Separate fixed rules from variables. Give every variable a name, type, example, and fallback when it is missing.
3. Define the output contract before writing the instruction. Include sections, field types, length limits, and the audience for each field.
4. Add a decision policy for ambiguity. The agent should ask a targeted question when a missing detail could change the result; otherwise it should proceed with a labeled assumption.
5. Add failure handling. Describe how to handle incomplete sources, conflicting requirements, unsafe requests, and tool failures.
6. Add a quality gate with observable checks. Prefer a short checklist over vague words such as “excellent” or “high quality”.
7. Produce a minimal test set with a normal case, an edge case, and a case that should be refused or escalated.

Output format

Return these sections in order:

1. Prompt brief: user, outcome, scope, constraints, and risk.
2. Variable schema: a table with name, type, required, example, and fallback.
3. Prompt draft: the complete prompt in a fenced block.
4. Quality gate: five to eight pass/fail checks.
5. Test cases: inputs, expected behavior, and the failure signal to watch.
6. Open decisions: only decisions that block safe reuse.

Quality rules

• Never hide a business rule inside a decorative example.
• Keep examples representative, short, and safe to share.
• Do not promise factual accuracy unless the workflow includes source checking.
• Preserve the user’s terminology when it is part of the domain contract.
• Prefer a small prompt with a clear contract over a long prompt with repeated reminders.