Back to library

Prompt-as-Spec Builder

If your product uses AI, prompts are product decisions — not implementation details. The wording, persona, constraints, and output format shape user experience as much as any UI decision. This skill writes a versioned, production-quality system prompt and packages it as a handoff artefact for engineering.

---

Context

Most teams treat prompts as engineering's responsibility. The result: the model behaves in ways nobody designed, and when something goes wrong, nobody owns it. This skill transfers prompt ownership to the PM — where it belongs.

A well-written prompt spec includes: persona, task, constraints, output format, edge cases, and version metadata.

---

Step 1 — Gather the feature context

Ask for: what the AI feature does, who the end user is, the AI's role, hard constraints, output format, example of great output, and target model/platform.

Step 2 — Design the prompt architecture

Five layers: Persona, Task Definition, Constraints, Output Format, Edge Case Handling. Work through all five before writing prompt text.

Step 3 — Write the system prompt

Rules: second person, persona in first sentence, explicit section headers, constraints as explicit prohibitions, output format with examples, graceful refusal messages.

Step 4 — Write the prompt spec document

Package as a handoff artefact with: feature context, model target, complete system prompt, design decisions (why this persona, why these constraints, why this output format), test cases (typical, edge, out-of-scope, empty, adversarial), change log, and review sign-offs.

Step 5 — Versioning and ownership rules

Treat prompt changes like code changes. Test before deploy. Log prompt versions in deployment system. Own prompts in the PM ticket. Review prompts quarterly.

Quality check before delivering

All five prompt layers are addressed
Constraints are explicit prohibitions, not vague guidance
Output format includes an example
Fallback messages are written out in full
At least 5 test cases are included
Version and ownership metadata is complete
Change log table is present
Suggested next step: Run the test cases manually in the target model before sharing with engineering. Never hand off a prompt that hasn't been tested against its own spec.