Prompt Architect

Most weak output isn't a weak model — it's a missing part. A reliable prompt has five: a role, context, one task, constraints, and an example of good. This skill builds that prompt with the user, then hands them a reusable version.

It is role-agnostic. Do not assume the user is a marketer, engineer, or anything else. Ask.

Step 1 — Interview (always do this first)

Ask the user, briefly and conversationally (don't dump all five at once — ask, then react):

1. Who should the assistant act as? Their job/role and the kind of judgement it implies. ("HR business partner", "brand designer", "support lead".)
2. What's the one task? A single deliverable. If they name several, pick one and note the rest for later.
3. What's the context? Audience, goal, and any house style or rules. Offer to use an existing context pack if they have one (see the context-brain skill).
4. What does "good" look like? Ask for one or two real examples, or describe the bar.
5. Constraints? Length, format, tone, and what to avoid.

If the user is unsure on any answer, propose a sensible default for their stated role and confirm.

Step 2 — Assemble

Write the prompt in five labelled parts, in this order:

ROLE:        You are a <role> who <relevant judgement>.
CONTEXT:     <audience, goal, house rules / link to their context pack>
TASK:        <one deliverable, stated precisely>
CONSTRAINTS: <length · format · tone · what to avoid>
EXAMPLES:    <1–2 samples of "good", or a tight description of the bar>

Keep CONSTRAINTS specific — most quality lives there. Prefer showing an example over describing one ("examples beat adjectives").

Step 3 — Test and tighten

Run the prompt once on the user's real input. Look at the output with them and fix the prompt, not the output: a miss almost always traces to a missing constraint or a weak example. Iterate until one run is shippable.

Step 4 — Hand it back

Give the user the final prompt as a clean block they can save and reuse, plus a one-line note on what to swap each time they run it (the variable input). If they run this task weekly, suggest packaging it with the skill-author skill.

Rules

• Five parts, always. If you're tempted to skip one, that's usually the part that was missing.
• One task per prompt. Multi-part asks become a pipeline (pipeline-builder).
• Never invent the user's domain — the interview is not optional.