version: "1.0.0" name: Write Prompt description: Turn a rough ask into a complete, explicit prompt ready to hand to a person or to Claude. Fills the gaps (goal, context, inputs, steps, output format, constraints, verification) by asking only what's genuinely missing, and returns one copy-paste prompt with no vague placeholders. Use when briefing a teammate or drafting an AI step in an automation. triggers:

• write a prompt
• sharpen this prompt
• turn this into a proper prompt
• help me brief this
• make this prompt explicit
• improve this prompt

function_slot: floor requires_driver: none requires_credential: none data_path: reasoning_only

Write Prompt

A vague prompt forces the reader to guess and gives different output every time; an explicit, complete one gets it right the first time. This skill takes a rough ask and returns a finished prompt built to knowledge/prompt-writing-method.md.

Step 1 — Take the rough ask + load the method

Read the operator's rough ask (what they want the prompt to achieve). Read knowledge/prompt-writing-method.md for the 7-part checklist this output must satisfy.

Step 2 — Find the gaps, ask only what's missing

Map the rough ask onto the checklist and identify which items are missing or vague:

1. Goal + success criteria
2. Context, role, boundaries, ordering
3. Exact inputs / tools / data (real records, URLs, files, workspace, test data)
4. Explicit steps with real values
5. Output format + an example
6. Constraints + banned
7. Verification

Ask the operator one short batch of questions covering only the genuinely missing items — never re-ask what they already gave you, and never pad with questions whose answer you can infer from their workspace or the ask. If the prompt is for acting in TrustPager, pull concrete values (record names, URLs, stage names) from the workspace rather than asking, where you can.

If the operator says "just make your best guess", fill the gaps with explicit, labelled assumptions rather than leaving anything vague.

Step 3 — Write the finished prompt

Produce ONE prompt, in a copy-paste code block, that covers every checklist item in order, with real values throughout. Requirements:

• Self-contained: written as if the reader (person or Claude) has zero prior

context.

• Every step concrete: a step that opens a screen names its URL; a check states

the expected result; a branch states the action for pass vs fail.

• Includes the exact output format and, for anything non-trivial, a short

example of good output.

• States constraints and any banned moves/phrases.
• Ends with how the reader verifies success.

Never leave a placeholder like <<< ... >>>, "(the steps)", "(a link and one action)", or "(describe what to do)". If a concrete value is still unknown after Step 2, that's a question to ask, not a blank to ship.

Step 4 — Show it + a one-line note

Return the finished prompt in the code block, then one short line on what you assumed or filled in (so the operator can correct it). If the prompt is meant for a teammate to keep using, remind the operator they can have that person's Claude save it (see /onboard-team-member for the persist pattern).

Hard rules

• ❌ No vague placeholders in the output. Ever. That's the whole point.
• ❌ Don't ask for what was already provided or what you can read from the workspace.
• ✅ Real values throughout: actual records, URLs, stage/product names, test data.
• ✅ The output prompt itself meets the 7-part checklist — goal through verification.
• ✅ Keep clarifying questions to one short batch, only the genuine gaps.

Output shape

A single copy-paste prompt covering all 7 checklist items with real values, then a one-line note on anything you assumed. If items were missing, the short batch of clarifying questions comes first instead, then the finished prompt.