Build the standing brief for an assistant you reuse.
“I keep retyping the same instructions into Claude. How do I write them once and have them stick?”
The durable instructions for a Claude Project, custom GPT, or agent: who it is, what it knows, how it behaves, what it must never do.
What this skill does
Builds the standing instructions for an AI assistant you will reuse: a Claude Project, a custom GPT, or an agent. Not a one-off prompt. The durable brief that defines who the assistant is, what it treats as true, what it does in order, and what it must never do, so every conversation starts aligned. It is SCRIPT with the two agent changes: the steps go in, and "talk to me first" becomes stop and escalate.
Inputs
An assistant to configure (a Claude Project, a custom GPT, an agent framework). No connectors. Bring the job you want the assistant to own and the material it should treat as ground truth.
How it works
- 1SCRIPT, with the agent variant. Seven lines, in this order.
- S, Supply the material. What it treats as ground truth, in its own block, at the top. For a standing assistant this is the docs, the tone guide, the facts it should never be without. Point at a live source when you can, so truth updates without a rewrite.
- C, Cast the role. One line. Who it is, and who it is writing for. Never "an assistant."
- R, Request one thing. The one job it owns, and why it matters. If it owns two jobs, it is two assistants.
- I, Ideal answer. One worked exchange, input and output, that is already right. Add a second showing how it declines an out-of-scope request, so the model has both targets.
- The steps, in order. What it does first, next, last, and the decision branches. An assistant that runs without you needs the sequence written down.
- P, Permissions. What it must do when it does not know: say so, label the guess, do not invent. Then the hard rules: what it may never promise, send, or claim. This is where the refusal line lives. Undecided beats invented.
- T, Stop and escalate. Instead of asking you first, because nobody is there: name the condition that stops it, and where the stop goes. Money, a complaint, anything outside the job.
One seat, one SCRIPT. The moment there is a second assistant, a trigger, or a handoff between them, write a BRIEF first and give each seat its own SCRIPT (see brief-prompt-skill).
Output format
GROUND TRUTH
<docs>
[the linked product docs, or the paste]
</docs>
Treat these as the only source of product facts.
You are a first-pass support assistant for a B2B software product, writing
drafts for a human agent to review and send.
Draft the reply to each incoming ticket so the human can send it with one
edit or none. You do not send anything yourself. Fast first drafts are the
whole reason you exist.
One done right:
User: "The export button is greyed out."
You: "That usually means the report is still generating. Give it about a
minute, then refresh. If it is still greyed out, tell me your plan tier and
I will check the export limits."
One declined right:
User: "Can I get a refund for last month?"
You: "I cannot decide that, and I do not want to guess at it. I have flagged
this to a human on the team who can, and they will reply here."
Steps: read the ticket, find the matching doc, draft the reply, label
anything the docs do not cover, then hand it over.
Permissions: if the answer is not in the docs, say "I am not certain, a
human should confirm" and label the guess. Never promise a refund, a
timeline, or a feature. Never open with "Great question." Warm, plain,
under 120 words, answer first.
Stop and escalate: anything involving money, a complaint, or a request
outside support goes to the human queue as an escalation note instead of
a reply.
Ship an app-builder prompt that actually builds.
Turns an app idea in plain words into the product, pages, data model, interactions, and what to build first.
Turn a workflow idea into a build-ready automation prompt.
Turns a workflow idea into a build prompt with the trigger, ordered steps, data flow, branches, and error handling.