# Saved Prompt Template

Once you have a weekly task the machine can handle, don't retype the
instructions every time. Save them somewhere permanent — a document you can
paste into a new session every week. Some tools call this a "system prompt" or
"custom instructions." Whatever your tool calls it, it's the same thing:
standing orders the machine follows every time you give it the weekly data.

Your saved prompt should include four parts.

---

## 1. Context paragraph — who you are, what your world looks like

Set this once so every task has context. Example:

> For context going forward: I manage a team of eight in facilities management.
> Our clients are commercial property owners, mostly office buildings and retail
> centers. When I say "tenant," I mean the business renting space, not a
> residential renter. When I say "PM," I mean property manager, not project
> manager. My reports go to the VP of Operations, who cares about cost per
> square foot and response time metrics, not technical details.

## 2. Task instructions — the specific steps for this task

Be specific about where the inputs come from, what you do with them, and where
the output goes. Example:

> Here is this week's data: [paste — sanitized]. Reorganize it into [this
> format], calculate [these metrics], and add commentary on anything that
> changed more than X% from last week.

## 3. Output format — exactly how you want it back

> Give it back as a one-page summary. Lead with the cost impact, not the
> technical detail. Keep each section under 150 words. Use a table for the
> numbers.

## 4. The checks — pulled from your Fire Log

These are the machine's known failure modes in your work. Bake the guardrails
into the prompt so it checks itself before you check it. Example:

> Before producing the final output, list every specific number, date, and
> factual claim in the document. Next to each one, note the source. If you
> cannot identify a source, flag it with [UNVERIFIED] so I can check it before
> sending. Never make up percentages. Always show your math on the totals.

---

The machine flags the claims. You verify the flags — against a source you open
yourself, not one the machine summarizes for you. That's two layers of review
instead of zero. That's governance. The one-hook version.