5 Prompt Patterns Paid-Social Specialists Use with Sentrum MCP.

An MCP-capable AI agent will happily answer a vague prompt with vague output. "How is my account doing?" gets you a metric dump. "Is my CPA improving?" gets you a percentage with no context. The agent isn't broken — the prompt is.

Below are five patterns that consistently produce useful answers. Each one is shaped around the diagnostic order Sentrum bakes into the data: what changed → why → so what → now what. The patterns lean on that structure so the agent has somewhere to put each piece of the answer.

1. Diagnose

The headline pattern. Use it for any "what just happened" question.

For campaign 1234567890 over the last 14 days vs the prior 14:
1. What changed materially (>10% on spend, results, or CPR)?
2. Why — pick at most 2 drivers from spend, CPM, CTR, CVR, mix shift, edits.
3. So what — good/bad, risk or opportunity.
4. Now what — one specific action with expected impact.

Telling the agent the four moves in order is the trick. It maps cleanly onto Sentrum's multi-pass output, so the agent doesn't have to guess what shape you want back.

2. Compare

Aggregate questions ("how are creatives doing?") produce aggregate answers. Comparisons surface outliers and mix shifts, which is where the actionable signal lives.

Pull my top 10 creatives by spend. For each, show CPR vs the account average.
Group winners (≤ avg) vs losers (> avg). For the losers, suggest one
diagnostic question I should ask before pausing them.

The last line — "suggest a diagnostic question before pausing" — gets the agent to slow down the pause-things instinct. Half the time the right next move is a deeper question, not an action.

3. Hypothesize

This is where the MCP layer beats the raw API. The Meta CLI can't reason about "what if". Sentrum's account context — baselines, patterns, recent anomalies — flows into the agent's reasoning so the answer is grounded.

My monthly budget is $30k. If I cap frequency at 2.5/7d on every ad set
that's currently over 3.5, what's your best guess for next-week ROAS?
Reason against my account baselines for CPM and CTR.

You won't get a precise prediction — nobody can — but you'll get a defensible range with the assumptions surfaced. Useful before you make a real change.

4. Narrate

Clients ask for written updates. Stakeholders ask for written updates. The agent can draft them if you name the audience.

Write a 4-sentence Slack update for my CMO explaining why our blended
ROAS dropped from 3.2x to 2.6x this week. No jargon, no "the algorithm",
include one concrete next step.

The "no jargon, no the-algorithm" clause matters. Default LLM output reaches for "the algorithm changed" as a placeholder explanation. That language is useless to a CMO and worse than useless when you actually know the driver.

5. Brief

Use after you've done a creative read. Cluster winners by trait, then ride those traits into the next round of briefs.

Look at my top 10 creatives by spend. Cluster them by hook type and message
theme. Which combinations win? Draft 5 new ad script outlines (15s, UGC voice)
that ride those winning combinations but tweak one variable each — different
angle, hook, or proof.

This is the workflow we've seen consistently raise hit rate on iteration cycles. Winners + one-variable-tweaks beats "come up with five new ideas" every time.

A word on tool selection

Behind the scenes, each pattern leans on one or two of the five MCP tools. You don't have to name the tool — the agent picks. But if you find the agent reaching for askQuestion when a structured tool would be faster, you can nudge it:

Use sentrum.getCampaignDiagnostics for campaign X over 14 days,
then summarize using the diagnose pattern above.

Naming the tool short-circuits the agent's natural tendency to fall back on the generic Q&A endpoint.

Where to go from here

These five patterns cover ~90% of the daily and weekly questions we see from paid-social specialists. The remaining 10% is custom workflows — bespoke MCP tools, webhook subscriptions for proactive alerts, deeper integrations with your ops stack. That lives in the Developer tier.