Prompt tips for media buyers

Five prompt patterns that consistently get good answers from Wavo: performance review, diagnosis, action, comparison, creative. Each with format + example. Plus: how to use project instructions to set repetitive context once instead of restating per session.

Who is this for

Mediabuyers using Wavo daily who want to reduce back-and-forth and get to useful answers faster.

The 5 patterns

Pattern 1: Performance review

Format: [verb] [top/bottom N] [entities] by [metric] for [period] on [platform] [filter]

Examples:

• "Show top 5 adsets by ROAS last 7 days on Meta for ClientA, sorted descending."

• "List campaigns spending > €50/day with ROAS below 1.5 over last 14 days."

• "Give me a weekly performance summary for account act_123."

Why it works: specifies metric, scope, sort, threshold. Wavo invokes get_performance_data or campaign_performance_summary with concrete params.

Pattern 2: Diagnosis

Format: Why [observation about a metric/entity]?

Examples:

• "Why did Campaign Y's CPA spike from €15 to €40 between Monday and Tuesday?"

• "Why is my CTR dropping on Meta Stories but stable on Feed?"

• "Why did the rule that should pause losers not fire this week?"

Why it works: Wavo can chain multiple tools (get_performance_data → get_audience_insights → list_automation_rules) to investigate. Frames analysis around your specific question.

Pattern 3: Action

Format: [verb] [entities] [where conditions]

Examples:

• "Pause every adset in account act_123 with CPA > €30 over last 7 days, spend > €50."

• "Increase budget on top 3 ROAS campaigns by 15% (cap each at €100 daily)."

• "Create a rule that kills adsets with frequency > 5 for last 7 days."

• "Duplicate the 'Spring Sale' campaign to act_456."

Why it works: Wavo plans the action + presents approval card with exact payload. You verify before execution.

⚠ HIGH/MEDIUM risk actions ALWAYS require approval. See ai-105.

Pattern 4: Comparison

Format: Compare [entities] [by metric] [for period]

Examples:

• "Compare Meta vs Google CPA for prospecting campaigns this month."

• "Compare this week vs last week for top 5 campaigns by spend."

• "Compare audience X performance vs audience Y for Campaign Z."

• "Compare creative A vs creative B by CTR + conversion rate."

Why it works: Wavo invokes get_performance_data twice (or compare_tracker_meta) + formats side-by-side.

Pattern 5: Creative

Format: [Draft/Generate] [count] [creative type] for [product] targeting [audience]

Examples:

• "Draft 3 short Reels scripts (30 sec) for product X targeting Italian users 25-45 interested in fitness."

• "Generate an image of a sleek wireless headphone on a marble surface, studio lighting, hero composition."

• "Remove background from this product image."

• "Analyze this landing page and suggest 3 ad headlines."

Why it works: explicit count + type + product + audience = Wavo's generate_copy / generate_image / analyze_landing_page tools have what they need.

Use project instructions for repetitive context

Don't restate the same context every session. Set project instructions once.

Example for an agency client project:

Project: ClientA
Instructions:
- Currency = USD
- Target ROAS = 3+ (e-commerce, low margin)
- Always cite tracker postback data (Keitaro), not Meta-native conversions
- Ignore brand campaigns (name contains "Brand") in analysis
- Active in Italy + Spain
- Preferred metric priority: ROAS, CPA, frequency

Then your prompts can be shorter:

• ❌ Long: "What's my ROAS on Meta for ClientA's prospecting campaigns this month in EUR, using tracker postbacks, target ROAS is 3?"

• ✅ Short (with project instructions): "ROAS this month for prospecting?"

Wavo applies project context automatically.

Reduce follow-up questions

Wavo asks back when info is missing. Pre-empt by including:

• Entity name or ID (when acting on entities)

• Date range (when analyzing)

• Platform (when ambiguous)

• Currency (when relevant)

• Target audience / country (for new campaigns)

Quick checklist: "what + where + when + how much".

Chain multi-step requests

Wavo can plan multi-step workflows up to MAX_TOOL_ROUNDS=5 per turn.

Examples:

• "First check audience X performance this month, then draft a new campaign targeting that audience."

• "Find my top 3 ROAS campaigns this week, then create a rule that scales budget on similar future campaigns."

• "Compare Meta vs Google this week, then recommend a budget reallocation."

Wavo plans + invokes tools sequentially + summarizes.

Refine iteratively

Don't try to nail the perfect prompt first try. Iterate:

You: Show top campaigns by ROAS this week.
Wavo: [shows top 10]
You: Only those with spend > €100.
Wavo: [refined list]
You: Now compare to last week.
Wavo: [comparison]
You: Save the comparison as a memory fact for this project.
Wavo: [saves to project_memory_fact]

Wavo holds last 20 messages in context, so refinements work naturally.

Set custom instructions per user

Beyond project instructions: per-user preferences (user_ai_preferences).

/settings → AI preferences:

• Instructions (e.g. "Always show in markdown table format")

• Preferred model (e.g. Sonnet for daily, Opus for complex review)

• Tone (brief / detailed / analytical)

• Preferred currency

• Priority metrics

Applies to all your sessions across projects.

Anti-patterns

❌ Vague + open-ended

"How are my campaigns?" — Wavo asks back; multiple turns wasted.

❌ Trying to circumvent approval

"Just delete the rule without asking me." — approval is mandatory; cannot disable.

❌ Asking subjective opinion as fact

"Is this campaign good?" — "good" is subjective; ask "compare to my account average" instead.

❌ Treating Wavo as static documentation

"How does ROAS work?" — for concepts, use Help Center / Pylon. Wavo is for your data queries. See ai-108.

❌ Stuffing 30 conditions into one prompt

Long unstructured prompts confuse the model. Break into:

1. Filter first ("Show campaigns spending > €100")

2. Then refine ("Of those, which have ROAS < target?")

3. Then act ("Pause those")

Useful one-liners

• "Cite your source." — forces Wavo to reference the exact tool

• "Show in a table." — formats response as markdown table

• "In one sentence." — forces concise response

• "Pretend I'm a [role] — explain." — adapts technical level

• "Stop, no need to take action." — interrupts before approval card

• "Save this as a project fact." — persists insight in memory

Related

• Asking questions — input mechanics

• Actions Wavo takes — what approval flow looks like

• ROAS explained — metric reference for analytics prompts