How to use the Prompt Library

This page is the starting point for anyone new to the Prompt Library. Once you know where to paste a prompt, how far it will go when you run it, and which ad product it targets, the rest of the categories are much easier to read.

How to run a prompt

Picaro.AI has two usage patterns, and the prompts in this library work as-is in both.

• Picaro.AI dashboard — paste the prompt into the AI chat inside the dashboard and send it. The connection to your ads account is established automatically.
• MCP connection (Claude / ChatGPT, etc.) — paste it straight into an AI agent that has the Picaro MCP plugin connected. Connection setup is required in advance.

Prompts marked with an MCP only badge — in the list or at the top of a page — are only for MCP connections. In the dashboard you can do the same things from the UI.

If you haven’t connected yet, start from Connect ad accounts.

The four interaction patterns

Every prompt states its interaction pattern. How far it proceeds automatically differs by pattern, so check it before you run.

Pattern	What happens	Changes to ad settings
Instant display	Send it and a table or summary comes straight back	None (read-only)
Analyze → Approve → Execute	The AI presents a proposal; nothing is applied until you reply “execute it”	Applied after approval
Interactive creation	The AI confirms parameters in dialogue, shows a preview, then executes	Applied after approval
Interactive logging	The AI collects information through dialogue, then records or updates it	Records only, such as action logs

Nothing is applied until you approve

“Analyze → Approve → Execute” and “Interactive creation” prompts write nothing to Amazon Ads until you approve. Conversely, the moment you reply with your approval, real bids, budgets, and keywords change. Read the proposal before you approve.

Note that caps and adjustment ranges (such as “up to ±20%”) are instructions written in the prompt text — they are not a limit the system enforces. If there is a range you need to stay within, write that condition into the prompt.

Supported ad products and data sources

Prompt Library categories are split by the ad product or data source they target. The terminology differs per target too, so knowing these distinctions makes picking a page much faster.

Target	Full name	Management units	Main categories
SP	Sponsored Products	Campaign / ad group / keyword / product target	Create Sponsored Products (SP) campaigns, Adjust bids, and others
SB	Sponsored Brands	Campaign / ad group	Handled alongside SP in the cross-product categories
SD	Sponsored Display	Campaign / ad group	Handled alongside SP in the cross-product categories
Amazon DSP	Demand-Side Platform	Order / line item	Analyze Amazon DSP delivery performance
AMC	Amazon Marketing Cloud	Customer level / ad-exposure level	Analyze cross-ad customer behavior with AMC

“Order” and “line item” are Amazon DSP terms

SP / SB / SD use “campaign”, “ad group”, “keyword”, “product target”, and “negative target”. Orders and line items appear only on the Amazon DSP pages.

Choosing between AMC and Amazon DSP

• AMC — analyzes the customers who touched your ads across SP / SB / SD / DSP, at the customer level: acquisition, repeat purchase, LTV, NTB, and brand halo.
• Amazon DSP analysis — analyzes delivery performance at the order / line-item level (impressions, CTR, CVR, ROAS, and so on).

“Which ad type is bringing in new customers?” is an AMC question; “which line item’s delivery efficiency dropped?” is a DSP-analysis question.

How to replace placeholders

The {{...}} markers in a prompt body are the spots you replace with your own values before running.

Analyze the SP campaigns from the past {{PERIOD (e.g. last 30 days)}}.

In the example above, you would rewrite {{PERIOD (e.g. last 30 days)}} as something like last 30 days. Each prompt’s “What to enter” or “Placeholders” field says what to fill in, with examples.

Watch out for placeholders you forgot to replace

If you send a prompt with {{...}} still in it, the AI either can’t identify the target and asks you back, or runs against a target you didn’t intend. Check that no {{ remains before you send.

For prompts that take a campaign name or ASIN, asking “list candidates from the real data” and picking from what it shows you is more reliable than typing a value from memory.

What to check before you run

• Whether you’re on the right account — if you work across multiple accounts, confirm this first so you don’t run against the wrong one.
• Whether the data for the period is final — same-day and previous-day data may not be finalized. Looking at a recently completed period is safer.
• Whether the prompt writes anything — check the “Interaction pattern” field.
• Whether KPI targets and labels are registered — doing Set targets for sales, ad spend, and ACoS and Classify ads with labels and naming rules first lets you omit parameters in later prompts and improves judgment accuracy.

How to sanity-check the results

Don’t take the returned tables and proposals at face value — check the following.

• Do the counts add up? — look for a caveat such as “M of N records were covered”. When there is a lot of data, only part of it may be referenced.
• Is the ordering right? — for top and bottom rankings, read the numbers and confirm the order is correct.
• Did you get the metric you asked for? — check that a metric which couldn’t be retrieved wasn’t quietly substituted with a different one. The correct behavior is to state “not retrievable”.
• Where did the numbers come from? — check that estimates weren’t invented to fill gaps in the data.

If something looks off, follow up with “show me the basis for this number” or “tell me how many records were covered and how many were excluded” — the AI will re-check and correct itself.

About the level of automation

Automation goes as far as “the AI proposes, a human approves” — that is the operational boundary. There is no fully automatic mode that keeps changing bids and budgets without approval, and there is no mechanism that automatically rolls changes back. See Automation phases for details.

If something isn’t working, check Troubleshooting; if that doesn’t resolve it, you can reach the dev team directly from Send feedback and support requests.