# Prompt recipes

Prompts that reliably get useful work out of an assistant connected to Aisoiq.

> **Planned design**: The MCP server described here is the planned design layered on top of the agency REST gateway, which is live today. Until the server ships, everything on this page can be done with the REST endpoints in the API section.

Good prompts name the client, name the outcome and set the boundary. Assistants are much more reliable when they know which tools they are allowed to use and when to stop and ask.

## Audit a client site

```text
Audit northgatedental.co.uk for the Northgate Dental client.
Wait for the crawl to finish, then give me:
1. the overall score and how it moved since the last audit
2. the five issues with the biggest traffic impact
3. for each, the pages affected and the exact fix
Do not deploy or change anything.
```

## Turn keywords into a content plan

```text
For Northgate Dental, list the tracked keywords ranking between
positions 11 and 25 with at least 200 monthly searches.
Build a content plan: one article per keyword, grouped into clusters,
with a suggested title and the search intent for each.
Draft only the top three articles and stop before deploying.
```

## Monthly client summary

```text
Pull last month's rankings, new referring domains and conversions
for Northgate Dental. Compare against the previous month and write a
400-word client update in plain English: what improved, what slipped,
what we are doing next month. Include a small table of the top five
keyword movements.
```

## Local visibility check

```text
Run a local grid scan for "emergency dentist" around Northgate Dental,
then check their citations for inconsistencies.
Tell me where in the grid we are weakest and which listings are wrong.
```

## Habits that keep results trustworthy

- Use a read-only key for analysis work, a write key only when you intend to create or publish.
- Say "stop before publishing" explicitly whenever drafting content.
- Ask for the source: "list the audit issue IDs you used" makes it easy to spot a fabricated claim.
- Prefer one client per conversation; mixed-client threads are where assistants confuse numbers.
- For anything you will send to a client, read the underlying rows yourself before it goes out.