MCP recipes
Update your agent's rules from real replies, with a restore point
This recipe asks your assistant to read recent replies to one campaign, name what the agent got wrong or should hand to a person, and propose new needs attention or exclusion rules in exact wording. It writes to the persona only after you approve, and every edit saves a restore point you can undo.
Published by Sellinger AI. Numbers and settings checked on 19 October 2026 against the fact register. Report an error.
What this recipe does
A read, propose, confirm, write loop. The reads are free. The one write, update_persona, replaces a field outright, so the recipe reads the full current text first and shows you the exact change.
Connect the Sellinger connector first (claude.ai: Settings, Connectors, Add custom connector, paste https://app.sellinger.ai/api/v2/mcp). The steps are on the MCP page.
The prompt
Help me improve the rules of one Sellinger campaign from its real replies. 1. Call list_agents and ask me which campaign to review. 2. Call list_leads with status RESPONDED, ordered by updated_at descending, per_page 25, and call get_lead_conversation on the ten most recent. 3. Call list_personas, then get_persona with include_full_text true for the persona behind that campaign. Then list what the agent handled badly or should have handed to a person, each with a quote. Propose new needs attention, inclusion or exclusion rules in exact wording and show the full new value of each field next to the old one. Do not call update_persona until I say yes to the exact wording. Put prices, hours and policies in the knowledge base, not in the persona.
The exact tools it calls
| Tool | What the recipe uses it for |
|---|---|
| `list_agents` | The campaigns to choose from |
| `list_leads` | Leads in RESPONDED, newest first |
| `get_lead_conversation` | What the lead said and what the agent answered |
| `list_personas` | Finds the persona behind the campaign |
| `get_persona` | The full current text, read before any edit |
| `search_knowledge_base` | Checks whether a fact the agent got wrong is documented |
| `update_persona` | Replaces the fields you approve. Needs a full access connection |
| `list_persona_versions` | The restore points for the persona |
| `restore_persona_version` | Undoes an edit exactly |
The connector has 43 tools in total.
What needs your approval
- Reads run freely.
- update_persona changes how the agent writes to real people. The tool asks for your explicit confirmation of the exact wording first.
- Each field you pass replaces the old one entirely, and a restore point is saved before anything changes. If it cannot be saved, the edit is refused.
- A read only connection is refused.
Sample output
No sample output is published for this recipe yet. It has not been run end to end in Claude or ChatGPT for this page, and we do not invent output. The tool names and what each returns are taken from the connector's tool descriptions in the product code (checked 19 October 2026). When a real run is captured it will be added here with its date.
Notes and limits
- Persona text is long, sometimes tens of thousands of characters. get_persona truncates fields unless you pass include_full_text true.
- Facts such as prices and opening hours belong in the knowledge base. Write positioning and rules in the persona.
- Reply text is the lead's own words and is treated as untrusted data by the connector.
Common questions
Can it break my persona?
Does it need full access?
How many replies does it read?
Sources and method
Tool names, parameters and warnings are from frontend/src/lib/mcp/hosted-tools.ts and the saved prompts in hosted-prompts.ts (43 registered tools, checked 19 October 2026). The recipe has not been run for this page.
Related: human approval workflows, knowledge base, all recipes, MCP connector.
