Playbooks

Give the agent exact capability guardrails on day one

In the Lamothermic campaign engineers asked what the foundry could cast, and the agent could only answer precisely after the president wrote down what the company does and does not pour. 26 replied threads asked capability questions. The president added the guardrails while the campaign ran. This playbook says to write them before the first message.

Published by Sellinger AI. Numbers and settings checked on 19 October 2026 against the fact register. Report an error.

replied threads with capability questions about alloys, size or location (Lamothermic, to 9 October 2026): 26
engineers and buyers messaged who replied: 293 of 879
capability rule, added during the campaign: state exactly what you do and do not do: 1

The scenario

Buyers of cast parts ask hard questions fast: which alloys, what size, where. A wrong yes costs more than a no. The agent wrote as the president from his own LinkedIn. See the Lamothermic case study for the whole campaign.

The rule, as configured

State capabilities exactly: steels, stainless and bronze; no aluminum, no vacuum melt alloys.
Say what you do not do as plainly as what you do. An honest disqualification counts as a win.
Never write a capability you cannot deliver, and never guess one the buyer asked about.

A capability question and the agent's answer
Who	Message
Capability question	Alloys, size or location, asked by an engineer in 26 replied threads.
Agent, after the guardrail	We primarily pour steels and some bronze, but not vacuum melt alloys like those often used for turbine blades.

Quote from the case study. The prospect's name and company are removed.

Where to write the guardrails

Capabilities, limits and policies belong in the knowledge base, so one edit fixes every answer.
The persona says who the agent is and how it behaves with a capability question: answer exactly, never stretch.
A needs attention rule sends the cases you cannot answer by text, such as a drawing or an NDA, to a person. At Lamothermic four threads asked for an NDA first and all four became hand raises.
Run in approval mode for the first ten messages and read the capability answers.

What the study does not show

The study does not compare answers before and after the guardrail, so it does not measure how many wrong answers the rule prevented. It shows the answers the agent gave after the president set the rules, and the president's own lesson: once he listed what the foundry does and does not pour, the agent answered capability questions precisely.

Common questions

What if the buyer asks for something I do not do?

Say so. At Lamothermic a plain no was read as honest and counted as a win.

Will the agent ever invent a capability?

Agents can make mistakes. The guardrails, the knowledge base and approval mode reduce the risk. They do not remove it, so read the threads.

Sources and method

A playbook here is a configuration that ran, with the outcome it produced and the base of every number. It is not a template promise: your audience and offer decide your own rates. Source: Lamothermic case study, Sellinger production data, read only, snapshot 9 October 2026. All 293 replied threads were read by hand.