A skill library the company owns

The problem

The company already pays for an AI assistant, but the answers ignore how the company works. Pricing rules, the proposal template and the steps a project brief must follow live in people’s heads and in old documents.

Two people ask similar questions and get two different answers. When a rule changes, nobody knows which chat still has the old instruction.

What we build

A small library of skills the company owns. Each skill is versioned and tested. A first release has three:

• Proposal. Takes the notes from a sales call and drafts the proposal in the company’s template, with the sections the firm always includes.
• Quote check. Reads a draft quote and checks it against the current price list and the discount rules. It flags a line that breaks a rule and leaves the quote unchanged.
• Project brief. Turns meeting notes into the one-page brief the delivery team expects, in their format.

Every skill packages four things: the procedure step by step, the domain rules, the tools it may call and the output format.

How it runs

• The skills live in the client’s repository, next to the agents that use them.
• The assistant the staff already use loads the library. Behaviour comes from the skill, not from the chat history.
• A new version is published only after it passes a set of examples: inputs, and the output the company considers correct.
• A rule change is a new version of one skill, with the examples updated to match.

What the client gets

• A library of tested, versioned skills.
• The examples that define “correct” for each skill.
• A short guide, so their own team can add the next skill.
• Ownership: the library stays in their repository when the engagement ends.

The result

The assistant follows the company’s process. A change to a price rule shows up the next time the skill runs, for everyone.