> For the complete documentation index, see [llms.txt](https://herd-security.gitbook.io/herd-security-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://herd-security.gitbook.io/herd-security-docs/ai-coach/ai-coach.md).

# AI Coach

{% hint style="warning" %}
**Beta — access by request.** AI Coach is currently in beta and enabled per organization. If you don't see **AI Coach** in your left-hand navigation, contact your Herd representative to request access.
{% endhint %}

## Overview

**AI Coach** helps security and enablement teams understand — and improve — how their people actually use AI tools like **Claude**, **ChatGPT**, and **Microsoft Copilot**. Once you connect your organization's AI accounts, Herd scores how effectively each conversation was prompted and turns the results into targeted, friendly coaching.

The goal is upskilling, not surveillance. AI Coach is explicitly a **coaching signal, not a performance evaluation** — it describes *prompting habits*, not the person.

To open it, click **AI Coach** in the left-hand sidebar. You land on the **Overview**, and the section is organized into three tabs:

* **Overview** — your organization's AI posture, leaderboards, and score trend (plus the Setup/Connected flow for admins).
* **Conversations** — the scored-conversation list and per-conversation scorecards.
* **Findings** — AI policy violations and coaching signals, with triage actions.

***

## How scoring works

Every scored conversation gets an overall **prompting-effectiveness score from 0–100**, plus a breakdown across eight prompting-craft **dimensions**. Each dimension is rated on an anchored 0–4 level:

| Level | Label     |
| ----- | --------- |
| 0     | Absent    |
| 1     | Weak      |
| 2     | Adequate  |
| 3     | Strong    |
| 4     | Excellent |

A dimension can also be **N/A** when it doesn't apply to a given conversation (for example, tool awareness in a plain browser chat).

### The dimensions

| Dimension                         | What strong prompting looks like                                                   |
| --------------------------------- | ---------------------------------------------------------------------------------- |
| **Goal clarity**                  | States the desired outcome unambiguously.                                          |
| **Context sufficiency**           | Provides the background, data, and inputs the task needs — enough, not excessive.  |
| **Constraints & requirements**    | Specifies explicit requirements, output format, and limits.                        |
| **Success criteria**              | Defines what a good or finished result looks like.                                 |
| **Front-loading**                 | Supplies complete requirements up front instead of drip-feeding them across turns. |
| **Iteration & course-correction** | Gives clear feedback and corrects course across turns.                             |
| **Scope & decomposition**         | Right-sizes the ask and breaks large tasks down.                                   |
| **Tool & environment awareness**  | Leverages the tools, files, and capabilities available to the assistant.           |

Herd also surfaces a **model-fit** signal on each conversation — whether the model the person chose was a good fit for the task.

***

## Privacy by design

AI Coach is built so you can act on *how* people prompt without reading *what* they wrote.

* **Coaching, not evaluation.** Scores and tips describe prompting habits and are framed as a growth signal — not a rating of the person.
* **Never chat content.** The summaries shown in the conversation list and on each scorecard are dimension-derived one-liners. Admins — even super-admins — triage on the **score and coaching**, never the back-and-forth of the chat itself.
* **Aggregated, scores-only leaderboards.** Team standings are averages, and the individual leaderboard (off by default) shows names, ranks, and scores only.
* **People see their own.** Users can review their own scored activity and get prompt rewrites through the Herd Slack app.

***

## What you can do with AI Coach

* [**Setting Up AI Coach**](/herd-security-docs/ai-coach/ai-coach/setup.md) — Connect your Claude, ChatGPT, or Microsoft Copilot account and map AI activity to your people.
* [**Reviewing AI Use**](/herd-security-docs/ai-coach/ai-coach/reviewing-ai-use.md) — Read the Overview's posture cards, browse scored conversations, read a conversation's scorecard, and coach an individual.
* [**Findings**](/herd-security-docs/ai-coach/ai-coach/findings.md) — Review AI policy violations and coaching signals, and triage them without ever exposing chat content.
* [**Leaderboards & Remediation**](/herd-security-docs/ai-coach/ai-coach/leaderboards-and-remediation.md) — Turn on leaderboards and self-service, and auto-assign coaching when a pattern of weak prompting emerges.