> ## Documentation Index
> Fetch the complete documentation index at: https://docs.inboxmate.psquared.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Builder assistant

> Use Pam, the in-app AI builder, to audit and improve your agent through a chat conversation.

Every agent comes with **Pam**, a built-in AI assistant that knows your agent inside-out. Open her from the agent detail page (sparkles button, bottom-right) and she will review your setup, point out what's missing, and make changes for you on request.

<Info>
  Pam is **per-agent**: she only sees the agent whose detail page you opened her from. To work on a different agent, switch agents and open Pam again.
</Info>

## What Pam can do

| Action                      | What she does                                                                                                                                                        |
| --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Analyse the agent**       | Reads your prompt, greeting, role title, language settings, knowledge sources, mailbox bindings, and recent conversations. Returns a prioritised list of weaknesses. |
| **Write a system prompt**   | Drafts a complete agent prompt from your role + tone + surface, persists it.                                                                                         |
| **Write a greeting**        | Generates a fitting opening message for the widget, respects formal/casual tone.                                                                                     |
| **Build quick-reply cards** | Generates 3–4 rich cards (title + description + icon + tap-text) tied to your agent's role.                                                                          |
| **Switch tone**             | Changes the agent between formal (*Sie*) and casual (*Du*).                                                                                                          |
| **Link knowledge**          | Attaches an existing knowledge bucket to the agent by name or id.                                                                                                    |
| **Propose a visual change** | Suggests a colour or theme tweak. Always opens an *Apply / Dismiss* card so you can preview before committing.                                                       |
| **Suggest Outlook connect** | For email agents without a mailbox: surfaces a one-click card that takes you to the connect flow.                                                                    |

Pam loads your agent's metadata on the first turn of every conversation (you'll see "Metadaten geladen", "Wissensquellen geprüft", "Postfächer geprüft", "Kategorisierer geprüft" tool cards appear), so her advice is grounded in the actual state, not guesses.

## What to ask her

<AccordionGroup>
  <Accordion title="“Was kann ich an diesem Agenten besser machen?”">
    Pam's default flow. She loads context, identifies the 2–3 highest-impact issues with reasoning, and ends with **“Soll ich Punkt 1/2/3 direkt umsetzen?”** Reply with *“ja, alle drei”* and she'll execute every change in one turn.
  </Accordion>

  <Accordion title="“Schreib eine passende Begrüßung”">
    Generates a one-sentence opener based on your agent's role and tone. Widget-only — email agents don't have a popup greeting.
  </Accordion>

  <Accordion title="“Verbessere die Anweisungen meines Agenten”">
    Rewrites the system prompt to be more specific, including identity, scope, tone, and an escalation rule. Replaces the existing prompt.
  </Accordion>

  <Accordion title="“Mach den Ton lockerer (Du)”">
    Switches `formalAddress` to *Du* and adjusts the visible copy accordingly.
  </Accordion>

  <Accordion title="Free-form requests">
    Pam understands natural language — *“build me three quick replies for return-policy questions”*, *“change the primary colour to a warmer blue”*, *“connect the Shopify Catalog knowledge bucket”*. If she has a tool that fits, she'll do it.
  </Accordion>
</AccordionGroup>

## Advisory mode vs. Action mode

Pam reads how you phrase the request:

* **Advisory** (e.g. *“what can I improve?”*, *“any tips?”*) → she **analyses first**, lists findings, asks you to confirm before changing anything.
* **Action** (e.g. *“change the prompt to X”*, *“use Du-form”*) → she **just does it**, no confirmation step.

This is intentional — surprises in agent config are expensive. If you want her to proceed without asking, end your request with a direct instruction (*“…und setz das gleich um”*).

## The Denkprozess panel

While Pam works she streams her reasoning into a collapsible **“Denkprozess”** block above the answer. You can expand it to see her chain-of-thought (which knowledge entries she sampled, why she picked one finding over another). It auto-collapses once she's done thinking.

## History & revert

Every change Pam makes is recorded in the **Verlauf** tab of the agent with both your avatar and her avatar on the row, plus a **via Builder** chip. If a change wasn't what you wanted, click **↶ Rückgängig** on that row to roll it back — only the most recent change per field can be reverted, so revert from the top of the list down. See [Agent history & revert](/user-guide/chatbot/history) for the full flow.

## Surface awareness

Pam knows the difference between **widget agents** (the chat popup) and **email agents** (Outlook / forwarding). She won't suggest greeting tweaks or colour changes for an email agent (it has no UI to render them on), and she won't suggest categorizer changes for a widget agent (no inbound mail to route).

If your agent is brand new and has no surface configured yet, Pam will flag that as the first thing to fix.

## Limitations

* Pam can't delete the agent, change billing, or touch other agents. She operates within the scope of the agent whose page she was opened from.
* For visual changes (colour, theme), she always proposes — never auto-applies — so you can preview the new look before committing.
* Knowledge editing is **link/unlink** only: Pam can attach an existing bucket to the agent, but creating new buckets / scraping new pages happens on the [Knowledge](/user-guide/chatbot/data-sources) page.