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

# Instructions, knowledge and lead categories

> Steer how the CRM AI classifies replies and writes drafts with instructions, a knowledge base and your own lead categories.

Three settings decide what the CRM AI does with a reply. **Lead categories** say how replies are sorted. **Instructions** say how to behave. **Knowledge** gives it facts about you. All three are per organization, and they apply to email, LinkedIn and WhatsApp replies alike.

## Instructions

Instructions are standing guidance added to the AI's prompt. Open **Settings → AI → Instructions**. There are two lists, shown as two tabs with a count each:

* **Reply drafting**: added to every draft the AI writes. Use it for tone, length, sign-off and what to offer or avoid. For example: "Sign off as Priya from Youse. Keep emails under 120 words. Never offer discounts; suggest a 20-minute call instead."
* **Reply classification**: added to every classification of a reply or call. Use it to settle edge cases. For example: "A referral to a colleague counts as Interested, not Other. Replies from agencies are Not interested."

<Steps>
  <Step title="Add an instruction">
    Click **Add instruction**. Give it a short title and write the guidance. The AI receives each instruction as its own titled section.
  </Step>

  <Step title="Organize">
    Expand or collapse instructions, duplicate or delete one, and use **Copy all** or **Export JSON** to keep a backup.
  </Step>

  <Step title="Save">
    Click **Save**. The page shows **Unsaved changes** until you do, and **Discard** drops your edits.
  </Step>
</Steps>

Limits: 50 instructions per list, titles up to 200 characters, each instruction up to 8,000 characters, and 20,000 characters across a list. The page warns before you reach them.

Instructions steer the AI but do not unlock anything. It still never sends a message itself, and still does not state facts that are not in the context it is given.

## Knowledge

Knowledge is the product, pricing and policy material the AI grounds its drafts in. Open **Settings → AI → Knowledge**, then click **Add document**.

Each document has:

* **Title**, such as "Pricing and plans".
* **Kind**: Custom, Company, Product, Pricing, FAQ, Scheduling, Objection or Case study.
* **Content**, written in Markdown. Headings and lists help the AI find the right part.
* **Tags**, comma separated, for example "pricing, onboarding".
* **Always include in draft context**. Documents marked **Always included** go into every draft. Keep these short.

Every saved edit creates a new version. **Archive** takes a document out of drafts, and **Restore** brings it back. The **Active** and **Archived** tabs list each.

For each draft the AI is given the always-included documents first, then the best matches for the conversation, up to a limit of about 2,000 tokens. Longer excerpts are cut off to fit. Under a draft, **Knowledge used** shows which excerpts it relied on, so you can check where a claim came from. The classifier also receives your knowledge as background about you, not about the person replying.

<Tip>
  Put the facts people ask for in separate short documents, such as pricing, scheduling and one common objection. A single long document wastes the limit on text that is not relevant.
</Tip>

## Lead categories

Open **Settings → CRM & leads → Lead categories**. Replies are sorted into four fixed categories: **Interested**, **Customer**, **Not interested** and **Other**. Under each are **subcategories**, which you can change. Each row shows:

* **Subcategory**, with how many active records it holds and its funnel stage, or "Not a funnel stage".
* **Classification description**: tells the AI when a reply belongs here. A subcategory with no description shows a prompt to add one. Be concrete: "The person explicitly asks for a product demo".
* **Sequence**: the [sequence](/crm/sequences) that starts for it, or **No sequence**.

Use **Add subcategory** in a category to create one. Its dialog asks for a **Name**, a **Classification description** and a **Funnel stage**. Use **Edit** on a row to change them. **Delete** removes a subcategory, but not one that still holds CRM records or classification history.

The **Funnel stage** is a number. The AI can move a lead only to the same or a higher stage. Leave it empty for outcomes that are not funnel stages, such as Out of Office or Not interested. See [Pipeline](/crm/pipeline) for the starter set and how stages interact.

<Note>
  New organizations start with a default set of subcategories but no sequences. Add yours and attach them here.
</Note>

## How they work together

When a reply arrives, the classifier sees your categories and their descriptions, your classification instructions and your knowledge. It returns a category, a subcategory, a confidence and a reason. After classification, the draft is written from the conversation, the sequence step's goal, your drafting instructions and the matching knowledge. If a result looks wrong, fix the wording of the description or the instruction rather than correcting every reply by hand.

## Related

<CardGroup cols={2}>
  <Card title="Sequences" icon="list-checks" href="/crm/sequences">
    Build the flows you attach to subcategories.
  </Card>

  <Card title="Action required" icon="inbox" href="/crm/action-required">
    Review the drafts these settings shape.
  </Card>

  <Card title="AI provider" icon="sparkles" href="/workspace/ai-provider">
    Choose the model that runs all of this.
  </Card>

  <Card title="Pipeline" icon="kanban" href="/crm/pipeline">
    Funnel stages and the default pipeline.
  </Card>
</CardGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.