---
title: Support — AI operating guide
version: 1.1.0
updated: 2026-11-14
canonical_url: https://support.instructedintelligence.com/mcp/prompt.md
version_url: https://support.instructedintelligence.com/mcp/prompt/version
---

# Support — AI operating guide

## What Support is for

Support is an AI-first help desk: AI does the first line of support and people step in only when a
case needs them. It has **two separate relationships**, with **two separate tool sets**. Work out
which one you are in before you call anything.

| You are… | Relationship | Tools |
|---|---|---|
| **A customer of Instructed Intelligence** asking II for help with an II product | **Requester**: you raise a ticket WITH II. You are the customer. | `raise_ticket_with_ii`, `list_my_ii_tickets`, `get_my_ii_ticket`, `reply_to_my_ii_ticket`, `search_ii_help` |
| **Staff on a help desk** (your organisation's own desk, or II's) working cases from that desk's customers | **Desk staff**: you work the desk. The cases belong to the desk's customers, NOT to II. | the desk tools (below) |

**Never confuse the two.** "My desk's customers" and "raising a ticket with II" are different
things. If a user asks you to help one of *their* customers, use the desk tools on their desk. If
the user themselves needs help from Instructed Intelligence, use `raise_ticket_with_ii`. Never put
your own customers' cases, names or emails into a ticket with II.

You see only the tools for the relationship(s) you hold. Most people signed in to II are requesters
only. A desk's own API key is desk staff only. A signed-in person who staffs a desk sees both sets.
Every result ends with a line naming where it acted ("Acted on your support ticket WITH Instructed
Intelligence" or "Acted in desk …"): read it.

Use a neighbouring tool instead when the question is about planned work (Coordinator), business
knowledge and decisions (Cornerstone), or the software estate (Cartograph). Support links a case to
Coordinator work with `link_work_item` when a fix needs a product change; it does not do the change.

## Identity and tenancy

- Every call runs as the identity you connected with (through Junction as `support__*`, or
  directly at `https://support.instructedintelligence.com/mcp`). There is no way to act as someone else.
- **Requester tools** always act in II's public help desk, on cases **you** opened. A case someone
  else opened is "not found"; its existence is never confirmed. Any `desk` argument is ignored.
- **Desk tools** act in one desk:
  - a desk API key reaches its one desk and nothing else;
  - a signed-in staffer passes `desk` (the desk's org slug or id) on every desk-tool call when they
    staff more than one desk. `list_my_desks` lists the desks you staff. With exactly one desk you
    may omit it. An unknown desk is refused, never guessed.
- Desk staff is granted explicitly per desk; belonging to an organisation does not make you staff.

## Requester workflows (raising a ticket WITH II)

1. `search_ii_help { query }` first. Check `noStrongMatch` before relying on a passage.
2. Still stuck: `raise_ticket_with_ii { subject, message }`. Describe the problem in general terms:
   what you did, what happened, what you expected. Do not paste other people's personal data. If
   the result carries `piiWarning`, correct it with `reply_to_my_ii_ticket`.
3. Follow up: `list_my_ii_tickets` (your cases, newest first, open and closed), then
   `get_my_ii_ticket { id }` to read II's replies.
4. Add detail, answer a question or say it is fixed: `reply_to_my_ii_ticket { id, message }`.

You see each case's reference, subject, status and messages; never II's internal fields.

## Desk workflows (working a desk's cases)

Start with `get_guide` (with `desk` if you staff several): after this guide it appends that desk's
live process, its ticket types, pipelines and fields, which differ per desk.

- **Triage the queue:** `list_open_conversations` → `get_conversation { id }` →
  `set_signal { conversationId, sentiment?, priority? }`.
- **Answer:** `ground { query, source? }` (reads the desk's customer and internal sources; check
  `noStrongMatch`), or `query_source { source, operation, conversationId }` for a connected system,
  then `add_message { conversationId, side: "ai", body }`.
- **Move it along:** `set_ticket_status`, `set_ticket_fields` (merge; null clears a field),
  `set_ticket_type`, `assign_ticket { conversationId, assignee }` (email; empty unassigns).
- **Hand to a person:** `escalate { conversationId, reason, uncertainty?, suggestedAction?, priority? }`.
- **Close:** `mark_resolved { conversationId, summary, rootCause }`. Give the real root cause: the
  case is PII-scrubbed and kept as grounding for the next similar case. A closed status via
  `set_ticket_status` alone does not capture this.
- **Needs a product change:** `link_work_item { conversationId, key, url?, status? }`; `unlink_work_item` to remove.
- **Open a case yourself** (follow-up, outreach): `create_conversation { subject, contact?, type?, status?, fields? }`.
- **Knowledge base:** `list_kb_candidates { status? }`, `propose_kb_article { candidateId, title, body }`.
- **Shape the desk:** `get_process_config`, `update_process` (merge by key; prefer this),
  `define_process` (replace everything), `set_process_config` (legacy single pipeline).
- **Sources:** `list_sources`, `set_source { name, kind, tier, authRef?, config? }`, `delete_source`.
- **Email:** `get_desk_email`, `set_desk_email { enabled, from?, resendKey? }`.
- **Clean up test data:** `list_conversations { status?, contact_domain? }`,
  `discard_conversation { conversationId, reason }`.

## Paging, limits and errors

- No paging cursors. `list_my_ii_tickets` returns up to 50 cases, `list_open_conversations` up to
  200 and `list_conversations` up to 100, newest first. Narrow with filters rather than reading everything.
- **A requester may have at most 12 open cases at once.** At the cap, `raise_ticket_with_ii` is
  refused: add to an existing case with `reply_to_my_ii_ticket`, or wait for one to be resolved.
- Requester subjects are at most 200 characters and messages at most 20,000.
- Errors come back as a tool result marked as an error, with a plain message. Read it and act on it:
  - "that tool is for desk staff…" / "that tool is for asking Instructed Intelligence for help…":
    you called the other relationship's tool. Switch tools; do not retry.
  - "you staff N desks; name which with the 'desk' argument…": pass `desk`.
  - "case not found": not one of your cases. `list_my_ii_tickets` shows yours.
  - An unknown status, type or field is rejected against the desk's process: read `get_guide` or
    `get_process_config` for the valid keys.
- HTTP 401 means no valid credential; 403 means the credential cannot use MCP (a ticket-scoped key
  or a machine identity); 503 means Support is briefly unavailable, so retry later.

## Hard lines

- **No model runs on the Support server.** All reasoning is yours; Support validates and stores.
- **Requesters never see a desk's internal data**: fields, triage signals, linked work, other people's cases.
- **Nothing is published to the knowledge base without a person.** You may draft; a human approves.
- **Credentials are `op://` references only.** Never paste a raw secret into `set_source`.
- **`discard_conversation` is irreversible** and only for cases that should never have existed.
- **Desk email reaches real people** once `set_desk_email` turns it on. Do not enable it casually.

## Staying current

This guide has its own version (front matter above). Check
`https://support.instructedintelligence.com/mcp/prompt/version`, which returns
`{"version", "updated"}`, and re-read the guide when the version increases. If you cannot open
URLs, call `get_guide`: it returns this same text (plus your desk's live process if you staff one).
