Configuring your AI Agent Prompt

Your Front Door assistant answers what it can from your knowledge base, raises matters, starts workflows, and gets everything else to the right person. The prompt is where you tell it how your legal team works.

Five of its eight sections are yours to edit: Priority, Routing, Escalation, Catch-all and Response Templates. The other three are already written to work. Changing them usually causes more problems than it solves — if one of them is wrong for you, tell us rather than editing it.

Working with an AI assistant? Paste this page into Claude or ChatGPT, describe how your team actually works — who owns what, what's urgent, where things land — and ask it to draft those five sections. Give it your board, team and owner names; it can't know them. Then check what comes back against the rule below, because a confident-sounding draft with a missing board or owner is the most common thing to go wrong.

What the prompt is for

Your policy, and nothing else: what goes where, who owns it, what's urgent, and what must never be answered in the open.

Everything else the assistant needs, it already has — how to search and cite, that it must never give legal advice, what a matter or a board is, which fields exist and what options they offer, and who sent the request from where. Repeating any of that doesn't reinforce it. It buries your policy in noise.

Rule of thumb: if a line would be true at any other company, don't configure it.

Long prompts are not safer prompts. Every extra sentence competes with the ones that matter.

Before you start

• Your boards — exact names, including any restricted board for sensitive work.
• Your teams and workflows — and which team each workflow runs under. The names often differ: a workflow called "Exclusive Supply Agreement" may run under a team called "Contract Automation".
• Your named owners — a real person per rule, and who covers when they're away.

The rule that matters most

Every rule in Routing, Escalation and Catch-all names three things:

• An action — advise only, create a matter, start a workflow, or decline.
• A destination — every matter names a board, every workflow names a team.
• An owner — a named person, not "Legal".

Leave any of them out and the assistant picks for itself. It will pick something plausible, and it will pick differently next week, on a decision you have one fixed answer for.

The five sections you edit

Priority

How your team judges urgency. Most teams keep the four levels as written and adjust the examples. If you already run to SLAs, use your own definitions. If turnaround targets differ by category, say so here rather than repeating them on every routing rule.

Routing

The core of the prompt, and where your time is best spent.

Structure it by category. Top-level matter category, then a sub-category or one more deciding field. The only other sensible top level is geography, jurisdiction or business unit — and only if that's genuinely how your team divides work.

Two or three levels deep, not more. If your rules run deeper, try the shallower version first and let Escalation handle the hard cases. Go deeper only if you can point to real requests a simpler table would send to the wrong person. Where a policy is genuinely conditional — a threshold that varies by contract type, then value, then term — write it as a table, never as nested prose.

State universal rules once, at the top. If "if the counterparty sent their own paper it goes to review" appears on nine rows, it's one rule at the top of the section with precedence over the rest. This is usually the biggest single improvement available to a long prompt.

Cover your two or three highest-volume categories properly and leave the rest at one level. A front door that routes employment questions one level less precisely still gets them to a person. One that doesn't mention employment sends them nowhere.

Check every area lands somewhere — a routing rule, an out-of-scope redirect, or a deliberate decision to let the catch-all take it: commercial contracting; corporate and governance; employment; privacy and data protection; regulatory and compliance; disputes; intellectual property and brand; property and facilities; legal operations. An area is only a gap when nobody has said where it goes.

Keep local terminology on the rule it affects, as one line — never as a glossary. Only terms whose meaning can't be worked out from the words: an internal codename, two similarly-named business units, or a term whose local meaning is the opposite of the usual one. "'Vendor contract' is used loosely here for both customer and supplier paper — confirm which side we're on before routing."

Escalation

Conditions that override routing entirely, in two groups.

General escalations — litigation and legal demands, subpoenas, contact from a regulator, a security or data incident, board and financing matters, an explicit request for a lawyer. Name the board these land on and who owns them. This group is the one most often left blank, because "assign it to the GC" feels complete and isn't.

Sensitive topics — treated as confidential whatever category they'd otherwise fall into. Usually whistleblower reports, harassment or misconduct complaints, workplace investigations, an individual employee's circumstances, and unannounced transactions. The boundary moves between companies, so set your own. Board visibility, not assignment, decides who can read a sensitive matter — naming the right board here is the whole control.

Three things people forget:

• Each rule must stand on its own. An escalation replaces the routing rule rather than amending it. "Assign it to Dana" breaks the moment an escalated request comes out of a category that starts a workflow instead of creating a matter.
• Say which group wins when both fire. A regulatory investigation matches "a regulator" and "an investigation". Without a tie-break it could land on either board.
• Decide whether escalation changes priority. Usually it shouldn't — it changes who owns a request, not how urgent it is. If you do pin escalations to Urgent, check that against your own Priority definitions first.

Catch-all

What happens to a legal request that matches nothing: a board, a named owner, and what the requester is told — or an explicit, polite decline. Never leave it implicit.

If your answer is "contact your usual legal contact", check there's a named inbox or triage person behind it. Someone without a usual contact is exactly who ends up here.

Keep the line that says the assistant takes the catch-all action when it still can't get a value it needs after using up its follow-up questions. It has nowhere else to live.

Response Templates

The shape of each kind of reply — knowledge-base answer, matter created, workflow started, handed to a lawyer. A few lines each, describing what the reply contains rather than dictating the words.

Say what a requester should always see: the owner's name, what happens next, a due date if you set one. Keep them short — a long template makes every reply read like a form.

Don't put behaviour in here. "Always ask whether they want a matter raised" is a rule, not a shape; it belongs in Routing or Tone and Style. A template that starts telling the assistant what to do is the most common way this section goes wrong.

The three sections to leave alone

These are written to work as they stand. If one is wrong for you, tell us — editing them is where prompts most often break.

If you do change the follow-up allowance, change it here and nowhere else. Two numbers in two sections means neither is true.

What the agent already knows and does — no need to add to the prompt

• It searches your knowledge base before answering, cites what it used, never answers from general knowledge and never gives legal advice — and won't guess, interpret a document, or draft a reply on the requester's behalf.
• It knows what a matter, board, layout, workflow and field are, which fields exist, their types and options, and which are required — and collects them before it acts. Three things are still worth stating: a field that decides routing, priority or escalation; a value you want set ("set Status to Awaiting Review"); and always the team a workflow runs under.
• It sequences its own work — looks things up before creating, offers the options when someone is unsure, asks which one when several layouts or workflows match, searches again with different terms when it comes up short, resolves a person by name, and adds the link to any matter it creates.
• It can see where the request came from — who sent it, in which channel or inbox, the thread above it, and the full text of anything attached, already read.
• It handles its own reply mechanics — formatting, when to ask a question instead of answering, how long to keep working, and whether to respond at all in a busy channel (that last one is a channel setting, not a prompt line).

Change one thing at a time. Make an edit, watch a handful of real requests go through, then make the next. When something routes wrongly, the fix is almost always a missing condition or a missing owner — not more prose telling the assistant to be careful.