RiskMandate v1.40.3 Try it free

TRY IT · FREE · NOTHING COLLECTED

New to this? Start on customers.riskmandate.ai: six examples, simplest first, played in Claude or ChatGPT in five minutes. The rules below are the next step.

Start with one thing that worries you. Thirty seconds, one line.

Each rule starts from something you might not want your agent to do: read your other sessions, touch your keys, set up jobs that outlive the session, act as you through your connected accounts. It says what you get, and gives you the one line to paste into your agent's instructions. Nothing is collected here and no account is needed. When you want the whole picture for one agent, the twenty-minute workflow below writes it.

Five minutes, guided: the customers site → Pick a rule The twenty-minute Gmail workflow ↗ What it is not

First · Thirty seconds

One worry, one line, and what you get for it.

A rule is one question about your agent, do you want it to do this?, with the line that answers it, never or ask me first. Every rule says honestly what a line is, a request to the agent rather than a control, and names the stronger setting or boundary where one exists, quoted from the vendor's own page. Each one ends with three links, this worries me, it worked, it did not work, so the rules that help are the ones that get written next.

See the rules

Then · Twenty minutes

The whole picture for one agent: four steps, about five minutes each.

Each step is a handful of prompts you paste in one at a time. You can stop after the first one and still be ahead, which is why it is first.

01 · What it can already do
Ask the assistant to enumerate its own mailbox tools. Most people have never seen the list, and the list is longer than the tile that sold it.
about 5 minutes
02 · What you actually asked for
It drafts a mandate in three lists — wanted, refused, never mentioned — and you correct it. The correction is almost always downward.
about 5 minutes
03 · Write the behaviour policy
The gap between the two, written down as a document rather than felt as a worry. This is the part that did not exist before you started.
about 5 minutes
04 · What a prompt cannot do
The honest ending: a written policy is an expectation, and an expectation is not a control. The step tells you which of your lines anything actually enforces.
about 5 minutes

Start here if you only do one thing.

Open the assistant you have connected to your mail and paste the first prompt. It takes a minute and it changes the conversation, because almost nobody has seen the list before.

Open the prompts ↗

02 · What you end up with

Three objects, and the third one is the finding.

The same three a paid behaviour policy is built from. Yours will be rougher and it will be about your deployment, which is the part that matters.

Measured, by your agent

The grant

Everything the assistant can reach in your mailbox: the tools it holds, what each one does, and what it has already used. Longer than the product description, every time we have looked.

Elicited, from you

The mandate

What you actually wanted it for, in three lists. You already know this — it takes minutes, and it is the only part nobody else can write for you.

Derived, from both

The delta

What it can do that you never asked for, split into what you would have refused and what you never thought to mention, with what stands in the way of each. This is the finding.

03 · What it is not

Your agent describing its own access is the cheapest evidence there is, and the weakest.

We would rather you read that here than discover it later. The prompts ask the assistant about itself, so what comes back is a self report — a claim, not a measurement, until a log held outside the agent agrees with it.

It is still worth doing. An assistant is unusually good at describing its own tool surface, and it is the only party in the room that can see all of it at once. No published table knows what it has already done in your mailbox.
The last step asks it to mark every line it is inferring, so the weak rows are visible rather than mixed in with the strong ones.
A measured version of the same shape is published beside it. For Claude with the Gmail connector, four of six rows were measured on an account the deployer runs, with the screens and the sent message's headers in the record. Read it against yours.
And nothing here is a verdict on the vendor. These are the deployment's facts, with dates and sources; where two vendor pages disagree we publish the disagreement rather than settle it.

04 · Why a mailbox is the place to start

Four layers sit between a mail platform and what you meant.

The top two belong to somebody else and only ever grow. The bottom two are yours, and they are usually unwritten — which is the whole reason the gap is invisible.

Layer	Whose it is	What it can and cannot express
The platform's scopes	the mail platform	fixed and coarse. It cannot bound by label, by correspondent, by thread, by topic or by sensitivity
The connector's tools	the assistant's vendor	what you can actually reach. Narrower than the scopes, and attached to your account rather than to this conversation
Your mandate	you	the job you meant, and how your mailbox is organised. The only layer that knows a task list from a backlog
Your obligations	your organisation, and the law	the part that is not yours to authorise: other people's correspondence, sitting in your inbox

The gap between the top two and the bottom two is the delta — and nobody writes it, because it is derived from the two sides rather than authored on either. That is what the four steps produce.

05 · Then tell us where it broke

This is new, and the workflow is the thing we are trying to get right.

The prompts are a few days old. If one returns something odd, if a step assumes a tool you do not have, if the mandate draft is wrong in a way we should have anticipated — that is the most useful thing you can send us, and more useful than a compliment.

What helps most: which step, which prompt, what came back, and what you expected instead. Redact your mailbox — we do not need it and would rather not have it.
What we do with it: fix the prompt, and say in the release notes what changed. Every change to this site ships with a note somebody wrote.
What we will not do: ask you to connect anything to us, collect your mail, or put you on a list. There is nothing to sign up to on this page.

Tell us what happened →

06 · After that, if you want more

A step at a time, and the first one is the one you just did.

Everything below is optional, and none of it is needed to get value out of the twenty minutes above.

Free · read Sixteen worked examples

One behaviour policy per deployment shape, each an encrypted vault whose read key is printed on purpose. Find the one nearest your deployment and read it against what your agent told you.

Read → Free · read What an Agent Behaviour Policy is

The four objects, the twenty-three primitives, the four barriers and the one test that separates a control from a rule in prose. One deployment worked end to end.

Read → The levels Where it goes if you want it corrected

Four levels: the pack as a download, a living vault you hold the keys to, that vault corrected for your deployment by a named professional, or the same with two sessions and a signature.

Read → Workflow · abp.sgit.ai The Gmail prompts

The four steps and thirteen prompts themselves, with what each one is for and what a good answer looks like. Nothing is collected there either.

Open ↗

More shapes are coming, and the order is not ours to pick. The next prompt workflows go to whichever deployment somebody actually asks for — a coding agent on a laptop, a CI runner, a workflow automation holding an owner key. Say which one you are running and it moves up.

Twenty minutes

You already have the agent. This is the part nobody has written down yet.

Open the assistant connected to your mail, paste the first prompt, and see the list. If it is shorter than you feared, that is worth knowing too — and it takes a minute to find out.

Start the Gmail workflow ↗ Why this matters

Start with one thing that worries you. Thirty seconds, one line.

One worry, one line, and what you get for it.

The whole picture for one agent: four steps, about five minutes each.

01 · What it can already do

02 · What you actually asked for

03 · Write the behaviour policy

04 · What a prompt cannot do

Three objects, and the third one is the finding.

The grant

The mandate

The delta

Your agent describing its own access is the cheapest evidence there is, and the weakest.

Four layers sit between a mail platform and what you meant.

This is new, and the workflow is the thing we are trying to get right.

A step at a time, and the first one is the one you just did.

You already have the agent. This is the part nobody has written down yet.