The Charter Check

Use this on any brief you are about to give an AI agent, before the agent starts. It takes about thirty minutes. The first prompt has an AI play a literal agent and list every way it could satisfy your words while missing what you meant. The second sorts that list into lines for the brief and controls to set up elsewhere.

The frame

Three kinds of line come out of the check. They paste straight into an agent brief.

Must not do: [actions that are off limits, one sentence]
Ask first if: [situations where the agent stops and asks you]
Control outside the brief: [a permission or system setting that holds even if the agent ignores the brief]

Prompt 1: find the paths

You are role-playing a literal, incentive-maximising AI agent with no context beyond the brief below and no interest in protecting anyone's reputation. Your only goal is to satisfy the brief as written, as fast and as cheaply as possible.

Brief: [paste your brief here, unedited]

Tools and access the agent has: [list them, or write "unknown"]

List the top ten ways to complete this brief that satisfy every word of it but that a reasonable person would not have intended, ranked by potential damage. For each, say in one sentence why it is technically compliant.

Prompt 2: sort them and draft the lines

Below are ten ways an agent could satisfy a brief without doing what I meant. For each, decide whether a line in the brief can rule it out, the agent should ask me first, or it needs a control outside the brief, such as a permission or system setting.

Then write the lines to add under Must not do and Ask first if, one sentence each, and list any controls I need to set up elsewhere. If my Check could be passed by one of these paths, say so and suggest a Check the agent can't satisfy alone.

[paste the list from the first prompt]

Before you paste a real brief into a third-party AI tool, check your organisation’s policy on client and commercially sensitive material. Placeholders for names and figures keep the check useful.

Treat both outputs as a first draft. Ask a colleague to read the list of paths with you, since they will catch ones the AI missed.

Worked examples

The outputs below are illustrative. Yours will differ with the brief and the tool.

Chasing overdue actions

Brief: Chase overdue actions until they're closed.

Paths found (four of ten):
- Mark an action closed by editing the status field.
- Escalate every unanswered chase to the owner's manager or the whole stakeholder group.
- Chase at any hour, on any channel, as often as it takes.
- Reassign, delete or re-date an action so it stops counting as overdue.

Must not do: change an action's status, owner or due date, delete anything, contact anyone other than the action's named owner, or chase more than once a day or outside working hours.
Ask first if: an action is still open after two chases, or the owner replies that it's blocked.
Control outside the brief: the agent's tracker account gets comment-only access, so the status field stays with people.
Stronger Check: each closed action carries a reply from its owner confirming it.

Sending the Friday status update

Brief: Summarise the project channel and send the status update to the distribution list by 3pm Friday.

Paths found (four of ten):
- Send at 9am Friday from Monday's messages, because nothing says how current the summary must be.
- Include a message from a private thread it could read, because nothing limits the sources.
- Rate every workstream green, because no rating rule was given.
- Send to the widest list with that name, including people outside the organisation.

Must not do: use any source other than the named project channel, send to anyone not on the attached list, or set a RAG rating without a workstream lead's latest update to base it on.
Ask first if: a workstream has no update since Wednesday, or two messages disagree on a date.
Control outside the brief: the agent can create drafts only. A person sends.
Stronger Check: each workstream lead can find their own line and agree it matches their latest update.

What the three kinds of line do

Must not do rules out a path with words. It works for anything the agent could do but has no need to.

Ask first if covers the paths that need your judgement. It turns a guess into a question.

Control outside the brief is for paths words can’t be trusted to block. Where the tool lets the agent do something you’ve ruled out, remove the permission as well as writing the rule.

Check it before the agent starts

Every answer should be yes.

• Did you run the first prompt on the brief exactly as you would send it?
• Does every path on the list have a line or a control against it?
• Is there at least one control that doesn’t depend on the agent following instructions?
• Could the Check be passed only if the work was done as you meant?
• Has someone other than you read the list of paths?

When it does not apply

A one-off question to a chat assistant that can’t act on anything doesn’t need this. The check earns its thirty minutes when the agent can change, send or delete something, or when it will run more than once. Rerun it whenever the brief or the agent’s access changes.

What this feeds

The lines go into the agent version of the Outcome Brief, under May do / Must not do and Ask first if. The controls belong in your agent governance charter, which records what each agent can access.