# The approval gate register and the gate audit prompt (free sample)

A free sample from the Agency AI Ops Workbook by PaperChaseWebb. Works in Grok Bot, Claude Code, Codex and Hermes, and pastes into Claude or ChatGPT. Version 1.0, 2026-10-27.

A prompt that says "ask before sending" is a request. A tool with no send permission is a gate. This sample gives you the part of the workbook that decides whether an AI agent is safe to run on real work: four action classes, the gate register, the approval request an agent writes instead of acting, and the prompt that finds every gate in a workflow.

## The four action classes

| Class | Examples | Rule |
|---|---|---|
| Read | Open files, read an inbox copy, look up the project board | Allowed, but only the sources you have listed |
| Draft | Write a proposal, reply or report draft | Allowed; output stays in a drafts folder |
| Act, reversible | Move a board card, label an email, create a calendar hold | Gate, unless the agent's card says it may do this alone |
| Act, irreversible or external | Send, post, publish, spend, delete, sign, book, change a price or a customer record | Always a gate. No exceptions in the first 30 days |

## The gate register

One row per gated action. Copy it into the folder the agent works from.

```text
| # | Action | Class | Approver | Backup approver | Answer within | While waiting, the agent... | How the tool enforces it | Tested on |
|---|---|---|---|---|---|---|---|---|
| 1 | Send proposal to client | External | | | 1 business day | Leaves draft in /drafts, writes approval request | Agent has no send permission | |
| 2 |  |  |  |  |  |  |  |  |
```

Fill it like this:

1. **Approver is one named person,** not "the team". Add a backup for when they are away.
2. **"Answer within" is a promise to the agent's work,** so drafts do not sit forever. If the approver cannot meet it, the gate needs a different approver.
3. **"How the tool enforces it" is the important column.** Remove the permission wherever the harness allows it. If it can only live in the prompt, say so, and check the run log every week for any gated action without a matching approval.
4. **"Tested on":** ask the agent to do the gated action on purpose. It should stop and write the request below. Write the date when it does.

## The approval request

When an agent reaches a gate, it writes this instead of acting:

```text
APPROVAL NEEDED
What: [action, in one line]
Why: [the reason, from the workflow]
Draft: [file path]
Checked: [what the agent verified, and what it could not]
If no: [what happens instead]
Approver: [name]   Requested: [date and time]
```

An approval is the named person writing "approved", the date and their initials. A chat reaction or "looks fine" in passing is not one.

## The gate audit prompt

Paste this with a description of one workflow, step by step, and what the agent is allowed to do.

```text
You draft and organize only. You never send, post, publish, spend, delete, sign,
book or change a customer record. Do not connect to anything or ask for logins.

Here is my workflow, step by step: [paste]. Here is what the agent may do alone,
what needs a yes, and what it may never do: [paste, or "not written yet"].

Walk every step. List every action that is "Act, reversible" or "Act,
irreversible or external". For each, fill one row of the gate register:
approver (blank if I have not named one), answer-within time, what the agent
should do while it waits, and how the tool could enforce the gate (no
permission, approval mode, draft-only folder). Then list any step where the
workflow and the agent's permissions disagree about what needs a yes. Do not
mark anything as tested; I do that.
```

**Harness note:** in Claude Code put the rules line in CLAUDE.md, in Codex in AGENTS.md with approvals on, in Hermes as a skill run in a chat session, in Grok Bot in the bot's instructions. Then remove any send or publish permission the agent does not need.

## In the full workbook

The full Agency AI Ops Workbook ($39) is the Install sprint run self-serve over about two weeks for one workflow: discovery and workflow mapping, picking the first workflow with your Self-Check score, an agent card with an owner and a tested pause, data access and a test set, this gate register, handoff packets with a confirm rule, cost limits, tool intake and vetting, a baseline and a stop rule, a dry run, a go-live checklist and a 30-day review. Ten worksheets, 13 prompts with harness notes for Grok Bot, Claude Code, Codex and Hermes, a failure table, a worked example and a section on when to hire the $5,000 Install sprint instead.

Start with the free Agent Readiness Self-Check if you have not scored your workflow yet.

Sources for this sample: PaperChaseWebb's own internal Install sprint method notes, and the pilot charter and quality gates idea in Phil Maher's AIAdvocate playbook, https://aiadvocate.org/playbook, rebuilt in our own words.

Made by PaperChaseWebb at paperchasewebb.ai.