All skills
Beta · FlagshipFieldwork & SamplingFieldworkWorkpaper

Control Testing Agent

Read a control description, design the test, retrieve the evidence, examine it, conclude with citations, emit the workpaper — the "robot senior auditor," end to end.

Plugin
dynamic-fieldwork-audit-testing
Version
v0.2.8
Impact
Very High
Complexity
Medium–High

How it works

Input
A control description (from the risk and control matrix or audit program) plus the prior test plan if this is a repeat test, then the sample evidence (PDFs, images, spreadsheets)
What it does
Drafts a test plan → a person reviews and approves it before anything else happens → pulls the real figures out of each piece of evidence → checks each item against the approved plan → scores it pass, fail, or exception
What has to hold true
No test plan can be used without a named approver and a timestamp. Dollar figures are never read off a picture of a document. Every row in the final workpaper has to link back to its source.

The part that makes it reproducible

Output
An approved, documented test plan, a record of what was checked for each sample item, and a finished, ready-to-review Excel workpaper

What it's made of

Each step runs on its own, so you can see — and stop at — any stage.

  1. 1

    control-test-designer

    Reads a bare control description and drafts test steps and what to check for. Nothing gets used until a person reviews and approves it — once approved, the test plan doesn't change mid-audit.

  2. 2

    evidence-examiner

    Pulls the real figures out of each piece of evidence (a PO, an approval email, an invoice) using structured extraction rather than eyeballing a scan — so a hard-to-read image can't be misread. Then checks it against the approved test plan.

  3. 3

    control-test-workpaper

    Rolls every result up into pass, fail, or exception, drops it all into the standard Excel workpaper, and won't finalize until every row links back to its source.

Building blocks it runs on

Shared, independently tested components. They ship bundled inside this skill — there is nothing extra to install.

What you'll need to change for your firm

This runs as published, but every audit function documents things its own way. These are the places it assumes a house format:

  • Your risk-and-control matrix column layout, so control descriptions are read from the right fields
  • Your workpaper template — header block, tick-mark legend, and sign-off line
  • Who is authorized to approve a test plan, and how that approval is recorded
  • Where sample evidence lives, and in what formats it arrives

Built for Claude

Install it into Claude Cowork and point it at your files.

  1. 1Download the skill bundle — everything it needs, including the shared building blocks, is inside.
  2. 2Add it to Claude Cowork as a plugin (dynamic-fieldwork-audit-testing).
  3. 3Point it at your own files and run it against one period you already know the answer to.
Request this skill

Free. We'll send the bundle and the setup notes.

Running on ChatGPT Enterprise or Copilot? We port these into your environment.

Want Control Testing Agent fitted to your methodology?

We adapt it to your control matrix, your workpaper template, and your thresholds — then hand it over so your team can run it themselves.