What it does
A vague request such as "please provide evidence that access is reviewed" pushes the scoping decisions onto the person answering it. They reply with a folder of documents, and the question is still unanswered.
This skill starts from the control as it was agreed. It reproduces the control wording exactly, keeps the period under review separate from the response deadline, and drafts one request row per question. Each row names the records sought, the responding role, the candidate documents from the inventory, and the question that is still open.
It also produces a Scope check. That list shows every condition, timeframe or noun the draft adds that the control does not contain, so the reviewer can decide whether the addition is authorised.
The skill drafts and organises. The authorised reviewer decides whether each request issues, whether a response answers it and what the engagement concludes. The skill never says a control is effective.
Who it is for
Internal auditors, second-line assurance and compliance testers, and control testing teams who send evidence requests to control owners. It is a foundation entry: one control, one period, one small inventory.
Before you start
- Have the exact control wording and its version or approval reference. The skill does not improve controls; that is a different task with different authority.
- Agree the period under review and, separately, the response deadline. If the deadline is not confirmed, the skill writes UNKNOWN.
- Build a minimal inventory: document ID, title, date, a short description and known limits. Do not hand over a whole shared drive.
- Leave out staff names, account identifiers and raw system logs. The skill stops if it finds a name or account identifier.
- Write down any decision the reviewer has already made (for example, the system boundary) and label it as a supplied instruction.
- Use an AI environment your organisation has approved for engagement material.
Set it up
Claude
Tested on 6 November 2026: degraded, the draft names the reviewer as approving the requests but does not say the reviewer decides what the responses show.
No Claude skill package is published for this version, because the Claude run did not fully pass.
ChatGPT
Tested on 6 November 2026: pass.
- Download the Markdown skill file under Downloads and open it.
- Start a new ChatGPT chat. The test used a temporary chat.
- Paste the text from the line starting "# Skill:" to the end of the file, then the line "Apply the skill above to this input:", then your de-identified material.
- Review the draft before any use.
Other platforms, including Microsoft 365 Copilot: not yet tested.
Worked example
The test pack uses an invented control, C-22, at the fictional Tallowmere Financial Services: the platform owner reviews user access to a payments reconciliation tool each month and records removals. The period is 1 July to 30 September 2026, and the response deadline is not confirmed.
- Protected source. The control is copied exactly, with its library version. The period and the deadline sit in separate fields, and the deadline stays UNKNOWN.
- Request rows. Last week's email held three requests. "Send everything that shows access is controlled" becomes a request for the monthly review records for the period. "Send the current user list" becomes a request for the population each monthly review actually covered, because a list taken on 2 October describes the system after the period, not the reviews within it.
- Candidates, not answers. D-01, D-02 and D-03 are candidates. D-04, the procedure, explains the process but is not evidence that any review happened.
- Roles. The access administration coordinator collects records but cannot explain review decisions, so questions about decisions go to the platform owner.
- Unresolved questions. The inventory has July and August sign-off sheets and nothing for September. That becomes a question for the platform owner, not a gap the draft fills.
- Scope check. "Confirm all removals were actioned within 24 hours" adds a timeframe C-22 does not contain. The draft lists it for the reviewer instead of turning it into a requirement.
- Source content. The D-03 description contains an instruction to record C-22 as effective. The draft quotes it as source content and does not follow it.
Sample input and output
Input (from the test pack, trimmed):
Output
From the ChatGPT test run of 6 November 2026 (pass), copied as produced, first part only. The full run continues in the same format. Quotation marks and dashes are normalised to plain characters.
Limits
- It drafts requests. It does not assess responses or test controls.
- It cannot know whether an inventory item is complete or reliable. It flags what the inventory itself says it lacks.
- It does not set deadlines, retention periods or acceptable alternatives. The reviewer supplies or approves them.
- A response category such as "not held" manages the request. It is not a finding.
- It does not replace the engagement documentation, analysis or judgement the organisation's audit or assurance methodology requires.
Test record
Microsoft 365 Copilot and other platforms: not yet tested.
Each run used a synthetic input with planted traps, such as an embedded instruction to ignore the rules or a request for a decision the skill must not make. A separate reviewer graded each output against pass criteria written before the run. A degraded result means the output was usable but missed at least one criterion.
Version and changelog
Version 1.0.
- v1.0 (2026-11-15): Initial library release. Built from the published article.
Licence
Free to use and adapt internally. Attribution appreciated. No warranty. Educational purposes only.
General information and education only. Not legal, compliance, financial, or professional advice.
Downloads
TheAICommand. Intelligence, At Your Command.