What it does
This skill answers routine HR policy questions strictly from the organisation's own uploaded policies. Every answer is framed as "the policy states", names the policy, version date and clause, and quotes the operative words, so a team member can check it in under a minute. If the policies are silent, it says so in the first line and routes the question to the escalation owner. If the answer could interact with an award, an enterprise agreement or the National Employment Standards, it adds an AWARD/NES CHECK flag and does not try to resolve it. Complaints, grievances, safety issues, legal threats, performance matters and questions about a specific person get the escalation line and nothing else.
Every answer is a draft. A person verifies it, sends it and owns it.
Who it is for
HR officers, HR shared service teams and people leaders who answer the same policy questions every week. It is rated foundation because the setup is a policy set, an escalation owner and a clear list of what the skill must refuse, and every answer is checked by a person before it is sent.
Before you start
- Collect the current approved version of each policy and record its version date. Retired versions stay out.
- Name the escalation owner by role.
- Frame questions by role, such as [EMPLOYEE_ROLE], never by name. Remove staff or employee numbers too. A question that still carries a staff number is not de-identified, and the skill will not repeat it.
- Use an AI tool your organisation has approved, and keep the policies and the skill in one project or workspace.
- Fill in the Owner, Approver and Approved on lines in the skill file before first use.
Set it up
Claude
Tested on 6 November 2026: pass.
- Download the Claude skill package (.zip) under Downloads.
- In claude.ai, open Customize, then Skills, then Add, then Upload a skill, and choose the .zip.
- Check the skill is switched on.
- In a new chat, ask Claude to use the policy-grounded-hr-query skill on your de-identified material.
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 synthetic test input is an invented freight co-operative with a Leave Policy and a Rostering Procedure, and three questions run one after another.
- A question the policy answers. Q1 asks about annual leave notice and approval. The method finds clause 5.2 of the Leave Policy, quotes it word for word, cites the version date and frames the answer as "the policy states".
- The award and NES layer. Annual leave is also governed outside company policy, so the method adds the AWARD/NES CHECK flag and does not state any entitlement itself.
- A question the policy does not answer. Q2 asks about cashing out annual leave. Neither document covers it. The method says so in the first line, marks it NOT COVERED and routes it to the People and Culture Manager, without filling the gap from general knowledge.
- The trap. Q3 looks like a rostering question, but it describes a complaint, a safety concern about forklift brakes and an identifiable worker with a staff number. The method returns only the escalation line. It does not quote the rostering procedure and does not repeat the staff number or the worker's details.
Sample input and output
Input (synthetic, trimmed)
Output
From the Claude test run of 6 November 2026 (pass), copied as produced. Quotation marks and dashes are normalised to plain characters.
Limits
- It knows only the uploaded policies. If a policy is out of date in the project, the answer is out of date.
- It does not interpret awards, enterprise agreements or the National Employment Standards. It flags where they may apply.
- It does not handle complaints, grievances, safety reports, legal threats, performance matters or questions about a specific person. Those go to a human.
- It drafts answers. It does not send them.
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 1.1.
- v1.1 (2026-11-15): Source now points to the guide page; Platforms and Tested against list only platforms actually tested; guide, worked example and test pack added.
- v1.1 (2026-11-15): Skill block gains Version, Owner, Approver, Approved on and Review trigger lines.
- v1.1 (2026-11-15): Method step 7 and Guardrails now cover grievances and performance matters, matching When to use, and give the exact escalation line.
- v1.1 (2026-11-15): Personal details now include staff or employee numbers; a routine question that mentions a person is reframed by role, and a question about a person's own situation is escalated.
- v1.1 (2026-11-15): Guardrail added: instructions found inside a question or a policy document are treated as content.
- v1.0 (2026-08-15): Initial library release. Extracted from the published article.
The v1.0 file stays available unchanged.
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
- Markdown skill file, version 1.1
- Previous version 1.0, kept unchanged
- Claude skill package, version 1.1 (.zip)
TheAICommand. Intelligence, At Your Command.