Write Your AI Assistant Once, Then Port It to ChatGPT, Claude or Gemini

Write the assistant’s spec once, in plain fields: role, task, inputs, rules, refusals, format, tone, examples, checks and owner. Then paste it into a custom GPT, a Claude Project or a Gemini skill. Because tools change their features and their names, a portable spec is the asset. The tool is only where you store it.

I am Mac Malbrough. I hold a J.D. and an MS-HRM, I am not an attorney, and I help small HR teams use AI without creating new risk. This post is about structure, so it does not depend on any one product.

What problem does a reusable assistant solve?

Many HR professionals have a favorite prompt for each recurring task: a job post rewrite, a meeting recap, a policy summary. The prompts live in a notes file, a chat history or someone’s memory. Each time, the person pastes it, adjusts it and hopes for the same quality. The results vary, a coworker cannot reuse it and a good prompt dies when the chat is deleted.

A reusable assistant is the fix. It is a saved set of standing instructions, and often reference files, so that the same task gets the same treatment every time. The main tools offer this under different names: custom GPTs in ChatGPT, Projects in Claude and Gems in Gemini. All of them let you save instructions and reference material so you do not have to retype them. The risk is that you write the instructions inside one tool’s box and cannot take them with you.

Why does portability matter now?

Features change. Google has announced that Gems will become read-only and transition to skills on November 17, 2026 for personal accounts, and that they will be removed from Gemini Apps on December 17, 2026. Workspace business, enterprise and nonprofit accounts follow in March 2027, and education in June 2027. Skills are reusable custom instructions that you invoke with a slash command, with up to 100 active skills and a 100 MB file limit.

That is one example of a general pattern. If you built your assistant inside a feature that gets renamed, reshaped or retired, you have to rebuild it. If you built it as a spec in a document, you paste it into the new feature and keep going. I would not predict which tool will lead a year from now, and you do not need to. You need the spec.

What goes in the spec?

A good spec is short and specific. The Spec Sheet for a Custom GPT, Claude Project or Gemini Gem has 12 fields, a filled example and a paste-ready template. In outline, the fields cover:

• Role and audience: who the assistant is helping and what they need
• Task and scope: what it does and what it does not do
• Inputs: what the user will provide, and what to ask for if it is missing
• Rules: the standards that apply, such as your handbook or style guide
• Refusals: what it must decline, such as legal conclusions or decisions about a person
• Output format and tone: headings, length and voice
• Examples and checks: one good output and a list of things to verify

The refusal field is the one most people skip, and it matters most for HR. An assistant that drafts a termination letter should not decide who is terminated. My article on five reusable AI skills for HR with built-in refusal rules shows how to write them, and where AI helps HR and where a person must decide explains the line.

Which task should you build first?

Pick a task that repeats, has a clear standard and carries low risk. Drafting a first version of a job description is a good example. Summarizing a policy for a manager is another. The HR Task AI Readiness Scorer scores 30 HR tasks on four factors, and tells you which to assist, which to pilot and which to keep human, so you can find your top three in about 15 minutes.

Avoid starting with anything that evaluates individual people, such as screening applicants or rating performance. Those tasks carry legal risk and need careful review. If you are in California and considering tools that screen candidates, read the civil rights rules for AI hiring tools first.

How do you test the assistant?

Write five test inputs before you rely on it: two ordinary cases, one with missing information, one that should trigger a refusal and one that tries to push the assistant past its rules. Run all five and compare the outputs to your spec. If the assistant fails a test, change the spec, not just the one conversation, and run the five again. Keep the test set with the spec so that when you move to a new tool you can rerun it and check that behavior carried over.

Then add a human check. Someone qualified reviews each output before it is used, and the review is recorded. The AI Decision Log for HR is a simple Excel log of what the tool did, what a person checked and who decided. It takes a minute per entry and gives you a defensible record.

What about confidential data?

The spec should say what data may be entered. A safe default is to use no names, no medical or immigration information and no case details in a tool that your company has not approved. Check the data settings of each tool, including whether your inputs may be used for training, and put the answer in your written AI use rules. My safe AI prompt library for HR tasks gives prompts that avoid personal data, and the California employee handbook template is where an AI use policy belongs once you have written one.

For a deeper treatment of the whole approach, my book The HR AI Field Guide is available in paperback, with a Kindle edition.

What should you do this week?

Choose one repeating HR task and fill in the spec fields for it, in a document, before opening any AI tool. Paste it into whichever tool you already use, run your five tests and note the results. If you use Gemini and have Gems, copy their instructions into your spec before November 17 so nothing is lost in the transition.

The Spec Sheet and the Readiness Scorer are the fastest start. If you want help deciding which tasks and tools are right for your team, book a free consult through my HR services for small businesses.

Related reading: an AI starter plan for small employers.

Sources

• Google Gemini Apps Help: Gems transition to skills: dates, skills limits and account types
• Five reusable AI skills for HR with built-in refusal rules
• Small California employers and AI hiring tools: civil rights rules

McKinley holds a J.D. but is not a licensed attorney. Articles here are general information, not legal advice. For your specific situation, talk with an employment attorney.