Your AI prompts can become exhibits. Write them like they will.

What the courts just signaled

Two things happened in early June 2026 that every litigator should file away. First, a court ordered the disclosure of an expert witness's AI prompts, as reported by Mayer Brown. Second, in a climate lawsuit, Shell sought the AI prompts used by the other side, a development Reuters covered on June 2, 2026. On June 8, Reuters Practical Law published an analysis of the discoverability and privilege-waiver risk that arises when AI tools become part of the legal work itself.

Read those together and the signal is clear. The prompts you type into an AI tool are now being treated as part of the record. They can be requested in discovery. They can be ordered produced. And the questions courts are starting to ask are not theoretical anymore. They are showing up in real cases with real parties fighting over exactly these documents.

This is not a tool review and it is not a reason to stop using AI. It is a litigation hygiene problem, and litigation hygiene problems have a known shape. You change how you work so the record you create is one you can live with.

Treat every prompt as a potential exhibit

Here is the single discipline that does most of the work. Before you type a prompt into any AI tool, ask one question. Would I be comfortable producing this in discovery and seeing it read aloud in a deposition.

If the answer is yes, type it. If the answer is no, rewrite it until the answer is yes, or do that part of the thinking somewhere a prompt log does not exist.

This is the same instinct good litigators already have about email. You learned years ago not to write things in an email that you would not want a jury to see, because email is discoverable and people forget that until it is too late. A prompt is email's younger, more talkative cousin. It captures not just your conclusion but your reasoning, your assumptions, the theory you tested and abandoned, and sometimes the weakness in your own case that you were stress testing. All of that is exactly what opposing counsel would love to read.

The practical version of the rule:

• Write prompts that state the task neutrally. "Summarize the deposition testimony on the maintenance schedule" is safer than a prompt that editorializes about how damaging that testimony is to your client.
• Do not narrate your strategy into the prompt. The prompt is not the place to think out loud about your theory of the case if that prompt may be produced.
• Do not paste in material you would not want attached to the prompt as context. The context travels with the prompt.
• Keep the speculative, candid strategy work in the channels that are most defensibly protected, and keep the AI prompts task-focused.

None of this slows you down once it becomes habit. It is a thirty second pause that can save you a very bad afternoon.

It also helps to remember what a prompt log actually contains. It is not a clean record of your conclusions. It is a running transcript of your process, including the questions you asked and then thought better of, the framings you tried, and the moments you pushed the model toward an answer you wanted. Read back in a hostile setting, that transcript can be made to look like coaching, like cherry picking, or like an admission about the weakness you were probing. None of that may be a fair reading, but you do not control the reading. You control what you put in the log. Write it so that even an uncharitable reading lands somewhere you can defend.

Where work-product protection holds, and where it can break

Lawyers reach for the work-product doctrine here, and often correctly. Work-product protection is real and broad, and a great deal of AI-assisted work will fall inside it. But the June 2026 cases are a warning that the protection is not automatic, and that bringing an AI tool into the work can create new exposure.

A few honest distinctions, stated carefully because this is the part where overconfidence hurts people.

• Protection can break when the AI output crosses into territory that is not pure attorney mental impression. Once a prompt and its output feed into something an expert relies on, you are closer to the world of expert disclosure than to core attorney work product. That is exactly the seam the expert-prompt disclosure order ran through.
• Privilege can be waived by how you use the tool. Sharing protected material in a way that breaks confidentiality, or building a workflow where the AI interaction sits outside the protected channel, can waive what you assumed was safe. The June 8 Reuters Practical Law analysis flagged privilege waiver as a live risk, not a remote one.
• The doctrine is still being worked out. Courts are reasoning through these questions in real time. That means the safe assumption today is the conservative one. Do not rely on a protection that has not yet been tested in your jurisdiction to shield a prompt you would regret producing.

The point is not that the doctrine is gone. It is that you should not treat it as a force field. Write as if the prompt might be produced anyway, and let the protection be your backstop rather than your only plan.

A short protocol you can adopt this week

You do not need a committee for this. You need a habit and a couple of agreements.

• Adopt the one rule firm-wide. Write prompts you would be comfortable producing in discovery. Put it on a card next to the workflow rules people already follow.
• Separate the channels. Keep candid strategy thinking out of AI prompt logs. Use AI for defined tasks: summarizing, organizing, drafting from facts you provide.
• Brief your experts. The expert-prompt disclosure order is the clearest warning of all. If an expert is using AI, their prompts may be ordered produced. Make sure they know that before they start, not after a motion to compel.
• Mind the upstream boundary too. What you put into the tool in the first place is a separate question from what happens to the prompt in discovery. For the rules on client confidentiality before anything is uploaded, see what is safe to put into AI. This briefing assumes that question is already settled and deals only with the prompt as a discoverable record.

The skill under the tool

The lesson here is older than AI. Every new tool that captures your reasoning creates a new discovery surface, and the lawyers who get hurt are the ones who treat the tool as private when it is not. Telephones, then email, then text messages, now prompts. The pattern repeats, and so does the defense. Assume the record can be read by the other side, and create a record you can stand behind.

The advantage was never the tool. It is the discipline around it. If you want the structured version for legal work, The Leveraged Attorney teaches how to use AI on real matters without creating exposure you will regret, and the two minute course quiz will point you to the right starting place. For the items that should never go into a tool at all, keep the never upload list close.