About 25 minutes to set up · claude.ai only

Clean up billing narratives before the invoice goes out

Today

Prebills come back full of entries like "review file" and "call re: case." The billing partner rewrites them at the end of the month, or the client writes them down.

With Claude

Each raw entry comes back as a narrative that says what was done and what it moved forward, with block billing, vague wording, and possible clerical work flagged for the timekeeper. Your time values are copied through untouched.

The prompt

Paste this into claude.ai and replace anything in brackets.

You are helping a law firm tighten up billing narratives before a prebill goes to the client.

Absolute rules:
- Never change, round, or add to the time value on any entry. Copy each one exactly as given.
- Never describe work that is not already in the raw entry. If an entry is too vague to rewrite honestly, return it unchanged and flag it.
- Do not merge entries and do not split them.

For each entry, output five fields: date, timekeeper, unchanged time value, rewritten narrative, flag.

Rewrite each narrative so it names the specific task, the document or person involved, and the purpose it served in the matter. Past tense, no filler, no adjectives.

Add a flag when an entry looks like block billing, when it is too vague for a client to approve, when it describes work that may be clerical and non-billable, or when it conflicts with the guidelines below.

Client billing guidelines to follow:
[paste the outside counsel guidelines here, or write "standard"]

Raw time entries:
[paste the raw time entries here]

What to skip in law firms

  • Client confidential material in a personal consumer login. Privilege and Rule 1.6 do not bend because a tool is convenient. Anthropic's consumer plans (Free, Pro, Max) use your chats to improve their models only when the model improvement setting is turned on in Privacy Settings, and commercial plans such as Team, Enterprise, and Claude for Work exclude inputs from training by default. ABA Formal Opinion 512 puts the work on you: read the terms yourself, judge the risk of disclosure, and get informed client consent where it is required, and it says boilerplate engagement letter language does not count as informed consent. Settle the account, the setting, and the consent question before a real matter goes in, and practice on redacted or hypothetical facts until you have.
  • Legal authority and deadline math. A language model can produce case names, docket numbers, quotes, and holdings that read perfectly and do not exist. In Mata v. Avianca the Southern District of New York sanctioned the lawyers under Rule 11 after a brief cited cases that were never real. Treat every authority Claude gives you as unverified until you have pulled it in your own research service and read it. Same rule for dates: never let Claude calculate a statute of limitations, a response date, or a filing deadline. Have it quote the dates in your material, then calendar the deadline yourself.
  • Anything that leaves the firm without an attorney reading it. Claude does not practice law and does not give legal advice, and none of this changes who signs the document. No pleading, motion, demand, opinion, or client letter should go out because Claude drafted it and it looked fine. Do not put Claude in front of clients or the public to answer legal questions, and remember your state's advertising rules cover whatever you say about AI on your own website.

Want all 4 workflows for law firms in your inbox?

Every prompt on this page plus the rest, so you still have them on Monday. One email, then four short ones on making it stick. Unsubscribe any time.

Or answer four questions and Claude writes three workflows for your specific business rather than the industry generally.

Before you try it

More for law firms

Full guide for law firms