Hand off the work

Run a firm skill

4 min read

In the last lesson you handed Cowork a one-off task. A skill is the same handoff, except the task is one the firm has already built and approved. Skills are pre-built workflows for jobs we do over and over, set up once so nobody has to explain the steps each time. You can run them wherever you use Claude, and here you will run one in Cowork. You do not configure anything. You point Cowork at the work, ask for the thing by name, and the skill runs.

The one you will run here splits K-1s. When a partnership, S corporation, or trust return is finished, each partner, shareholder, or beneficiary needs their own K-1 package for their personal return, and pulling each one out of the full return by hand is slow and easy to get wrong. The skill does it: point Cowork at the folder that contains the return and it copies each recipient's complete K-1 package, the form page plus that person's supplemental statements, into its own PDF, named by year, entity, and recipient. It works on our Drake returns and on returns other firms prepared, and it never changes the original. It only writes copies.

Here is a run.

You

I have the finished 1065 for a four-partner return in this folder. Split out the K-1s, one PDF per partner.

Claude

I detected a 1065 and split it into four partner packages, copied from pages 18 through 41 of the source. Each file is named with the year, the entity, and the partner. Two flags worth your eye: one partner's name was read off the bare K-1 form rather than a cover letter, so double-check that filename, and the last partner's package is my best guess at where their pages end, so confirm it did not run long. The original return is unchanged.

Look at how that result differs from the summary in the last lesson. It does not just say done. It tells you the form type it detected, how many packages it made, the page range it copied from, and then a short list of flags. Those flags are the most important part of the message. The skill is telling you exactly where it was less sure: a name it had to read off the form itself, and the last recipient's package, where it has to judge where one person's pages stop and the return's own pages begin. It is not hiding the soft spots. It is handing them to you to check.

So you check them. An agent just did several steps of work without you watching, which is precisely when your judgment matters most. Open every package it produced and run down four things.

  1. 1Count

    One file per recipient, matching what you expected.

  2. 2Assignment

    Each file holds the right person's K-1.

  3. 3Completeness

    Each package has the full K-1 and its statements.

  4. 4Flags

    Every flagged name, tail, or field is resolved.

None of this takes long, and the safety net is real: because the skill only ever copies, the original return is always there to fall back on if a package came out wrong. But a misassigned or short K-1 that goes out to a client is a real problem, and the only thing standing between the agent's draft and that mistake is the few minutes you spend looking. The skill made the packages. Whether they are right is still your call.

Your turn

Try it on a sample return

Find a sample or practice return that has K-1s in it, point Cowork at it, and ask it to split out the K-1s. When it finishes, open every package and run the four checks above before you trust a single one.

Cowork runs in the desktop app, so do this one there.

That is the whole path. You started not sure how to open Claude, and you can now hand an entire task to an agent, let it work on its own, and decide whether what it gave you is right. Everything in between, the prompting, the use cases, the verifying, the workflows, was building toward that one move: handing off the work while keeping the judgment.

That last part does not change, and it is worth saying plainly at the end. The tools will keep getting faster and more capable. What clients pay BC&V for was never the document. It is the judgment that the document is right, and that judgment stays with you no matter how good the tool gets.

You are now AI fluent. Not because you can run the exact steps this training showed you. Those steps will change as the tools do. You are fluent because you have what lasts: you know how to ask for what you want, how to check what comes back, and when to keep the judgment for yourself. When the next tool arrives, and it will, you will not be starting over. You will pick it up the way you picked up this one, already knowing which part to hand off and which part is yours to keep.