Record it once, run it every month

You've explained the monthly variance process four times this year. Once to a new analyst. Once to Claude, in a chat, from scratch. Twice to yourself, reconstructing what you did last quarter because nobody wrote it down. Record a skill exists so you do that explaining once, on camera, and then never again.

What you're recording

Record a skill is a Cowork feature: you turn on screen recording, perform a workflow the way you normally would, narrate your reasoning as you go, and Claude turns the whole demonstration, the clicks, the typing, the spoken logic, into a reusable Skill. You can review it, save it, and invoke it on a later cycle.

One distinction matters here: this isn't a macro. A macro replays clicks at fixed coordinates and breaks the moment a menu moves. What gets generated instead is structured guidance, closer to a written SOP than a recording, that Claude follows and adapts on each run.

For a strategic-finance professional, that's the useful way to think about it. You're not capturing "summarize this file." You're capturing "take this month's export, update the standard workbook, run the variance checks, and write the commentary the way we always write it."

Where it fits in finance work

A Skill earns its keep on work that's repeatable, rules-driven, and runs through a browser or desktop rather than a clean API. Six workflows show up often enough to name directly.

The real value sits underneath all six rows. It kills the re-explain-the-process-every-month tax. A well-built Skill encodes your source order, your definitions, your materiality thresholds, your naming conventions, your exception handling. Not a generic version of the task. Yours.

Example: a monthly variance pack

Say you run a standard budget-versus-actuals review across cost centers or revenue lines, a review nearly every FP&A desk owns in some form. You'd record yourself downloading the latest actuals export, opening the standard reporting workbook, and updating the source tabs. Then reviewing variances against budget and against prior month. Then, out loud, the part that matters: "Flag revenue or spend variances above 3% or $250K. Split volume-driven effects from rate or price-driven effects. Don't call something material if it's just a known timing shift." Then updating a commentary template with the same shape every time: performance, drivers, risks, follow-ups.

That threshold is the whole point of narrating instead of just clicking. A screenshot shows what you did. Only your voice explains why a $180K miss gets waved through and a $260K one gets flagged, and that reasoning is exactly what the Skill carries into next month, when you're not in the room.

How to record one

The feature lives in Cowork in Claude for Mac, on Pro, Max, and Team plans, with code execution turned on. Not in regular chat, not on Windows, not on Free or Enterprise. macOS will ask you to grant Screen Recording and Accessibility permissions before it works.

The mechanics are short. Open Cowork, select "+ → Record a skill," or go to Customize → Skills → Add → Record your screen. Start recording and do the task the way you normally do it. Narrate the logic behind your choices as you go, especially thresholds, definitions, checks, and what you do when something looks wrong. Select Done, and Claude proposes a Skill for you to inspect before it saves, or you can discard the take instead.

One step nobody skips twice: test the saved Skill against a non-sensitive, representative historical period before you point it at a live reporting cycle. The first run tells you what your narration captured, not what you meant to say.

How to make it useful

Recording quality is the whole game here. A rushed take produces a Skill that's confidently wrong in the same way every month.

Start narrow: a task that happens at least monthly, runs 20 to 60 minutes, and has a defined "done" state. Use a sanitized or prior-period data set for the first pass, especially if your systems touch customer, compensation, or non-public financial data. Narrate decision rules, not just clicks, meaning materiality thresholds, metric definitions, known exceptions, and your source-of-truth hierarchy. State your validation gates out loud: tie total revenue to the GL control total, don't overwrite formula cells, stop and flag anything above your threshold instead of guessing.

The line that matters most: separate the mechanical work from the judgment. Let Claude gather, format, reconcile, and draft.

Keep forecasts, accounting conclusions, external communications, and executive recommendations as yours. Treat the saved Skill the way you'd treat any other finance artifact with your name on it: name it clearly, keep its templates and reference files attached, document who owns it, and re-test it after the source system or the model changes underneath it.

A strong first Skill

Start with a monthly budget-vs-actuals variance pack, not your full close or your full forecast. It has repeatable inputs and outputs, the time saved is visible the first time you run it, and it's easy to validate against something you already know.

A narration script that does the job:

Update the monthly variance file using the approved exports. Reconcile total revenue and total spend to the control totals. Flag variances greater than 3% and $250K, explain whether each is volume, rate, mix, or timing, and produce a one-page management commentary. Do not change assumptions, formula logic, or final numbers without flagging them for review

Update the monthly variance file using the approved exports. Reconcile total revenue and total spend to the control totals. Flag variances greater than 3% and $250K, explain whether each is volume, rate, mix, or timing, and produce a one-page management commentary. Do not change assumptions, formula logic, or final numbers without flagging them for review

Update the monthly variance file using the approved exports. Reconcile total revenue and total spend to the control totals. Flag variances greater than 3% and $250K, explain whether each is volume, rate, mix, or timing, and produce a one-page management commentary. Do not change assumptions, formula logic, or final numbers without flagging them for review

Claude Skills already work across Excel, PowerPoint, and PDF, which covers most of what a recurring reporting package touches. Record the process once, and the real payoff shows up later: the third analyst who inherits this desk never has to ask how it's done.