I’ve been living in Claude for Excel the last three weeks, and I can truly say that this is what the cavemen must’ve felt when they discovered fire.

But before you can put rock to rock and create that initial spark, you must have the proper setup.

This guide walks you through installation, the parts of setup that aren’t obvious from the docs, and the questions worth having answered before you start relying on it for real modeling work.

What you need.

1. A paid Claude plan — Pro, Max, Team, or Enterprise. The free tier doesn’t get the add-in (if you’re a reader and you’re on a free plan, kindly, what are you doing?)

2. Excel for web, Windows, or Mac. No dependency on Claude Desktop — this installs and runs on its own.

3. If you’re on a Team or Enterprise seat (through an employer), expect the add-in to show up within minutes, but budget up to 24 hours for full org-wide rollout before assuming something’s broken.

The install.

Five steps, and the only one that trips people up is the first — the add-in isn’t where you’d guess.

1. Open the Home tab ribbon

2. Click Add-ins on the far right (sometimes labeled Get Add-ins, red windows icon).

3. Search “Claude by Anthropic for Excel,” click Get it now, then Add.

4. Sign in on the web. You’ll get redirected out of Excel to sign in or create a Claude account.

5. Pair the code. Excel’s sidebar shows a pairing code — copy it into the browser where you’re signed in, then approve the permissions prompt. That’s the last gate. Once you clear it, the sidebar is live.

Two systems that look the same and aren’t.

This is the part worth slowing down for, because getting it wrong doesn’t throw an error — it just means your rules aren’t being applied.

Claude for Excel has two separate places to store standing preferences that (as of August 2026) don’t talk to each other.

Instructions live in the Excel sidebar (Settings, the ⋮ menu). They’re a single block of always-on text, applied automatically to every conversation in that app. This is where a short, durable rule belongs — the kind of thing you’d never want to restate.

Skills are structured folders — a SKILL.md plus optional scripts and references — encoding a full multi-step workflow. They live in your account’s skill library, sync everywhere Claude runs, and apply either automatically when relevant or on demand via a slash command. This is the format for something like a full monthly-close package: structure, QA checklist, worked example.

Here’s the distinction that matters: Instructions are app-specific, Skills are global. Excel Instructions and PowerPoint Instructions are two different fields — set one, and it says nothing to the other. Skills, once uploaded, are the same everywhere.

Think of instructions as those rules you give to your analysts on Day 1.

Never hardcode assumptions or coefficients as static numbers — always link them to a driver cell or formula so they update dynamically.

Format all Excel models using IB conventions — blue font for inputs, black font for formulas, green font for cross-sheet links.

If this wasn’t intricate enough, the Excel Instructions field and the web Profile Preferences field (claude.ai → your initials → Settings → Profile → “What preferences should Claude consider in responses?”) are not shared. If you split work across web for writing and research and Excel for modeling — which most Strategic Finance people will — put your core standing rules in both places. Setting a rule once and assuming it travels is the mistake (and one of the core reasons I decided to write this guide).

One more wrinkle (I promise this is the last one): Skills you’ve already built for other work — a “Your Voice” file, say — apply on web and chat but won’t surface inside Excel unless they’re Excel-relevant. If you want a finance-specific skill available in the sidebar, upload it the same way you did your other skills; it doesn’t show up there by default. You can also build one from inside Excel directly — type /skill-creator in the sidebar, describe the workflow in plain English, and Claude drafts the SKILL.md for you. Worth doing for anything recurring, like building a teaser or a model-health check.

Cross-app context: what actually carries over.

If you work across Excel, PowerPoint, Word, and Outlook in the same sitting, you can sync updates between one app to the other — with a caveat.

Files open at the same time — this part is solid. Live sharing works. Claude can read from a workbook open in Excel and write straight into a slide open in PowerPoint, no re-explaining the structure in between. However, Claude can’t open or switch files for you. Both need to already be open.

Files closed, reopened later — this part isn’t settled. From my experience, if you close a file (like an excel model) and re-open it later, you’ll need to re-point to that file (ex: “reference the model I just saved as Q3_Close.xlsx”) rather than assume Claude remembers.

The scenario that matters more in practice: a full day of context-building, then picking the file back up four days later. The likely answer is you’ll need to re-establish interpretation and context. Here’s the mechanic:

• What Claude reads live from the open file — tabs, formulas, values — reloads automatically every time you reopen it.

• What you explained in conversation — which cells are drivers versus overrides, why an assumption sits where it does — is session-bound, and probably doesn’t survive four days. That said, Claude does have a chat history feature (found at the top of your chat windows) where you can start again from an existing chat.

The fail-proof fix maps onto something you’re likely already doing: building a “Read Me” tab inside the workbook itself — input definitions, driver map, tab dependencies, override log, a “don’t touch these” note. Since Claude re-reads the open file every session, that tab loads for free, every time, with nothing re-explained. Pair it with the account-level Instructions for the rules that are always true (driver-based formulas, IB formatting), and the two together cover what conversational memory fails to hold.

You can automate this entire process by asking Claude to load all relevant context, edits, and other nuances throughout your chat history into the Read Me tab. I suggest prompting Claude to refresh the Read Me tab before you close out each session.

Here’s an example prompt

Create a new worksheet in this Excel workbook called README.

The purpose of this tab is to preserve the important context, decisions, edits, assumptions, and nuances from our work in this session so that another analyst or a future AI session can quickly understand the workbook and continue working without losing context.

Review:

1. Our entire conversation from this session.

2. Every change you made to the workbook.

3. The workbook's current structure, formulas, assumptions, and outputs.

4. Any issues we identified, investigated, or resolved.

5. Any decisions I made or guidance I gave you during the session.

Then create a concise but comprehensive README containing the above.

Connectors: installed but not connected.

Claude for Excel supports outside-data connectors — S&P Capital IQ, FactSet, Moody’s, PitchBook, Daloopa, Dualipa (kidding), Aiera, plus custom ones like an internal MCP server. This does not mean that you automatically get access to FactSet (or Dualipa) — the app still needs authentication (an associated account) before Claude can actually pull data through it.

To see what’s installed: Claude Excel sidebar → + icon → connectors

To connect one:

1. claude.ai → Settings → Connectors

2. Find it, click Connect

3. Complete the OAuth-style sign-in to the provider

Once connected, it works automatically in Excel — nothing further to configure on the Excel side. All connector management happens centrally on claude.ai; the sidebar is a viewer, not a configuration surface. Same pattern as Skills.

One thing worth a second look before flipping it on: a custom connector — a self-configured MCP server, say — can introduce real risk. Worth knowing what it actually does and where it sends data, especially given what these workbooks tend to hold.

So if you work in healthcare as I do, the approval process can be ~lengthy~.

Skills are not easily acquired.

Ok, listen up! This is a nuance worth paying attention to, especially for us finance nerds.

Web Skills (claude.ai → Settings → Skills) shows only a handful of skills plus anything you’ve personally uploaded. It was never built to surface plugin-bundled skills individually (ex: /dcf-model)

Web Plugins (claude.ai → Settings → Plugins) shows only a few plugins, but each plugin does contain several skills. For instance, the Finance plugin contains 8 bundled skills.

Third-party marketplaces and Anthropic’s GitHub page contain plugins you can download and upload to your Claude workspace. There are thousands of plugins available, but not all can be trusted.

Claude for Excel’s pre-loaded skills are a third path entirely, baked into the add-in and independent of both catalogs above — things like /audit-xls, /clean-data-xls, /dcf-model, /comps-analysis. They’re likely sourced from the same finance-plugin ecosystem (the /dcf naming lines up with the marketplace’s /dcf command), but they ship inside Excel by default rather than through either catalog.

Net effect: if you want your own custom skills available inside Excel, you upload them the same way you did your voice files. They don’t show up there on their own.

Common errors, and how to fix them.

A rule you set on the web isn’t showing up in Excel. Instructions and Profile Preferences are separate fields. Set the rule in both places, or accept it only applies where you put it.

A skill you built for writing doesn’t trigger in Excel. Skills sync globally, but relevance still gates whether one fires — an Excel-irrelevant skill (ex: your voice or writing style) won’t surface in a spreadsheet context even though it’s technically available “everywhere.”

Like an MBA Associate, Claude doesn’t seem to remember what you discussed about a model from a few days ago. Expected, per the cross-app section above — conversational memory is the part that doesn’t reliably persist. Re-point Claude to the file explicitly, and lean on a Read Me tab in the workbook for anything that needs to survive the gap.

You went looking for “Audit Support” in the web Skills catalog and it’s not there. It’s not a standalone skill — it’s bundled inside the Finance plugin. Install the plugin, not a skill.

What this buys you, and what it doesn’t.

The whole installation from start to finish takes about 10 minutes once you know the ribbon location. What takes longer is deciding what goes in Instructions versus what belongs in a Skill, and getting your standing rules set in both the places they need to live. None of this replaces documenting the model itself. A live workbook with no Read Me tab still hands Claude a black box every session, connector or no connector, skill or no skill. The setup gets Claude into the room. What it does once it’s there still depends on how well the workbook explains itself — and ultimately how well you know the workbook.