The AI Chief of Staff Workflow: Earmark + Claude

A paste-ready guide to combining Earmark and Claude into the workflow most teams are missing - a per-account memory layer that gets sharper with every call.

What most "AI chief of staff" tools get wrong

Most products marketed as an AI chief of staff do one of two things. They transcribe your meetings into searchable notes, or they generate a fresh summary at the end of each call. Both are useful. Neither is a chief of staff.

A chief of staff carries state. They walk into the Tuesday call already knowing what was committed on the Friday call. They remember that the VP of Sales is leaving in six weeks, that the renewal is up in Q3, and that the customer is sensitive about the word "scope." The job isn't summarization. It's continuity.

This guide shows how to build that continuity with two tools - Earmark for capture, Claude for reasoning - and one living document per account we'll call the Account Brain. You can stand it up in an afternoon.

Worth flagging up front: Earmark's own docs acknowledge that deal-level synthesis across many calls isn't a one-click action inside Earmark today (see the "What this workflow doesn't do" section of the sales calls workflow guide). This post is the practical workaround they hint at, written out end-to-end.

The division of labor: Earmark captures, Claude reasons, you approve

Each tool does one job:

• Earmark sits in the meeting. It transcribes live (no bot joins the call), generates structured artifacts during the conversation, and - on the desktop app - auto-saves a markdown transcript to your computer the moment the call ends. It's the capture layer.

• Claude holds the per-account memory. It reads each new transcript, compares it against the existing Account Brain, and proposes an updated version. It's the reasoning layer.

• You review the proposed update, fix anything wrong, and commit. You're the ship layer.

The third role is the one most "AI chief of staff" tools quietly skip. Earmark's operating principle is that nothing auto-sends - the human is always the final author. The Account Brain inherits the same rule. Claude proposes. You ship.

The feature that makes this easy: local transcripts

The whole workflow hinges on one Earmark feature you may not have turned on yet: auto-saved local transcripts.

On the Earmark desktop app (Mac or Windows, v1.4.1+), every meeting transcript is written to your computer as a markdown file the moment the call ends. The default save location is:

~/Documents/Earmark/Meeting Transcripts

Each file has YAML frontmatter (title, date, participants, workspace) followed by a ## Transcript section. You can change the folder under Settings → Personalization → Meeting files - the only constraint is that it has to live inside your home directory (~).

That means there's no "export" step. Earmark already drops a clean, structured markdown file on your disk after every call. The setup below picks up from there.

If you're on the web app, the file doesn't land on disk - you'll copy the transcript out of Earmark manually after each call. The shape of the workflow doesn't change; only the friction does.

One-time setup: pick how Claude reads the transcripts

There are two flavors of this setup depending on which Claude surface you're on. Pick the one that matches.

Path A (recommended) - Claude Cowork or Claude Code

Cowork and Claude Code read directly from a folder on your computer, which means no upload step ever. The Earmark docs spell this out - point Cowork, Claude Code, Codex, or Cursor at the transcripts folder and it just works.

1. Point Cowork (or Claude Code) at ~/Documents/Earmark/Meeting Transcripts/. Earmark is already writing to that folder after every call.

2. Create one account-brain.md per account in the same workspace. An accounts/ folder beside the transcripts works fine - e.g. ~/Documents/Earmark/accounts/acme-health/account-brain.md.

3. Paste the prompt further down into wherever the surface keeps durable instructions - a Cowork session prompt, or a Claude Code CLAUDE.md in the workspace root.

How you scope each call to the right account is a small choice:

• Filter at read time. Leave all transcripts in the default folder and let Claude pick the right ones using the participants field in the frontmatter. Simplest setup; fine until you have a few hundred transcripts.

• Per-account subfolders. Move each new transcript into accounts/{name}/calls/ after the call ends. One extra drag per call, but Claude only ever reads the relevant transcripts and the brain stays cleanly co-located with its calls. (You can't have Earmark write directly into per-account folders today, since each meeting only knows its own metadata, not which account it belongs to.)

Path B - Claude Projects on claude.ai

Projects don't sync with a local folder - files have to be uploaded. So you can't point Earmark at a Project, and you'll keep a small upload step in the loop.

1. Create one Project per account.

2. Add a file called account-brain.md. Leave it empty for now, or seed it with the seven-field structure below.

3. After each call, upload the latest transcript from ~/Documents/Earmark/Meeting Transcripts/ into the Project.

4. Paste the prompt further down into the Project's custom instructions.

Either path gives you the same outcome: Claude has the new transcript, the current brain, and a contract for how to update it.

The five-step loop, after every meeting

1. Take the call in Earmark. Earmark transcribes live. If you pre-seed a task like the Client Call template, you'll also walk out with a structured artifact - decisions, action items, owners - without writing a prompt. (For sales calls specifically, the MEDDPICC-shaped template in Earmark's sales calls workflow guide is more thorough.)

2. Make sure Claude can see the transcript. Path A (Cowork or Claude Code): nothing to do - the file is already in the folder Claude is reading, or in the right per-account subfolder if you organized that way. Path B (Projects): upload the latest transcript from ~/Documents/Earmark/Meeting Transcripts/ into the Project.

3. Ask Claude to update the brain. One message: "Update the Account Brain for {{account}} using the transcript from {{date}}. Apply the rules in the instructions."

4. Review the proposed update. Claude returns the full updated brain in the same seven-field structure, with each change tagged. Read it against the current account-brain.md. Edit anything that's wrong, exaggerated, or missing nuance.

5. Commit the new brain. Replace the contents of account-brain.md with the approved version. Done. The next meeting reads from it.

Three to five minutes per call once it's muscle memory. The work compounds in the brain instead of dissipating in your inbox.

The seven fields the Account Brain has to carry forward

The brain isn't a summary. It's a fixed-structure snapshot, and the structure is what stops Claude from regenerating it from scratch each time.

1. Stage of relationship. Onboarding, scaling, steady, at-risk, winding down. One word. Changed only when the call gives explicit evidence to change it.

2. Current spend tier or scope. What the account is paying for or committed to. Carried across calls unless this call agreed to change it.

3. Health, with a one-sentence justification. Green, yellow, or red, plus one sentence on why. The justification is what stops "red" from being meaningless three weeks later.

4. Active workstreams. For each project or initiative: name, status, what changed this call. Carried forward in priority order.

5. Open commitments - both sides. What you owe them. What they owe you. With dates if dates were said out loud. This is the field that separates a real chief of staff from a glorified note-taker.

6. Decisions made and decisions pending. A decision made on Tuesday is different from a decision still pending on Tuesday. The brain tracks both, and migrates items from "pending" to "made" as the relationship moves.

7. What I should remember next time. One to three lines of human texture - how the person is feeling, what they're under pressure on, anything that would change how you run the next conversation. This is the field that earns the "chief of staff" label.

Every update either modifies a specific field or carries it forward intact. No silent deletions. No quiet rewordings.

The Claude prompt (paste-ready)

Paste this into wherever your Claude surface keeps durable instructions - a Cowork session prompt, a Claude Code CLAUDE.md, or the custom instructions on a Claude Project. It's the contract Claude follows every time you ask for an update.

You maintain a persistent Account Brain for a single account.

Inputs available to you:
1. account-brain.md - the current Account Brain for this account.
2. Meeting transcripts from Earmark - markdown files with YAML
   frontmatter (title, date, participants, workspace) and a
   ## Transcript section. They may live in a folder you can read
   directly, attached to a Project, or pasted into the chat.
3. The most recent transcript, which the user will name when they
   ask you to run an update.

When the user asks you to update the Account Brain, produce the full
updated brain in this fixed seven-field structure:

## Stage of relationship
## Current spend tier or scope
## Health (with one-sentence justification)
## Active workstreams
## Open commitments - both sides
## Decisions made and decisions pending
## What I should remember next time

Rules:
- If a section was present in the prior brain and the call did not
  touch it, carry it forward unchanged.
- If the call contradicts prior state, prefer the call. Tag the
  change with "(updated {{CallDate}})".
- If the transcript does not contain the information needed for a
  section, write "Not discussed on this call" rather than inferring
  it.
- Never name an owner the transcript does not name. Use
  "Ownership unclear" instead of guessing.
- Quote the speaker verbatim when capturing commitments or sensitive
  language. Paraphrasing is where memory drifts.

Output the full updated brain only. Do not summarize the call
separately - Earmark already produced the call artifact for that

Three constraints are doing the heavy lifting here: carry-forward, contradiction-tagging, and the no-inference rule. They turn Claude from a summarizer into a memory layer.

Two shortcuts inside Earmark itself

You don't always have to leave Earmark to get continuity. Two built-in features cover lighter-weight versions of the same idea:

• Customize context (for the next meeting). Before a follow-up call, open the meeting in Earmark, click Customize context, and paste the relevant fields from your current Account Brain (Health, Open commitments, What I should remember next time). Every artifact generated during the call will be grounded in that context. This is the fastest way to make Earmark itself "remember" the last call.

• Cmd+K command menu. Search across every past meeting, artifact, and task for this account. Useful when you need a single quote or decision and don't want to round-trip through Claude.

Use the Account Brain for the durable memory. Use Customize context to inject that memory into the next call. Use Cmd+K when you just need to look one thing up.

What this looks like after four weeks

The first week, the brain feels redundant. You already remember last week's call. By week three, you start to notice things you would have lost: a commitment from two calls ago the customer never followed through on, a workstream that's been "paused" longer than you realized, a quiet shift from "scaling" to "steady" nobody flagged.

By week six, What I should remember next time is the first thing you read before any call. A summarizer-style tool would have buried that line in last month's notes. An Earmark transcript would have captured it. Claude, running this prompt against an accumulating brain, surfaces it back to you at exactly the moment it matters.

That's the difference between an AI tool and an AI colleague.

Where to start

If you want to stand this up this week:

1. Turn on local transcript auto-save in Earmark (Settings → Personalization → Meeting files) if it isn't already.

2. Pick three accounts. Not all of them. Three.

3. Pick your path. If you're on Cowork or Claude Code, point Claude at the Earmark transcripts folder and create three account-brain.md files. If you're on claude.ai, create three Projects. Paste in the prompt either way.

4. For two weeks, after every meeting on those three accounts, run the five-step loop. Review every Claude update before committing.

5. At the end of two weeks, read the brains. If they feel like a real picture of where each relationship stands, the pattern is working. If they feel like a stack of summaries, tighten the rules in the prompt - usually the carry-forward rule or the no-inference rule is the one that needs more weight.

The discipline is layering, not regenerating. Earmark gets you the transcript. Claude does the layering. You stay the final author.

A chief of staff who remembers is worth ten who summarize. Earmark and Claude, run together with the Account Brain in the middle, are the smallest combination that gets you there.