Good morning.
I had four conversations this week that were interesting.
A prospect call, a partnership pitch, an advisory client review, and a podcast booking.
In each one, the person on the other side noticed something specific about their business in the first two minutes.
Each one came from an agent I built once and have been refining for three months.
Most owner-operators are still treating prep as a thing they do (or skip) before a conversation. The ones getting ahead have stopped doing it themselves and have stopped skipping it. They built an agent that runs the prep for them, every time, on a context file they own.
Today's issue is that build. Full prompts, the architecture behind them, the platform decision you need to make before you start, and worked examples for the four business types I see most often on this list
— Sam
IN TODAY’S ISSUE 🤖
Why the Briefing Agent is the right first agent
The platform decision (read it once, decide, move on)
The 4-Pass Briefing Architecture
The Protocol: Seven steps, full prompts included
What this looks like for agencies, SaaS, ecommerce, and experts
How to hand this to your team without losing the edge
A bonus build if your team lives in Google Workspace
Let’s get into it.
Why The Briefing Agent Is The Right First Agent
You have a small number of conversations a week that move the business:
Sales calls and renewals
Partnership conversations
Investor or board updates
Customer interviews
Press and podcast bookings
The difficult internal conversations with senior team members that you keep putting off
The output of those conversations is wildly uneven. Some weeks you walk in sharp, recall the right detail, and leave the room with the outcome you wanted. Other weeks you forget a name, miss a piece of public news the other side expected you to know, and burn goodwill you have to earn back later.
Preparation is the variable that decides which kind of week you have. It's also the work you routinely skip, because the time cost of doing it manually is too high for the perceived upside on any single conversation.
A Briefing Agent eliminates the trade.
You describe the meeting once. The agent does the research. A one-page brief lands in your inbox a few minutes before the meeting starts. The marginal cost of preparation drops to near zero, and the compound effect across fifty conversations a quarter is enormous.
Three reasons this is the right first agent to build:
It pays for itself the first time it surfaces something you would have missed. That usually happens on run one.
It works across every part of your business that involves another human. Sales, customer success, partnerships, hiring, press. One agent, dozens of meeting types.
It lives entirely on your own context. No CRM integration required, no analytics warehouse, no clean data. You can build version one with nothing more than your calendar and email.
This is also where the Digital Employee frame starts to make sense.
A spreadsheet is a tool. An agent that reads your emails, looks up the person you're meeting with, pulls their last six months of activity, and writes you a one-pager before your 2pm is something else.
You manage it the way you would manage a new hire:
A clear job description (the system prompt)
The context they need to do the work (the context file)
A review cadence on the output (you read the briefs, you correct what's off)
Expanded trust over time (more meeting types, longer leash, scheduled runs)
That mental model is the one that produces operators who get real leverage out of agents, instead of operators who just collect subscriptions.
The Platform Decision (Read It Once, Decide, Move On)
Three serious paths for building a Briefing Agent right now. Pick one. Run it for thirty days before considering a second.
Claude Code In Cowork Mode
This is the path I run on, and it's the path most readers of this newsletter will get the most out of.
Runs on your desktop with direct access to a designated folder
Reads connected applications, can be scheduled
The context file lives as a markdown file inside the folder the agent has access to, so you can edit it, version it, and inspect everything the agent is doing
Pricing is a Claude subscription plus light usage costs
If you have already built any kind of workflow around Claude, this is the path of least resistance.
The Codex App
OpenAI rebuilt Codex earlier this spring. The April release turned it from a coding tool into a general agent workspace.
Computer use on macOS, an in-app browser, scheduled automations, persistent memory
Plugins for Microsoft 365, Slack, Notion, and Jira
Included with every paid ChatGPT plan, native desktop app on Mac and Windows
If you're deep in the Microsoft stack, or you run the business out of Slack and Notion and don't already have a Claude workflow, this is the path I'd recommend instead. The build that follows works in either platform with minor adjustments. (Introducing the Codex app)
OpenClaw or Paperclip
Stay out of this one for now. The hype cooled hard once enterprise security teams started auditing third-party skills and finding real exfiltration risk in publicly available extensions.
Operators who deployed it ninety days ago are now either running it inside heavy sandboxing (NVIDIA's NemoClaw add-on or similar) or have moved to one of the managed paths above.
Open-source agent frameworks will probably get there. Today they're not the build I'd start with if you have a real business to protect.
Quick Decision Rule
Already in Claude? Use Cowork mode.
Already in Microsoft 365, Slack, or Notion? Use the Codex app.
The 4-Pass Briefing Architecture
The mistake most people make when they try to build a research agent is asking the model to do too much in one prompt. The output reads like a Wikipedia summary because a single-pass prompt produces a single-pass output.
Useful conversations need layered context, and the agent has to build that context in stages.
The 4-Pass Briefing Architecture is the structure I run on every meeting that matters. Each pass is a separate call, each one narrows the focus, and the final synthesis is built on the work of the prior three.
The four passes at a glance:
Company. What the company looks like, factually, today.
Person. Who the human on the other side of the table is.
History. What your own record shows.
Synthesis. The three priors compressed into one scannable page.
Pass 1: Company
What does the company on the other side of the table look like, factually, today? Funding history if relevant, public news in the last 90 days, leadership changes, product launches, current positioning, top competitors.
This is the broad scan, and the prompt is written to suppress generic framing in favor of specific findings.
Pass 2: Person
Who is the human in the meeting? What have they said publicly, written about, posted about, or appeared on in the last twelve months? What can the agent infer about how they think, what they care about, and what's currently on their mind professionally?
This pass produces the texture that makes you sound like you did your homework, because you did.
Pass 3: History
What does your own record show about this company and person? Prior emails, calendar events, CRM notes, and any internal messages that touch this company.
The history pass is what separates a Briefing Agent from a generic research bot. It pulls from your own context and surfaces the thread of an existing relationship.
Pass 4: Synthesis
All three priors combined into a one-page brief you can scan in 90 seconds:
Company overview
Person overview
History with you
Three talking points
Two questions worth asking
One watch-out
You read it walking from the kitchen to the office, and you walk in sharp.
Why The Sequence Matters
Trying to compress all four passes into one prompt produces flatter output, because the model never gets to think about company before it thinks about person, and never integrates history until it's already committed to a frame.
The four-pass build is slower by a couple of minutes and noticeably better in the final brief.
The Protocol
Seven steps. Every prompt below is usable as written.
Step 1: Pick The Platform And Confirm Access
Pick Claude Code in Cowork mode or the Codex app based on the decision section above. Confirm the agent has access (via MCP) to:
Web browsing
Your calendar (Google Calendar or Microsoft 365)
Your email (Gmail or Outlook)
The folder or workspace where the context file will live (a local folder for Cowork, a workspace in Codex)
You do not need any CRM or analytics connections to run version one of this agent. Those connections raise the ceiling later. The agent works without them on day one.
Step 2: Build The Context File
This is the single most undervalued step in the build. The Briefing Agent is only as sharp as the context you hand it.
Create a markdown file the agent will reference on every run. Name it briefing-context.md and put it in the folder or workspace the agent has access to. Fill it out once. Update it monthly.
# Briefing Context
## About Me
- Name, business name, role.
- One paragraph on what the business does and who it serves.
- One paragraph on what I sell, at what price band, to what kind of buyer.
- The three positions I take publicly that distinguish me in my market.
- The two positions I do not yet hold publicly but believe.
## How I Sound
- The two or three writers, podcasters, or operators whose voice I would be flattered to be confused with.
- Three sentences from my own writing that capture how I sound. Paste real lines from things you've already written, not aspirational ones.
## What I'm Currently Working On
- One paragraph on the engagement, project, or build that's eating most of my attention this quarter.
- One paragraph on the one or two things I'm preparing to say or ship in the next 90 days.
## Who I Want To Be Read As
- One line on the kind of operator I want the person in the room to leave a conversation thinking I am.
- One line on what I never want to be mistaken for.
## What The Briefing Agent Should Always Surface
- Anything in the last 12 months suggesting the person is hiring, raising, restructuring, pivoting, or in the middle of a transition.
- Anything they've said publicly about AI, agents, or operational rebuilds.
- Anything we've discussed in prior emails or meetings, even if it was a year ago.
- Any mutual relationships or shared portfolio.
## What The Briefing Agent Should Never Do
- No generic "founded in X, headquartered in Y" filler unless the founding story is specifically relevant.
- No filler adjectives ("innovative", "leading", "groundbreaking").
- No talking-point suggestions that sound like sales coaching. I want the substance, not the script.
- No invented quotes, statistics, or attributions. If a fact isn't sourced, mark it as inferred.This file is the difference between a generic research output and a briefing that sounds like you wrote it.
The Tacit-to-Explicit Knowledge work happens here. Everything you know about yourself and your market that you haven't yet written down lives in your head, where the agent can't reach it. The context file is your first deliberate act of making the tacit explicit.
Step 3: Install The System Prompt
This is the agent's job description. Put it in a file called briefing-agent-system-prompt.md in the same folder.
You are the Briefing Agent for [Your Name].
Your job is to produce a one-page briefing document before a meeting or conversation [Your Name] is about to have.
Read briefing-context.md before every run. Everything in that file takes precedence over your defaults.
You will receive a briefing request that includes:
- The name of the person being met
- The name of the company they work for
- The meeting type (sales call, partnership, investor update, customer interview, media interview, internal review, other)
- The date and time of the meeting
- Any context [Your Name] adds in the request
You will run four sequential research passes before producing the final brief.
PASS 1: COMPANY
Research the company. Focus on public news in the last 90 days, funding history if applicable, leadership changes, product launches, positioning, top three competitors, and any signals of strategic movement (hiring patterns, restructuring, new market entry). Cite sources inline as ([Source Name](URL)).
PASS 2: PERSON
Research the person. Find their LinkedIn, recent public writing, podcast appearances, public talks, social posts, and anything they have said about AI, agents, business architecture, or the topic of the meeting in the last 12 months. Cite sources inline.
PASS 3: HISTORY
Search [Your Name]'s connected accounts (email, calendar, CRM if connected) for any prior contact with this person or company. Surface prior conversations, the dates, the topics, and any commitments made. If no history exists, say so explicitly.
PASS 4: SYNTHESIS
Combine all three priors into a one-page brief using the exact template below. Do not deviate from the template. Do not add sections. Do not pad. If a section has nothing to report, write "None found" rather than filling it with adjacent material.
TEMPLATE:
# Briefing: [Person Name] / [Company Name]
Meeting type: [Type]
Date: [Date and time]
## Company in 90 Seconds
[Three short paragraphs. Factual. Cite at least two sources.]
## The Person in 60 Seconds
[Two short paragraphs. Specific. Cite at least one source.]
## Our History
[Bulleted list of prior touchpoints with dates. "None found" if none.]
## Three Talking Points
1. [Specific, grounded in research, not generic.]
2. [Same.]
3. [Same.]
## Two Questions Worth Asking
1. [A question that signals you've done the work.]
2. [Same.]
## One Watch-Out
[A sensitive topic, a recent reversal, a public position that might come up, or a known dynamic to handle carefully.]A couple of things this system prompt is doing under the hood:
It tells the agent to suppress filler, and the context file backs that up by banning specific adjectives. The combination is what kills the corporate-research feel of generic AI output.
It tells the agent to write "None found" instead of filling space. That single instruction prevents the common failure mode where the agent hallucinates context because it feels obligated to produce content.
Step 4: The Four Pass Prompts
These are the runtime prompts you send when you want a briefing produced. The system prompt above governs the agent's behavior across all four passes. The prompts below trigger each pass in sequence.
Pass 1 prompt (Company):
Run Pass 1 of the briefing for the meeting below.
Person: [Name]
Company: [Company]
Meeting type: [Type]
Date and time: [Date / time]
Additional context I want you to know: [Anything specific you want surfaced]
Produce the COMPANY section only. Pull from public sources. Focus on the last 90 days where possible. Cite inline. Do not generate any other sections in this pass.Pass 2 prompt (Person):
Run Pass 2 of the briefing.
Same person and company as Pass 1.
Produce the PERSON section only. Look for: LinkedIn, recent writing, podcast appearances, public talks, and anything they have said about AI, agents, business architecture, or the topic of this meeting in the last 12 months. Cite inline.Pass 3 prompt (History):
Run Pass 3 of the briefing.
Search my connected email, calendar, and (if connected) CRM for any prior contact with this person or this company. Surface dates, topics, and commitments. If no history exists, say "None found" explicitly.
Do not invent context. Do not infer prior contact from similar names.Pass 4 prompt (Synthesis):
Run Pass 4. Produce the full one-page brief using the TEMPLATE in your system prompt.
Pull from the work you did in Passes 1, 2, and 3.
No filler. No generic talking points. If a section has nothing real to surface, write "None found".
Deliver the brief as a markdown document I can read on my phone walking to the meeting.Step 5: Run Your First Briefing
Pick a real meeting on your calendar this week. Not a test. A real one.
Send Pass 1. Read the output.
Send Pass 2. Read the output.
Send Pass 3.
Send Pass 4. Read the final brief.
The first run will be 80% there. The company section will be sharp. The person section will be roughly right. The history section will be right or empty. The synthesis will sound a little too neat.
That's expected. Take the brief into the meeting anyway. After the meeting, write down two things:
What the agent surfaced that genuinely helped
What it missed that would have helped
Step 6: Tune After Three Runs
Three real meetings in, you'll have a clear sense of where the agent is over-producing and where it's under-producing. Update the context file with the corrections. Sharpen the "what the briefing agent should never do" section based on what it kept doing anyway.
Four common tunes after three runs:
Too much old news. Add a date filter to the context file: "Anything older than 12 months is not relevant unless I ask for it."
Company section reads like a press release. Sharpen the context: "I do not need positioning language. I need facts and signals."
History is being invented (the agent inferring prior contact from a similar name). Tighten Pass 3: "If you have not found a direct email, calendar, or CRM record, you have not found history. Do not infer."
Talking points sound like sales coaching. Add: "Talking points should be observations or questions a peer operator would raise, not openers a sales rep would use."
Step 7: Set The Schedule
The Briefing Agent should run automatically. You should not be firing off four prompts on the morning of every meeting.
Both Claude Code in Cowork mode and the Codex app support scheduled automations. Set the agent to run the full four-pass briefing at a fixed time each morning, for every meeting on that day's calendar where you're hosting an external person.
Every morning, your inbox or designated folder fills with one briefing document per outbound meeting. You scan them with coffee. The day is sharper before it starts.
What This Looks Like By Business Type
The Briefing Agent stays the same. The meeting types and surfacing priorities change.
If You Run An Agency: The Client QBR Brief
The most useful brief for a quarterly business review pulls together:
Public news about the client's market over the last 90 days
Hiring patterns at the client (a hiring spike on the sales side often signals expectations the agency hasn't been briefed on)
Any recent leadership changes
Your own prior email and Slack history with the client over the last quarter
Your Three Talking Points section becomes the second half of the agenda. The One Watch-Out section is where the agent flags a Slack message from six weeks ago where the CMO expressed frustration. You should not walk in pretending not to have seen it.
If You Run A SaaS Company: The Investor Update Brief
The brief reads more like a market scan:
Pass 1 picks up the investor's recent portfolio activity, public commentary on the SaaS category, and what they have said about AI agents in their most recent letter or podcast.
Pass 2 produces a read on the partner's current thesis evolution.
Pass 3 surfaces the commitments you made on the last update.
The Two Questions become the questions you ask back, which is half the point of a good investor update. The Watch-Out flags anything the partner has publicly said that contradicts the direction you're about to present, so you can address it directly instead of stepping into it.
If You Run An Ecommerce Business: The Supplier Negotiation Brief
The brief is tactical:
Pass 1 surfaces recent public news about the supplier's business (capacity expansion, raw materials cost pressure, public commentary on margins).
Pass 2 produces a read on the specific account manager, including how long they have been in role and whether they have moved up or sideways recently.
Pass 3 pulls your order history, prior pricing changes, and any commitments either side made on the last call.
The Watch-Out is usually a pricing pattern in the supplier's recent activity you can leverage in the negotiation without showing your hand.
If You're An Expert Or Creator: The Podcast Or Media Brief
The brief is the most personal of the four:
Pass 1 covers the show or publication, recent episodes, audience composition, and the topics the host has been pulling on lately.
Pass 2 is a deep read on the host, including positions they hold publicly that you should be ready to engage with.
Pass 3 surfaces any prior interaction with this host or publication.
Your Three Talking Points become the angles you want to be sure you get into the conversation. The Watch-Out is whatever topic the host has been chasing lately that could derail your prepared angles if you don't see it coming.
The pattern across all four: the brief is short, dense, and built on real inputs. You read it once and walk in with a clearer read on the room than the person on the other side, who probably did not run an agent against you.
Deploying This Across Your Team
The Briefing Agent works as a personal build. It compounds when your team is using it too.
Four pieces of ownership keep this clean:
You own the architecture. Context file, system prompt, and briefing template are your IP. They live in a shared folder the team can read but cannot modify. That single permission boundary keeps output consistent across every team member who runs it.
Anyone who runs external meetings runs the agent. Account managers, BD, sales, CS leads, sometimes the EA. They use the same upstream artifacts, which means the brief that lands on an AM's desk looks indistinguishable from the brief that lands on yours.
The EA owns the schedule. Their morning routine becomes: confirm the agent ran overnight, scan the briefs that landed, route the ones with sensitive Watch-Outs directly to you, and let the team-facing ones sit in the shared folder for the team member who owns the meeting.
You own the review. Once a month, read three or four random briefings the agent produced and look for drift. Drift usually shows up as press-release language sneaking back in, or as over-reliance on the same three sources. Ten minutes of context file tuning holds the bar for another thirty days.
The compound effect is straightforward. Every external conversation starts at the same level of preparation. Variance in customer experience drops. Your team gets sharper without anyone having to teach them anything, because the agent is doing the teaching by example.
Bonus: The Gemini Workspace Build For Teams Living In Google
If your team lives in Google Workspace and you don't want to set up a separate agent platform yet, there's a lighter version of this build that runs inside Gemini in Docs and Sheets. Less powerful than the full Briefing Agent. Real value with zero setup beyond a Workspace account.
Gemini In Docs: The Lightweight Briefing Prompt
Create a one-page meeting brief for my [Meeting Type] with [Person Name] at [Company Name].
Pull from:
- My emails in the last 12 months mentioning [Person] or [Company]
- My calendar events with [Person] or anyone at [Company]
- Public information about [Company] in the last 90 days
- Public information about [Person] in the last 12 months
Format:
- Company in 90 seconds (three short paragraphs)
- The person in 60 seconds (two short paragraphs)
- Our history (bullet list of prior touchpoints with dates)
- Three talking points (specific, not generic)
- Two questions worth asking
- One watch-out (sensitive topic or known dynamic)
No filler. No marketing language. Cite sources for public claims. Use "None found" if a section has nothing real to report.
Gemini In Sheets: The Meeting Tracker
Build me a meeting tracker with the following columns:
- Date
- Person (with company in parentheses)
- Meeting Type
- Briefing Generated (Yes / No)
- Outcome (short text field)
- Follow-up Owed (date)
- Status (Open / Closed)
Pre-populate the tracker with all external meetings on my calendar for the next 30 days.
Add conditional formatting: yellow if Briefing Generated is "No" and the meeting is in the next 48 hours, red if Follow-up Owed is past due.
Add a pivot at the bottom showing meetings per type per week, so I can see whether I'm spending my time on the right kinds of conversations.The Docs prompt produces the brief on demand. The Sheets prompt makes your conversation portfolio legible.
Run together, they replicate a meaningful slice of the Briefing Agent build with no setup beyond a Workspace account.
Run the Briefing Agent for thirty days and you stop being the person in the room who forgot the detail.
You walk into renewals remembering the last commitment. You walk into investor calls remembering the objection the partner raised six months ago that nobody else in the room caught.
Build it this week.
Talk soon,
Sam Woods
The Editor
.
