What’s waiting when I open Slack at 7am.

I opened Slack this morning and four messages were already waiting from my own bot. One had validated my Gmail connection and pulled the inbox. One ran a daily digest on a channel I follow. One fetched the last 48 hours of email and flagged what needed action. The last one was checking Salesforce activity logs for yesterday’s customer meetings to figure out what I should do today.

I hadn’t typed anything. It just runs.

For the last few months I’ve been building a sales operating system inside Slack, one skill at a time. It’s the real thing I work out of every day. This week it crossed the line from “thing I’m building” to “thing that runs whether I show up or not,” so it felt like the right time to write down how the whole thing fits together.

Two rhythms, two chains

Sales work splits cleanly into two kinds of time.

There’s the quarterly work. The stuff you do once when you inherit a territory or start a new year, then mostly forget until the next planning cycle. And there’s the daily work, the live execution of moving deals forward. Most “AI for sales” content only ever touches the second kind, and usually just one piece of it. I wanted skills for every step of both.

So the system has two chains.

The whole system, one diagram. Early chain on top, daily chain on the bottom.”]

The early chain: territory intelligence

This one runs once a quarter. Eight skills, in order, each output feeding the next step’s input.

It starts by building a complete account inventory for my territory. Then it scores those accounts against my ideal customer profile so I know where focus belongs. From there it maps the whitespace, finds the key contacts and the buying-committee influencers, and detects the signals that say an account is heating up. It drafts a strategic account plan with next best actions, researches the personas I’m about to reach out to, and ends by building a personalized, multi-touch outbound sequence.

One rule sits underneath the whole chain: account status drives the motion. Where an account sits tells the system what to do with it. I’m not making a hundred small judgment calls. The status makes most of them for me.

The output of that chain is a territory I understand and a set of sequences ready to send.

The handoff

Then something works. A meeting gets booked.

That’s the single point where the two chains connect. The moment a meeting lands, the daily skills take over and carry the conversation from there.

The daily chain: deal execution

This is the half that now runs on a schedule without me.

At 7am on weekdays, a Daily Priorities skill checks my Salesforce activity logs, looks at yesterday’s meetings, and tells me where the day should go. At 8am an Email Digest validates the inbox connection, pulls recent mail, and surfaces what needs a reply. Monday at 7:30 and Friday at 3, a Weekly Review runs the same way. Those are the auto-runs. I don’t trigger them. They show up.

The daily side, on a schedule.

Then there’s a set I trigger by hand when the moment calls for it. A discovery question bank before a call. A call summary after. Pipeline forecasting and deal coaching when I’m working the number. An activity logger that writes the work back into Salesforce so I’m not doing CRM hygiene at 6pm.

There’s a whole library of these skills now. Deal Brief, Account 360, Account Coaching Overview, Account Resolver, the Activity Logger. I reach for them by name the same way I’d reach for a teammate who’s good at one specific thing.

The test I keep running it against

The skills a new seller would inherit on day one.

Here’s the question I use to know if it’s working: could a brand new AE sit down at this on day one and produce?

Not “could they learn the tools.” Could they open Slack and immediately know what to focus on, where their time should go, who to message, and how to keep every conversation moving forward? That’s the bar. The judgment that usually takes a couple of years to build is starting to live inside the flow itself, where a new seller can borrow it from the first morning.

That’s the part I care about most. Building the skills was the fun part, and the easy part. Getting a system to the point where it runs on its own, gets used over the old way, and survives a busy week, that’s the harder problem. It’s also the only one that matters. A skill nobody opens twice is just a demo.

This one gets opened every morning. So far, so does the rest of it.

Want one for your role?

The system above is built for a seller, but the method underneath it isn’t sales-specific. You map the role into its real steps, build a skill for each one, and then comes the hard part: getting it to run on a schedule so it survives a busy week. I’ve run that loop on my own job and a few others, and it holds up the same way every time.

If you want one of these inside your own Slack, that’s the work I do. We start by mapping your role into the two chains and naming the handful of skills worth building first. Then we build them where your work already happens and get them to stick.

Two ways in. Want to build it yourself?

Subscribe — the newsletter ships skill templates and role blueprints as I write them.

Anthropic now gives you four models in the picker: Haiku 4.5, Sonnet 4.6, Opus 4.8, and the new Fable 5. Most people I see do one of two things. They never touch the dropdown, or they run everything on the biggest model because bigger must be better.

Both approaches burn the same thing: your usage limits, spent on the wrong work.

Here’s the 2-second fix. Ask yourself: what does it cost you if the output is wrong?

If wrong means a quick retype or a rerun, go small. If wrong means a bad decision in front of a customer, your VP, or your team, go big.

That’s the whole framework. Everything below is just that question applied four times.

Haiku 4.5 — the bulk lane

Haiku is the fastest and cheapest of the four, and Anthropic says it matches older Sonnet-tier performance on a lot of coding and agent tasks. It’s built for high-volume work where wrong costs you a few seconds, not a few hours.

Think extracting action items from long email threads. Cleaning notes into tables. Tagging a few thousand records. If you’d describe the task as “tedious but obvious,” Haiku does it in seconds and barely touches your limits.

Use Haiku when you’re doing bulk classification, tagging, or format conversion. Use it when you’re summarizing low-risk content for your own understanding, and the worst-case outcome is “I run it again.”

Skip it the moment a task needs judgment, nuance, or more than one real reasoning step.

Sonnet 4.6 — the default

Sonnet is the workhorse, and it’s where most of your day should live. Drafting posts. Rewriting emails in your voice. Turning meeting notes into plans. Summarizing long docs with actual context. It delivers the best quality per token for general work, which matters if you’re in Claude all day and care how fast your credits disappear. I am, and I do.

Use Sonnet when the output touches other people but isn’t life-or-death. Use it when you want strong reasoning and style control without paying Opus or Fable rates, and when you’re iterating fast enough that speed plus quality beats absolute depth.

Only leave Sonnet for the hardest 10% of your work.

Opus 4.8 — high stakes

Opus is slower on purpose. It thinks longer and earns that back in accuracy on multi-step work where correctness is the whole point.

Use it where errors compound: pressure-testing a launch plan, walking a contract clause by clause, building a financial model where one bad assumption poisons everything downstream.

I treat Opus like a senior reviewer. Not for everything, but for the work I can’t afford to redo. If you’re making a decision you can’t easily undo, or reviewing something dense with real downside, this is the tier.

Fable 5 — the agent

Fable is the new one, and it belongs in a different category than the other three. It’s Anthropic’s first Mythos-class model made generally available; the Mythos line was previously restricted to a small set of approved partners. Where Opus thinks harder about one thing, Fable runs an entire project. It plans its approach, works for days inside an agent harness like Claude Code, delegates to sub-agents, checks its own progress, and refines until it hits the goal.

It’s also the priciest model Anthropic sells: $10 per million input tokens and $50 per million output, which is double Opus 4.8 at the token level. So don’t ask it what time zone Lisbon is in.

One thing worth knowing right now: Fable 5 is included on Pro, Max, and Team plans at no extra cost through June 22. If you want to feel the difference before it starts drawing from your credits, this is the window.

Five Fable builds that actually justify the tier

Here’s where Fable stops being a launch headline and starts looking like a different tool.

1. Codebase-wide migrations in a day. Anthropic says one fintech customer ran a migration across a roughly 50-million-line codebase in about a day, work they’d estimated at two months for an engineering team. Fable planned the migration, updated code across services, and kept tests in sync instead of patching one file at a time. You don’t need this tier for a bug fix. You reach for it when the job is “rewrite a core abstraction everywhere and don’t break anything.”

2. Multi-day autonomous coding runs. Earlier models could push a project forward if you built a lot of scaffolding around them. Fable maintains its own plan. Teams are handing it a repo, a goal, and some guardrails, then letting it work through multi-file features end to end without constant nudging.

3. Senior-analyst-grade research. Early finance benchmarks put Fable at senior-analyst-level reasoning on dense, document-heavy work. The interesting use isn’t “summarize this PDF” (Sonnet handles that). It’s the job that looks like what a senior analyst would do over a week: deep dives across filings, internal docs, and data tables, coming back with structured analysis, scenarios, and stress tests.

4. Vision-heavy reverse engineering. Anthropic has shown Fable reconstructing web apps from screenshots and pulling precise values out of dense charts and scientific figures. For basic OCR you’re overpaying. It earns the tier when the visual input feeds a bigger reasoning or coding task, like looking at a UI and proposing the code behind it.

5. Long-horizon agents for research and ops. Fable is showing up as the brain behind agents that pursue a goal across many steps: running experiments, reconciling conflicting sources, orchestrating internal workflows over days instead of giving one-shot answers. The closest comparison I’ve found is a junior team that keeps its own to-do list and actually works through it.

The rule I actually run

My routing is simple. Default to Sonnet. Step down to Haiku when it’s bulk and low-risk. Step up to Opus when the cost of being wrong is high in a single document or decision. Escalate to Fable when the task looks like a project a small team would own.

Then I try to remove the decision entirely.

Limits got reshaped this year and Fable runs at twice Opus’s token price, so model choice now has real economic weight. Inside my own systems, the routing is baked into the tools. The daily Slack skills run light. Most work defaults to Sonnet. Only top-risk or project-scale tasks dispatch to Opus or Fable.

Because in the moment, people don’t optimize. They click whatever’s selected. Defaults beat decisions, and that’s exactly where most AI rollouts stall: the demo was great, but nobody wired in sane defaults. That’s a RUN-phase problem, and it’s the part almost nobody teaches.

Set your default once. Encode the routing into your workflows. Let the picker fade into the background.

If you’re done treating AI like a chat box: I map roles into AI Operating Systems. Map. Build. Run. A MAP Session is 75 minutes and you leave with your role rendered as a system blueprint. Reply to this email if you want one.

— Ant

▶ FIELD NOTES · ISSUE №002 · THE LONG LENS

The setup

This week’s AI Daily Brief ran an episode called “Why Agents Still Need Humans.” It dropped into the same stretch as a pile of 2026 labor data — Microsoft’s Work Trend Index, LinkedIn’s Labor Market Report, an NBER study, a Gartner forecast. Read them together and they cut against the headline everyone’s been writing since agents went mainstream.

The headline says agents replace people. The data says something quieter, and more useful: automate a task and the human work around it doesn’t disappear. It moves up a layer.

I’ve spent 2026 watching how the genuinely AI-native companies operate — not the ones that bought seats and called it a strategy. The pattern is consistent enough to build on.

What most people are missing

Everyone’s counting the work agents remove. Almost nobody’s counting the work they create.

LinkedIn’s 2026 Labor Market Report puts it at at least 1.3 million AI-related roles created in the last two years — data annotators, AI engineers, forward-deployed engineers. Jobs that barely existed five years ago. That’s not the shape of a technology that erases work. It’s one that reshuffles it upward.

Here’s the part you only see from inside the work: when every team runs the same agents on the same training data, the output converges. Sameness is the default setting of an agent economy — and sameness is what makes difference expensive. The slop is already everywhere. The scarce thing is the person who can frame the problem, judge the output, and push it past what the model would do on its own.

The mechanism

Here’s how the work actually restructures. Call it the human sandwich.

Humans set the frame → the agent collapses the task → humans judge and extend the result. You’re not cut out of the workflow. You move to both ends of it, as the decision layer. IBM Research keeps landing on the same result: human-AI teams beat either the human or the agent alone.

And the people doing the work already behave this way. In Microsoft’s 2026 Work Trend Index, 86% of AI users say they treat AI output as a starting point, not a final answer. The judgment seat doesn’t empty out as the models get better. It gets busier.

Three forces sit underneath this:

Shared agents beat personal agents. The every.ai team learned it the expensive way — solo agents break constantly, and the upkeep lands on one person. Agents wired into a team’s workflow are where the return actually shows up. The unit of leverage is the team, not the seat.

The floor rises before the ceiling does. An NBER study of customer-support reps given agent access found a 14% productivity gain — concentrated among the average and newer workers, not the stars. Agents close the gap between your best people and everyone else, which frees your best people for the work only they can do.

Delegation scales. Gartner projects that by 2028, 15% of day-to-day business decisions will run autonomously. That’s not replacement — it’s delegation, the move every good leader already makes. The question shifts from “will an agent decide this” to “which calls do I hand off, and how do I check the work.”

Who wins, who loses, what to watch

• Who wins: Teams that build the human skill of steering agents — framing, judgment, extension. Operators who run agents as a system, not a feature they switched on.

• Who loses: Companies optimizing for headcount reduction alone. They hit the efficiency ceiling, ship the same converged output as everyone else, and wonder why the savings never turned into growth.

• What to watch (6–12 months): Whether the net-new AI roles in the LinkedIn and Microsoft data keep compounding or flatten. Climbing confirms the delegation thesis. Hiring stalling while agent spend climbs hands the replacement camp its evidence.

What this means for the operator

One decision this quarter: stop scoring your AI rollout by cost removed. Score it by decisions delegated and floor raised.

Concretely — take one workflow your team already runs. Mark where a human sets the frame, where the agent collapses the task, where a human signs off. Make that agent shared, not personal. Then watch whether your average performers move up. That number tells you more than any efficiency metric will.

Efficiency is table stakes. Growth is the game.

Your artifact this issue: the human-sandwich frame — humans frame → agent collapses → humans judge and extend. Run it on one workflow this week.

Subscribe for more field notes from inside the work — where AI runs as an operating system, not a chat window.

— Ant

Here’s the exact 3-step workflow behind the clip — tools, prompts, and what broke on the first attempt.

Last week I posted a video that made my mom call me to ask if I had actually gone to Augusta.

I had not.

I was at my desk. I had no camera crew, no production budget, and no prior video editing experience. What I had was a clear idea, three AI tools, and about two hours.

The clip: a fan in the gallery at a major golf championship gets spotted by the pro golfer — who looks a lot like Tiger Woods on a Sunday — pointed out from the crowd, summoned under the rope, handed the putter, and asked to attempt the winning putt on the 18th green. In front of 40,000 people. On live television.

The fan makes it.

The crowd loses its mind.

The fan — who is me — cannot process what just happened.

The whole thing is 45 seconds. It looks like something that went viral off a real broadcast feed. And every single frame of it was generated with AI.

Here is exactly how it was built.

The 3-Tool Stack

Before I get into the steps, here are the tools:

• ChatGPT Image Gen 2.0 — created the reference still that anchored everything

• Claude in Cowork — wrote and iterated the full production document before a single frame was generated

• HeyGen Avatar Shots (Seedance 2.0) — generated the actual video clips

That’s it. No Premiere Pro. No After Effects. No camera. No green screen.

Step 1: Build the Scene Before You Build the Video

The first thing I needed was a reference image — not a headshot, not a phone photo, but something that put me visually inside the broadcast environment.

I uploaded my photo to ChatGPT Image Gen 2.0 alongside this prompt:

“Masters broadcast screenshot. A fan in a navy blazer has just been summoned from the gallery onto the 18th green. Lower-third graphic reads: FAN SUMMONED TO ATTEMPT THE WINNING PUTT. Leaderboard score bug bottom-right. Augusta atmosphere — azaleas, Georgia pines, green wooden scoreboard visible in background. The fan looks surprised and slightly amused. Live television — not a commercial.”

The output looked like it was pulled directly from a CBS Sports feed.

That image became the anchor for everything else. When you upload a reference photo into HeyGen, it gives the AI model your face, your outfit, and the surrounding environment all at once. One good reference image does more work than any text description can.

If the output looks like a stock photo or a render, add: “broadcast compression artifacts” and “720p television quality” to the prompt. That usually fixes it.

Step 2: Write the Production Document Before You Touch HeyGen

This is the step most people skip, and it’s why most people waste credits.

HeyGen Avatar Shots charges credits per generation — 4 credits per second, with no refunds for bad results. A 15-second clip costs 60 credits. Generate three of them badly and you’ve burned 180 credits on footage you can’t use.

The production document I wrote before generating a single clip had four parts:

1. The Consistency Block A chunk of text you copy-paste into every single shot prompt. It locks the subject appearance, environment, broadcast aesthetic, and lighting so every generated clip looks like it came from the same shoot. Without this, your clips will look like they were generated in three different universes.

2. Shot-by-Shot Prompts with Timestamps Each shot is a maximum of 15 seconds. Within that 15 seconds, you define specific beats using timestamps — [0s-5s], [5s-10s], [10s-15s] — and describe exactly what should happen at each moment. Camera angle, avatar action, crowd behavior, audio.

3. The Broadcaster Voice-Over Script Three acts, two voices. Play-by-play (higher, faster, chaotic) and color analyst (lower, slower, reverent). The energy arc across the three shots is what makes it feel like real television:

• Shot 1: “Wait... wait wait wait — is he pointing into the gallery? He IS. I have NEVER seen this in 30 years covering this tournament.”

• Shot 2: (barely above a whisper) “Forty thousand people. Not a sound. This man has never putted in front of more than his weekend foursome.”

• Shot 3: “IT’S IN!!! IT WENT IN!!! I DON’T BELIEVE WHAT I JUST WITNESSED!!!”

I recorded the VO in ElevenLabs and layered it over the clips in HeyGen AI Studio after generation. HeyGen also generated its own broadcaster-style audio from the scene description — it sounded nearly identical to the real thing.

4. The Broadcast Realism Line This is the single most important thing I learned. This paragraph — included in every shot prompt — is the difference between “looks AI-generated” and “wait, is this actually real?”:

“NTSC broadcast color profile, over-saturated greens, 720p compression grain, chroma artifacts on high-contrast edges. Masters-style lower-third graphic bottom-left (dark green, gold text). Score leaderboard bug bottom-right. Network watermark top-right, semi-transparent. Authentic live television — not cinematic, not commercial.”

Telling the model what kind of footage you’re making — technically, specifically — changes the output entirely.

Step 3: Generate, Fix, Generate Again

The first two shots came out well. Shot 3 — the tension shot, the one where the subject stands alone on the green before putting — had problems on the first attempt.

Here is what went wrong and why:

Double putter in frame. When you describe a prop from multiple angles across several timestamps in the same generation, the model loses count. It generated two putters. Fix: add “ONE putter only, held in both hands” explicitly and reference the prop as little as possible across multiple angle descriptions.

Avatar speaking without any script. Close-up face shots trigger lip-sync behavior even in Voice-Over mode. The model sees a face near the camera and assumes it should be talking. Fix: add “lips remain closed and still throughout, no dialogue, no lip movement” — explicitly, in caps if needed.

Double holes on the green. This one surprised me. When you describe the cup from multiple perspectives — including a ground-level camera insert looking across the green — the model generates multiple cups. Fix: add “ONE hole on the green, located 20 feet directly ahead of the subject” and never ask for ground-level camera angles in the same generation as overhead or medium shots.

The root cause of all three: I asked for five distinct camera angles within a single 15-second generation. That’s too many perspective shifts for the model to maintain spatial consistency. I simplified Shot 2 to three beats — wide, medium close-up, medium — and it generated cleanly on the second attempt.

The Final Structure

Three shots. Hard cuts between each. No transitions.

Shot Duration Energy Shot 1 15s The summoning — golfer points, fan crosses the rope, putter handoff Shot 2 15s The tension — alone on the green, 40,000 people silent Shot 3 15s The putt, the drop, the eruption

The tonal whiplash between each hard cut — chaos to silence to chaos — is what makes it feel like live television. Transitions would have ruined it.

What This Actually Means

Two years ago, a clip like this required a production crew, a golf course, release forms, and a real budget.

Today it requires a clear idea, the right prompts, and two hours.

I’m not saying AI is replacing creativity. It’s doing the opposite. It’s removing the distance between what you can imagine and what you can actually make. The constraint used to be resources. Now the constraint is clarity of vision.

The people who figure out how to think precisely — how to describe what they see in their head in terms a model can act on — are going to have an enormous advantage in the next few years.

That’s what AI with Ant is about. Practical workflows, real outputs, nothing theoretical.

Subscribe now

Anthony Brown is an Account Executive at Salesforce and the founder of AI with Ant — a Substack covering practical AI workflows in sales, content, and business..