DEV Community: T.M. Gunderson

Your AI Instructions Keep Dying in Chat Threads: Put Them in a Repo Instead

T.M. Gunderson — Thu, 05 Nov 2026 09:00:40 +0000

You've probably had this experience: you spend an evening teaching ChatGPT or Claude how your business works — your pricing, your tone, how you want customer emails handled — and it goes well. Then the next morning you open a fresh chat, and you're back to zero. Re-explaining everything. Re-pasting the instructions. The knowledge didn't survive the conversation.

Most small business owners treat this as an AI limitation to work around. Figma's product-management team treats it as a filing problem — and their fix is worth copying in a much smaller version.

The problem your chat window can't solve

Chat context is disposable by design. Close the tab and it's gone. Open a new thread and the AI starts blank. That's fine for one-off questions. It's terrible for operating knowledge — the stuff that should never get forgotten: how you quote jobs, what you promise in writing, which customers get special terms.

So people do the obvious thing: they keep pasting their instructions into every new chat. Some keep a "master prompt" document open on a second monitor. The knowledge exists, but it lives in the worst possible place — a place that resets every time you need to use it.

What Figma's PM team actually did

Ezra Mechaber, a growth product manager at Figma, described the setup on the How I AI podcast: an internal system the team calls a "PM OS" — not a product, just a repo. It holds about 31 documented skills covering the recurring work of being a PM: writing documents with the right reviewers looped in, briefing on open decisions, building small dashboards, even a first pass at coaching conversations.

His goal, in his words, was to be "lazy in the right places" — help with jobs he did poorly or forgot entirely, like looping support into a launch or finding who owns a piece of the product. Podcast host Claire Vo's framing is worth keeping: AI should "raise the floor," not just "take off the ceiling." The glamorous demos get the attention, but remembering to involve support removes a predictable failure from the operating system.

Two details matter more than the scale:

The skills live in a repo, not in anyone's chat window. Shared parent context, private workspace per person. When one PM improves a skill, everyone gets the improvement.
They prune it. Figma's PMs were invited to "roast" the skills at an offsite; the roasting decided which ones were worth keeping. Mechaber has been removing instructions as models improve. The repo is maintained operating context, not a permanent monument to prompt engineering.

Why "skill" is the right word

A skill is not a prompt. A prompt is what you say once. A skill is documented know-how that says when it applies and what it does. Four parts:

When to activate. "When a new review appears on my Google Business Profile" — not "whenever."
What to do. The exact steps: check the rating, classify the complaint, draft a response.
How to sound. Friendly, plain, never defensive — your tone, written down once.
What to reference. Your pricing sheet, your service area, your refund policy, so drafts cite real facts.

If you can't write those four lines for a task, you don't have a skill yet. You have a vibe — and vibes are exactly what dies in chat threads.

The five skills worth writing down first

A small business doesn't need 31. Five files cover most of the recurring operating knowledge in a service business:

The morning briefing. Open estimates, unanswered reviews, overdue invoices, today's appointments. Mechaber's most-used skill is exactly this — he runs a "What's Up" briefing about six times a day. One detail worth stealing from their instructions: the briefing reads whole threads, not isolated messages, because a briefing is only useful if it can tell a new request from one that was resolved ten messages later.
Customer replies. How you sound, what you promise, what you never put in writing — plus your FAQ: hours, service area, lead times.
Quote and estimate follow-ups. Your rules for chasing: how many touches, over how long, when to let go.
The weekly review. What shipped, what stalled, what the numbers say — run from your records, not from memory.
"How we do things here." The employee-handbook skill: escalation rules, who can promise what, the standard a draft must meet before it goes out.

Five files. An evening of writing, once.

The repo part matters more than any single skill

Where the files live is what makes this work:

One shared location, not five chat accounts. A single folder the whole team can see — a cloud drive, a repo, even a well-kept notes app. If your instructions are scattered across three chat threads and a text file on your laptop, they are already lost.
One owner per skill. When someone spots a bad instruction, they fix it at the source — and every future use inherits the fix. Feedback through conversation, not a suggestion box that never gets opened.
Pruning is the maintenance. A skill that keeps producing output you have to fix is worse than no skill. Figma's roasting session is the right instinct at any size: use, edit, delete what doesn't earn its place.
The AI changes; the files transfer. Figma's team started in one coding tool and now mostly works in a different one. The durable investment was the context and skills, not loyalty to an interface. Whatever model you use next year, the skills come with you.

How to hand a skill to an AI agent

Written files only pay off when you point the AI at them. Two sentences is usually enough: "Before you draft customer replies, read replies.md and follow it exactly. Ask me about anything it doesn't cover."

Then let the skill do the delegation. "Handle new Google reviews per reviews.md — draft only, show me before anything posts" delegates the whole routine, not one message at a time. The same move works for a person: a new hire who reads "how we do things" on day one asks better questions by day three.

Two honest boundaries

Size isn't the point. Figma has roughly 60–70 PMs using this. A solo operator with five files gets the same benefit — the mechanism (documented, referenced, edited) doesn't require a team.

None of this is measured. Mechaber says he tried to build a benchmark and couldn't; he doesn't know how much the setup outperforms a vanilla agent. The honest case is operational: documented instructions beat instructions you keep re-pasting — because you stop losing them, and because your team inherits them. If that isn't worth an evening of writing, no benchmark will convince you.

The 30-minute version tonight

Pick your highest-friction recurring task — the one where the AI's output usually needs fixing.
Write the four parts: when it applies, what to do, how you sound, what to reference.
Save it as a file in one shared place. Name it clearly.
Point your AI at it next time the task comes up: "Read [the file] and follow it exactly."
When the output disappoints, fix the file — not the chat. That edit is the whole system compounding.

Want a head start? The Boring Automation Pack from SMB Scale Up includes plug-and-play automation blueprints plus the standing-instructions worksheet — the boring half of this setup, pre-built.

Day Spas: The Quiet AI Automation Wins Hiding in Plain Sight

T.M. Gunderson — Thu, 05 Nov 2026 05:05:31 +0000

Day spas run on rhythm. The same treatments, the same rooms, the same staff shifts — week after week. That predictability is why AI automation fits here better than in most businesses. Not because spas are futuristic, but because the work is patterned.

This post covers three workflows where small day spas are quietly using AI to reduce no-shows, fill empty rooms, and bring clients back without hiring a marketing person. No hype, no big-tech budget required.

1. Pre-Appointment Intake: Stop the Paperwork Pile

Most spas still hand clients a clipboard on arrival. The delay is not just the five minutes of writing — it is the back-and-forth when a contraindication shows up after the therapist has already started.

What some spas are doing instead:

An AI-powered intake form sent 24 hours before the appointment
Conditional logic: if a client marks "pregnancy" or "recent surgery," the system flags the right therapist and adjusts the treatment plan automatically
A simple summary card that lands in the therapist inbox 10 minutes before the session

The honest boundary: this works for common conditions. Rare medical cases still need a human review. The win is not replacing judgment — it is surfacing the routine stuff early so the therapist walks in prepared.

Tool path that does not require IT: Most booking platforms (Square Appointments, Acuity, Vagaro) already support conditional intake forms. The AI layer is a simple prompt that reads the form responses and writes a one-line summary. No custom code needed.

2. Room-and-Staff Matching: The Tetris Problem

A four-room spa with five therapists and twelve service types is a daily scheduling puzzle. One mismatch — a deep-tissue therapist in a room without the right table, or a facial booked in a room without a sink — and the whole day unravels.

The automation angle:

A daily run that matches tomorrow bookings to room requirements (table type, equipment, ambient temperature preferences)
Flagging gaps: if three afternoon slots are empty but one room is overbooked, move a portable service to the underused room
Staff skill-to-service matching: ensuring the therapist certified in hot stone is not accidentally assigned a Swedish massage

The win is not a perfect algorithm. It is catching the two or three obvious mismatches before the day starts. Studies commonly report that service businesses lose 15–20% of productive capacity to scheduling friction — not no-shows, just poor room utilization.

Tool path: A simple spreadsheet or Airtable base with room attributes, updated once a week, plus a weekly AI-assisted audit of the booking-to-room fit. The automation is not the scheduler — it is the sanity check.

3. Post-Visit Rebooking: The 21-Day Window

Spa clients who rebook within three weeks of their last visit have dramatically higher lifetime value. But most spas rely on the therapist verbal "see you in a month?" which clients forget the moment they leave.

What works:

A personalized rebooking message sent 14–18 days after the visit — not a generic blast, but one that references the specific treatment
A short link to the booking page with the same therapist and treatment type pre-selected
For package clients: a reminder when they have two sessions remaining, not when they are at zero

The honest boundary: aggressive rebooking pushes annoy clients. One message per visit, timed to when the treatment benefits are naturally fading, tends to perform better than weekly newsletters.

Tool path: Most email platforms (Mailchimp, Constant Contact, ActiveCampaign) support date-based triggers. The AI layer is writing the message copy — referencing the specific treatment and therapist by name — rather than a generic "We miss you!"

Why Day Spas Specifically?

Day spas are an interesting automation case because they are not emergency-driven like HVAC or compliance-heavy like dental. The value is not speed — it is smoothness. A client who books without calling, fills out forms without confusion, and gets a timely rebooking reminder is a client who stays for years.

The technology is already here. Most spas already have booking software, email tools, and spreadsheets. The gap is usually a small workflow bridge — an intake summary, a room audit, a rebooking nudge — not a $5,000 platform replacement.

A Practical Starting Point

If you run or advise a day spa, here is the 30-minute audit:

Intake: Are clients still filling out paper forms on arrival? If yes, move the form online and add a simple AI summary for therapists.
Rooms: Do you ever discover a room mismatch after the client has arrived? If yes, run a weekly room-to-service matching check.
Rebooking: Do most clients leave without a next appointment booked? If yes, add a 14-day personalized rebooking trigger.

Pick one. The tools are cheap or free. The payoff is fewer interruptions, fuller rooms, and clients who actually come back.

If you are building AI workflows for small businesses, the Boring Automation Pack includes ready-to-use templates for intake automation, rebooking sequences, and service matching — written for non-technical owners who just want the thing to work.

5 AI Automations for Small Hotels and B&Bs That Actually Save Hours

T.M. Gunderson — Tue, 03 Nov 2026 09:01:09 +0000

5 AI Automations for Small Hotels and B&Bs That Actually Save Hours

Running a small hotel or bed-and-breakfast means being the front desk, the reservationist, the concierge, the housekeeper, and the bookkeeper — sometimes all before lunch. Guests expect instant communication, seamless check-ins, and personalized service. But when you're a team of two (or it's just you), the repetitive tasks pile up fast.

The good news: AI agents can now handle the operational busywork that used to require a full front-desk staff. Not as a replacement for hospitality, but as a first-line filter that frees you up for the human moments guests actually remember.

Here are five automations that make sense for small lodging operators right now. Each is something you can set up without a developer, and each gives you back time for the work that moves revenue.

1. Instant Guest Messaging That Doesn't Sound Robotic

The problem: Guests message at odd hours asking about parking, late check-in, or nearby restaurants. You wake up to a queue of unanswered texts, or you're interrupted during turnover.

The workflow:

Connect an AI agent to your messaging platform (SMS, WhatsApp, or your booking-channel inbox)
Feed it your property guide: parking instructions, Wi-Fi password, check-in window, local recommendations
Agent answers common questions instantly, in your property's tone
Escalates only for complaints, special requests, or refund inquiries

What changes: A guest who messages at midnight gets an answer in seconds instead of silence until morning. You keep control over tone and boundaries; the agent just handles the repeat questions.

2. Dynamic Pricing Without the Spreadsheet Obsession

The problem: Your rates are either static (leaving money on the table during events) or manually adjusted (consuming hours you don't have).

The workflow:

Agent monitors local event calendars, seasonal patterns, and competitor rates from your OTA dashboard
Once daily, it suggests rate adjustments within boundaries you set (e.g., "never below $89, never above $249")
You approve with one reply; the agent updates your booking system

What changes: You go from set-it-and-forget-it to responsive pricing — capturing event surges and shoulder-season demand — without the hourly rate-checking habit.

3. Review Response Automation That Sounds Like You

The problem: Positive reviews go unthanked. Negative reviews fester. Both hurt search ranking and future bookings.

The workflow:

Agent monitors new reviews across Google, Booking.com, Airbnb, and TripAdvisor
Drafts a personalized thank-you for 4-5 star reviews, referencing specific details the guest mentioned
Flags 1-3 star reviews with suggested response tones (empathetic, corrective, or "let's talk offline")
You review and approve before posting

What changes: Your response rate jumps from 20% to near-complete. Future guests see a host who listens, and review platforms reward engagement with better visibility.

4. Housekeeping Coordination That Doesn't Rely on Group Chat

The problem: Turnover days mean frantic back-and-forth about which rooms are ready, which need deep cleaning, and where supplies are low.

The workflow:

Agent reads check-out/check-in data from your property management system or calendar
Generates a daily housekeeping board: room number, checkout time, next check-in time, special notes (pet-friendly, extra towels)
Sends to your cleaner's preferred channel (SMS, WhatsApp, or email) each morning
Collects "ready" confirmations and flags any delays

What changes: No more "Is Room 3 done yet?" texts. Your cleaner gets a clear list; you get visibility without micromanaging.

5. Pre-Arrival Guest Onboarding That Reduces Check-In Friction

The problem: Guests arrive unsure where to park, how to get in, or what the Wi-Fi password is. You spend the first 10 minutes of every stay repeating the same information.

The workflow:

Agent triggers a pre-arrival message 24 hours before check-in
Includes: directions, parking, access code or lockbox location, Wi-Fi credentials, and one local recommendation
Collects estimated arrival time to coordinate with your schedule
Flags any accessibility needs or special requests you might have missed

What changes: Guests arrive informed and confident. You spend less time on logistics and more on hospitality.

A Reality Check on ROI

These aren't magic. An AI agent is only as good as the instructions you give it and the property data you feed it. Start with one automation — guest messaging is usually the highest-impact first step — run it for two weeks, measure the time saved, then add the next.

Studies commonly report that small lodging operators spend 30-50% of their week on repetitive admin and guest communication. Even a 20% reduction in that load is several hours back — hours you can spend on guest experience, maintenance, or simply not working.

Want the Playbook?

I've been documenting how small business owners can build these automations step-by-step — without hiring a developer. The AI Agent Owner's Playbook covers agent setup, prompt templates, and the exact workflows that work for operations-heavy businesses.

If you want a faster start, the Boring Automation Pack includes pre-built agent configs for guest messaging, review responses, and daily coordination tasks.

What's your biggest time sink as a lodging operator? Drop it in the comments — I'll suggest a specific automation for it.

Three Background Agents Now Compete to Initiate Your Work: Run the Delegation Audit First

T.M. Gunderson — Tue, 03 Nov 2026 09:00:39 +0000

The category is no longer optional

Three of the biggest AI vendors now ship roughly the same thing: a background agent that initiates work instead of waiting to be asked.

Meta announced Muse first, describing it as a personal agent that carries out long-running tasks and keeps working in the background. OpenAI's Dots can text you proactively when something you asked it to watch — a product's price, a travel itinerary, a deadline — changes. And on October 21, Anthropic opened a research preview of Claude Duet: a proactive agent built around Opus 5.5, with access to the over 6,000 connectors in the Claude Marketplace, a morning brief of what needs attention, and a rule that it checks with you before sending a message, making a payment, signing anything, or deleting anything.

The competitive coverage treats this as a horse race — who answered whom, which vendor is ahead. That's the wrong read for a business owner. The real change: agents that come to you are now table stakes, so the interesting decision is no longer whether to use one. It's which of your recurring tasks should hand off — and that's a decision to make before you enable anything, not after.

The chat-agent question was "can I trust it with a task I assign?" The background-agent question is different and harder: "which of my tasks should an always-on agent own?" Here's the audit to run first.

Step 1 — List what hands off cleanly: trigger-checkable work

Not every recurring task is delegable, even to a competent agent. The test is the trigger. A task hands off cleanly when all four are true:

It recurs on a pattern, not a deadline. Invoice chasing, review responses, appointment follow-ups, price-drop watches. Not "redo the website."
The trigger lives in a connected system. The agent's superpower is checking accounts on a schedule — Duet checks connected accounts every few hours, Dots watches what you hand it. If the trigger is visible in email, a bookkeeping app, a CRM, or a review site, the agent can see it start. If the trigger lives in a phone call or a hallway conversation, it can't.
A wrong answer is cheap to catch. A draft reminder you read before it goes costs nothing. A misfiled tax election doesn't.
Nobody at the business enjoys doing it. Low morale, high frequency — that's the sweet spot.

Anthropic's own task list for Duet is a useful map of what passes this test: finding missed Gmail and Slack messages and helping reply, checking QuickBooks or Stripe for unpaid invoices and drafting reminders, drafting follow-ups from HubSpot deal history. Notice what those have in common — read, draft, prepare, then stop for approval. That's the shape of a first safe delegation.

Step 2 — Write the never-hand-off line before the agent offers to cross it

Background agents can do more than read-and-draft now — Duet can prepare payroll in Gusto and, with your approval, pay with a virtual card through Link. That's the point of the product. It's also why the boundary has to be written down first, while nothing is connected.

The line worth keeping: money moving and promises made stay with a human. An agent can draft the payment; a person releases it. An agent can draft the client reply; a person sends it. Duet's own design concedes this — before it sends, pays, signs, or deletes, it shows you what it's about to do and waits. A few things are off-limits even with approval, like trading securities or moving money between accounts.

Vendors building the boundary in is good. It's not the same as you deciding where your boundary sits. Write the list before the first connection, because the moment an agent surfaces a half-finished tempting action is the wrong moment to invent policy.

Step 3 — Set an interruption budget before week two

Here's the failure mode nobody covers in the launch posts: the agent that interrupts more than it saves.

A background agent's whole mechanism is interrupting you — that's what "proactive" means. A Dots agent texting about a price drop you no longer care about, a Duet ping about an email you'd have ignored anyway — each one is individually reasonable and collectively a second inbox. Anthropic's announcement is unusually candid about the tradeoff: Duet checks your connected accounts every few hours because checking nonstop would use up your plan's limits much faster, and testers are told "it still might misread a thread or miscalibrate what's urgent."

So treat notifications like a budget. Rule of thumb: if a delegated task generates more than one interruption a day, the task wasn't delegable — or the notification threshold is wrong. Fix the threshold once, and if it pings twice, revoke the task. The task list should shrink toward the few that pay for themselves, not accumulate. A weekly review of what the agent actually surfaced — not what it says it did — is the companion discipline once things are running.

Step 4 — One owner per surface, or three agents will fight for your background

If you trial all three (don't — see boundaries below), the practical rule: never point two agents at the same app. Muse watching your email, Dots watching your email, and Duet drafting from your email is three sets of credentials, three notification streams, and a real chance two of them act on the same thread. One app, one agent, one owner — noted in the same place you wrote the never-hand-off list.

This sounds obvious and it's the first thing that breaks, because each agent's setup flow encourages connecting your whole stack on day one. Connect only the app that feeds the one task you're trialing.

The 30-minute version, tonight

List five recurring tasks with triggers in connected systems (Step 1's four tests — most owners find two of five pass).
Write the two lines: the never-hand-off line, and the one-owner-per-app rule.
Pick one task on one agent, on a two-week trial with an interruption budget of one ping a day.
Diary the revoke date. If at two weeks the interruptions outweigh the catches, disconnect the task — that's a successful audit, not a failed experiment.

Honest boundaries

Everything above is a research preview or a first release. Duet is waitlist-only, US, on paid Claude plans, and shares usage limits with the rest of the product — background checks count against your plan. Muse and Dots are consumer-first personal assistants; Duet is deliberately work-focused. Expect churn in all three.
These are subscription products with usage costs, not one-time purchases. A background agent that checks four apps every few hours, every day, is a line item that grows. Price the trial accordingly.
Don't rebuild your business around the tool. Vendors will iterate, merge, and rename these products inside a year. The audit — what recurs, what never hands off, what the interruption budget is — transfers to whichever agent wins. The tool choices don't.
The audit is pre-delegation. It decides what starts. It's the pair of, not a replacement for, a weekly review of what's already running.

Want the delegation worksheets? The Boring Automation Pack includes the recurring-task inventory and the permission-line worksheet — the boring half of this audit, pre-built.

Gutter Cleaning Season Is a Sprint, Not a Marathon — AI Can Handle the Dispatch

T.M. Gunderson — Tue, 03 Nov 2026 06:16:25 +0000

The ladder-and-leaves calendar

A gutter cleaning business lives and dies by two signals: the first heavy rain of fall, and the first freeze of winter. Between those points, every crew is maxed out. Every missed call is a booking that went to the competitor who answered first. Every route planned by gut instead of data is a gallon of gas and an hour of daylight wasted.

The margins are thin, the season is short, and the work is inherently local and physical. AI won't climb the ladder. But it can strip out the scheduling chaos, the reminder busywork, and the upsell opportunities that slip through the cracks when everyone's in the truck.

Here are three workflows that fit gutter cleaning operations right now.

1. Weather-Triggered Outreach That Books Before the Storm

The problem: Most homeowners don't think about gutters until water is pouring over the edge. By then, every crew in town is booked two weeks out.

The workflow:

An AI agent monitors local weather feeds for specific triggers: sustained leaf-drop conditions (wind + temperature drops), first freeze warnings, or heavy-rain forecasts 48–72 hours out.
It drafts a short, personalized SMS or email to past customers in the affected zip codes: "Forecast says 2 inches of rain Thursday. Want us to check your gutters before it hits? Reply YES for priority scheduling."
Confirmed replies flow straight into the dispatch board with a geo-tagged slot.

What changes: You shift from reactive (phone rings after the damage) to proactive (booked before the storm). Studies commonly report that service businesses using weather-triggered outreach see 15–25% higher booking rates in peak season — but the simpler win is that you're not competing for the same panicked caller as every other crew.

Honest boundary: This only works if you have a clean customer list with service history and zip codes. If your records are in a shoebox, start there.

2. Route Density Over Speed: The AI Dispatch Board

The problem: A crew drives 20 minutes between jobs, backtracks across town twice, and eats lunch in a parking lot because the next job is 35 minutes away. The day feels busy. The revenue-per-hour says otherwise.

The workflow:

Feed the AI your job list each morning: addresses, estimated durations, ladder-access requirements (single-story vs. multi), and any time windows the customer requested.
The agent optimizes for density, not just shortest-path: cluster jobs by neighborhood, respect time windows, and flag same-street or adjacent-block opportunities for same-day add-ons.
It outputs a sequenced route with drive-time estimates and a "nearby prospects" list — customers who had estimates but didn't book, located within a few blocks of confirmed jobs.

What changes: Less windshield time, more gutter time. A tighter route means one extra job per day in peak season — which, over a six-week fall rush, is material revenue.

Honest boundary: The optimization is only as good as the address data and time estimates. If you routinely underestimate how long a three-story colonial takes, the AI will pack the day too tight. Calibrate your durations honestly before you automate the board.

3. Annual Reminders That Book Next Season Before This One Ends

The problem: Most gutter cleaning relationships are transactional. The homeowner books once, pays, and disappears until the next emergency. There's no retention system because there's no time to build one during the rush.

The workflow:

After each job, an AI agent notes the property details (gutter length, problem spots like pine-tree overhangs, whether guards were discussed) and schedules a follow-up.
In late summer — before the rush hits — it sends a personalized reminder referencing the prior visit: "Last October we cleared 40 feet of backed-up gutters on Maple Street and flagged the pine overhang. Want to get ahead of the leaves this year?"
For customers who booked spring cleanings, it offers a fall prep bundle. For one-time customers, it offers a recurring annual plan.

What changes: Recurring revenue in a seasonal business. Even a 20% annual retention lift turns a six-week sprint into a more predictable operation.

Honest boundary: The reminder must reference real details from the actual job. Generic "Time to clean your gutters!" emails read like spam and convert poorly. If your field notes are thin, invest there first.

What this isn't

It isn't a replacement for the person who climbs the ladder. Gutters require eyes, judgment, and a willingness to get dirty. AI handles the desk. The roof stays human.
It isn't a fix for unreliable crews. If your team no-shows or your quality is inconsistent, automation just schedules the disappointment faster.
It isn't a demand generator in the off-season. Most homeowners genuinely don't need gutter work in March. The play is to lock in the fall rush before it starts, not to invent work where none exists.

The 20-minute start

If you want to test whether this fits your operation, don't buy software yet. Do this:

List your last 20 jobs — date, address, drive time from your shop, and how you got the lead (repeat, referral, Google, etc.).
Color-code them on a map. See the clusters and the outliers.
Count how many were within 5 minutes of another job that day.
Write one actual text message you'd feel good sending to a past customer before a forecasted storm.
Check whether your current scheduling tool can send automated messages — most can, even the basic ones.

That map and that message are your automation spec. Everything else is connecting the dots.

Honest boundaries

No first-person claims. This is research and analysis, not lived experience.
No guaranteed results. Route efficiency gains depend on job density and geography.
No tool is free. Even cheap automation has a setup cost in time and attention.

Want the templates? The Boring Automation Pack includes the weather-trigger outreach script, the route-density checklist, and the annual reminder sequence — built for seasonal service businesses, not software companies with flat demand curves.

Anthropic Filed to Go Public Losing $80 Billion in Six Months: Hedge Your AI Bill Now

T.M. Gunderson — Mon, 02 Nov 2026 09:00:10 +0000

The filing, in the parts that matter to a business owner

Anthropic's S-1 went public on October 26 for a Nasdaq listing under the ticker ANT, and the headline numbers got the finance world arguing. Per the filing, as reported by CNBC, Fortune, and The New York Times: $16.3 billion in revenue for the first half of 2026 — up 1,304% from $1.16 billion in the same period a year earlier — against a net loss of $80.7 billion.

Read past the headline before repeating it. Fortune's coverage notes that $72.4 billion of that loss is an accounting charge rather than cash out the door. The more durable facts are the commitments: Fortune reports plans to spend up to $567 billion on compute, including roughly $125.2 billion in five-year lease payments, excluding interest; the Motley Fool's read of the filing counts about $417 billion in future minimum payment obligations, with CNBC noting another $125–150 billion expected for capacity delivered between 2028 and 2031. For all of 2025, the company lost about $41.9 billion on $4.6 billion of revenue, per CNBC.

None of this is a reason to panic, and none of it is a stock opinion. It's a reason to notice what your own business is exposed to — and to do something cheap about it.

Why an SMB should read a cap-table story at all

A frontier lab's balance sheet eventually reaches its customers' invoices. Not this quarter — the point is direction. Three links in the chain:

Compute is the cost floor. Every token you buy sits on top of a lease and a chip that somebody financed.
Obligations are the floor's financing. Commitments measured in hundreds of billions get serviced from somewhere, at scale, continuously.
Pricing is the lever that services them. When a lab goes public, expectations start being marked to market quarterly — and pricing discipline is how a company answers.

For an enterprise, repricing is a negotiation with an account team. For a small business on a card-subscription, it's a line item that changes without anyone inviting you to a meeting. That asymmetry is the whole argument for hedging now, while it costs an evening instead of a crisis.

The honest counterweight: AI prices have also been falling across this market for two years. So treat this as a dependence audit, not a prediction. The hedge below is cheap in a world where prices double and cheap in a world where prices collapse. That's the only kind of bet worth making from incomplete information.

The four-hedge checklist

One evening, four hedges, in order of effort:

Hedge 1 — Lock in annual rates where they're offered. If a vendor you already trust offers a meaningful annual discount, trading a little flexibility for price certainty is the cheapest hedge available — it converts a variable line into a fixed one for twelve months. Two conditions make it a good trade rather than a reflex: you've watched your usage for at least a quarter, and you'd keep the tool even if a competitor launched something shinier next month.

Hedge 2 — Keep a tested fallback per task family. Not "a fallback model" — one per family: writing, coding, analysis. The families have different leaders, so a single substitute quietly fails one of them. For each family, name the alternative, run it against your real work (your trickiest email, your actual codebase, your genuine analysis — not a demo prompt), and write down the date you last tested it. Refresh quarterly. A named-but-untested alternative is a preference, not a hedge.

Hedge 3 — Revisit open weights the moment you have a compliance driver. The own-vs-rent question now has a dated example: Moonshot has said the full weights of Kimi K3.1 Pro — a frontier-adjacent model benchmarking fourth on Artificial Analysis's independent index — ship by November 4. That makes self-hosting a recheckable question with a real number attached, not a stance. The recheck trigger is specific: regulated or confidential data, predictable high volume, or a renewal where a credible alternative buys leverage. No trigger, no project — keep renting and move on. But write the trigger down, so the recheck happens when the trigger fires instead of never.

Hedge 4 — Never build a workflow that breaks if one vendor's price doubles. This is the portability rule, and it's the one that pays off in every scenario. Prompts, instructions, and context live in files you own — not trapped inside a single vendor's memory. Context gets exported on a schedule, not on moving day. The swap to your fallback is documented as a procedure — configuration, environment variables, a checklist — not a rebuild. The test: if moving your most-used task family to your second-choice model would take more than an afternoon, you don't have a fallback. You have a favorite.

What the filing's risk language means for you

The S-1 carries more than 90 pages of risk factors, per Variety's reporting — including, as Fortune and WIRED both quoted, the prospect that increasingly autonomous systems could "take harmful actions independently and at scale without human intervention," and language The Verge flagged about potential harms that couldn't be underwritten by the entire insurance industry in aggregate.

You cannot hedge existential risk with a checklist, and this piece won't pretend otherwise. But the business-continuity translation is mundane: if a frontier vendor's service degrades, pauses, or reprices — for any reason from an outage to a policy change to an actual incident — your business keeps running on the same boring infrastructure: documented fallbacks, exported data, no single points of failure. The catastrophic-risk section of the filing and the price-risk section of this post end at the same answer.

Honest boundaries

The big loss number is not cash burn. Most of the $80.7 billion is an accounting charge. Quote the obligations — the hundreds of billions in payment commitments — if you want the durable fact, not the headline.
Nobody knows when repricing lands. This is insurance, not a forecast. The same market that produced these obligations also produced two years of falling prices. The hedge is cheap in both worlds; that's the only claim being made.
Annual lock-in cuts both ways. Locked in at a high rate in a falling market is a loss too. Lock in what you'd keep anyway.
This isn't IPO analysis. Whether the stock is a buy is a different question with different inputs. The only question here is what your own business does on Tuesday.
Hedging is not hoarding. A drawer of untested AI subscriptions is a cost problem wearing a risk-management costume. One tested alternative per task family. That's the whole prescription.

Tonight's version

Fifteen minutes, four lines on one page:

Which of your current subscriptions offer an annual rate — and what the discount actually is.
The named fallback for each of your three task families, with the date you last tested it.
Whether a compliance driver is live. If no: write "rent, recheck on trigger" and stop thinking about it.
How long a swap of your most-used tool would take today — answered honestly.

If line four's answer is "I don't know," that's not a failure. That's the finding, and it's the first thing to fix.

Want the worksheets? The Boring Automation Pack includes the vendor-dependency audit and the quarterly fallback-test worksheet — the boring half of this checklist, pre-built.

The Inventory Problem Party Rental Owners Know Too Well

T.M. Gunderson — Mon, 02 Nov 2026 05:03:56 +0000

The seasonality trap

A party rental business runs on two calendars: the one your customers see (weekends, graduations, weddings) and the one you live inside (tent-cleaning Tuesdays, chair-repair weeks, the annual inventory audit that always slips to January).

Between May and September, the phone doesn't stop. Between October and April, you spend half your energy chasing the next booking. The work that actually determines profit — pricing by event size, routing deliveries efficiently, confirming counts the day before — gets buried under the volume.

AI doesn't change the calendar. It changes what you have to touch personally.

The three workflows that actually matter

1. Instant quotes that don't cost you a phone call

Most rental inquiries arrive with the same five variables: event type, guest count, date, location, tent or no tent. An AI agent that reads the incoming form or email, checks real-time availability against your inventory sheet, and returns a written quote with delivery windows — that single workflow removes the 15-minute back-and-forth that happens 20 times a weekend in peak season.

The honest boundary: AI can price standard packages accurately. It cannot judge a backyard slope from a photo, or a venue's load-in restrictions. The quote needs a human-review gate for anything non-standard — and that gate should be the exception, not the default.

2. The week-before confirmation pass

Every rental owner has a story about the event where the chair count was off by 20, or the delivery truck showed up to a locked gate. An agent that runs the confirmation checklist three days out — verifying counts, sending venue contact info to the driver, flagging any order that changed since booking — turns the frantic Sunday-evening phone marathon into a five-minute review.

The tool exists in plain sight: any scheduling platform with a webhook and a language model can run this. The hard part is writing the checklist once, not building the automation.

3. Seasonal pricing that actually reflects reality

Most rental businesses price by gut — last year's number, bumped 5% if things felt tight. An AI that reads your actual booking density, labor cost per event, and inventory depreciation can recommend price adjustments per package with the math visible. You still decide. The difference is you're deciding from data instead of memory.

Studies commonly report that small businesses using demand-based pricing see 10–15% revenue lift in seasonal industries — but the real win is simpler: you stop underpricing your busiest Saturdays because you set rates in February.

What this isn't

It isn't a replacement for the human who loads the truck. Delivery day has too many variables — weather, access, the customer who swears the driveway is "totally flat."
It isn't a fix for bad inventory records. AI reads what you give it. If your chair count in the spreadsheet is off by 30, the quote will be wrong with confidence.
It isn't a silver bullet for the slow season. Automation makes peak season survivable. It doesn't create demand in December.

The 20-minute start

If you want to test whether this fits your operation, don't build anything. Do this:

List the last 10 inquiries you received — email, form, or text.
Count how many of them asked for a variation of a package you already offer.
Time how long you spent on the back-and-forth for each.
Pick the one question you answer most often ("Do you have the 20x40 tent available August 12?").
Write the answer once, in your actual voice, with the conditions included.

That fifth item is your first automation spec. Everything else is plumbing.

Honest boundaries

No first-person claims. This is research and analysis, not lived experience.
No guaranteed results. Pricing lifts vary by market, competition, and execution.
No tool is free. Even "cheap" AI has a time cost in setup, testing, and maintenance.

Want the worksheets? The Boring Automation Pack includes the quote-automation spec template and the week-before confirmation checklist — the boring half of these workflows, pre-built.

Black October Hit Small Business Too: The 30-Minute Account Audit

T.M. Gunderson — Sun, 01 Nov 2026 09:00:11 +0000

A bad week got a name

Between Oct 19 and 22, 2026, a cluster of security disclosures landed close enough together that the industry gave the period a name: Black October. It wasn't one coordinated campaign. It was a flaw in the PHP interpreter (PHP2Shell) that could let unauthenticated attackers run code on vulnerable web servers — researchers demonstrated it against WordPress, Drupal, and Joomla, and the NSA issued an urgent patching advisory. It was a software supply-chain compromise of LiteLLM, an open-source gateway businesses use to route requests between AI providers, whose malicious release was built to steal browser credentials, cloud service secrets, developer tokens, and API keys used by AI services. And it was the recovery of weak certificate-authority keys that forced emergency responses from browser vendors.

Two numbers from the aftermath deserve an owner's attention:

27 seconds. On the No Priors podcast, CrowdStrike CEO George Kurtz described the fastest "breakout time" the company has observed — the gap between a first compromised machine and movement toward the systems that matter. CrowdStrike's threat reporting puts the average at 29 minutes, down from 48 minutes; the fastest observed case is 27 seconds.
~2,900 login attempts per minute. Microsoft Threat Intelligence recorded one October password-spraying campaign at more than 151,000 failed attempts against 327 accounts — the work of a group Microsoft tracks as Storm-0940, which it says used an open-source AI-powered penetration-testing framework. Honest footnote: Microsoft observed no successful logins in that specific campaign. Password spraying mostly fails. But it runs cheap, forever, against every account that exists — including the ones you've forgotten.

And the most small-business-flavored detail of the month: Fox News reported a startup founder waking up to roughly 50 billing notifications and $449 in OpenAI and Anthropic charges he didn't recognize, after stolen API keys were exploited during the wave. That's the modern version of finding strange charges on the company card — except the card was an AI account with an effectively unbounded spending limit.

None of this requires you to be a target. The LiteLLM malware harvested whatever credentials it found on the machines it landed on. Password sprayers work through whatever logins exist. The wave doesn't aim; it drifts. What determines your damage is what was lying around reachable.

The 30-minute account audit

One sitting. No security vendor, no new purchases. The goal is not to stop a breach — it's to shrink the blast radius, so that any single stolen credential is worth as little as possible.

1. Inventory every account a person OR an AI tool touches (10 min).
List everything with a login: email, banking, accounting, domain registrar, social media, team tools. Then the part everyone misses: the accounts your AI tools hold. Every assistant connected to your inbox, every agent with an API key, every automation with a saved password. These count double — they store credentials and carry sellable keys of their own. The October supply-chain attack went after exactly this layer.

2. Rotate anything stale (5 min).
Any password carried over from a previous job, a previous decade, or a vendor's default. Any API key created once and never rotated — the kind the LiteLLM malware was built to find. Modern practice is a named key per tool rather than one master key, so revoking one doesn't mean revoking everything.

3. Turn on MFA where it's missing (5 min).
Password spraying is defeated by the cheapest control that exists. Work the inventory from step 1 top-down: money first, email second (email resets everything else), everything else third.

4. Sweep your footprint (5 min).
Old accounts on services you stopped paying for, forgotten landing pages, test accounts, the personal email used on business tools years ago. Anything with a login that no one watches is a door with no one behind it.

5. Check what has standing access to money (5 min).
Confirm no AI tool or automation holds an unattended connection to anything that moves funds — payments, invoicing, banking. The $449 founder's problem wasn't that his key was stolen; it's that the stolen key could spend without asking. A key that can't spend is a nuisance. A key that can is an incident.

What this drill honestly is and isn't

This is hygiene, not defense. It will not stop a supply-chain compromise upstream of you, a zero-day in your web platform, or a determined attacker inside your walls. What it does is remove the easy paths: the forgotten login, the never-rotated key, the AI assistant with standing access to your payment stack. When the average attacker breakout takes 29 minutes and the fastest takes 27 seconds, response time is not a small business's strength — low value per breach is. The audit is how you lower it.

The cadence that makes it stick: run the full drill quarterly, and rerun just step 5 whenever you connect a new AI tool to anything.

Want the worksheets? The Boring Automation Pack includes the account-inventory and access-audit sheets that walk this drill end to end.

How Tax Prep Firms Can Automate the Grind Before Season Starts

T.M. Gunderson — Sun, 01 Nov 2026 05:08:19 +0000

The seasonal trap

Most tax preparation firms look like a gym in January: packed for three months, quiet for nine. The revenue is concentrated, the pressure is concentrated, and the mistakes are too.

The American Institute of CPAs notes that tax practitioners commonly work 60–80 hour weeks during peak season. Studies also suggest error rates on manually prepared returns can run 10–20% on complex filings. The problem isn't competence — it's volume against a fixed clock.

Automation won't replace a tax professional's judgment on a client's multi-state rental property or a business's R&D credit. But it can strip away the repetitive intake, document chasing, and data-checking that currently eat the first half of every engagement.

Here are five patterns tax prep firms are using to reclaim 15+ hours weekly before the next filing wave hits.

Pattern 1: Automated Document Intake & Sorting

The problem:
Clients send W-2s, 1099s, K-1s, brokerage statements, and handwritten notes in every format — email attachments, photos, portal uploads, and fax (yes, still fax). Someone has to open each one, name it correctly, and route it to the right folder.

Time drain: 6–10 hours/week during intake season

Automation approach:

AI-powered document classification that reads incoming files and sorts by form type (W-2, 1099-NEC, 1099-INT, etc.)
Automatic client-folder routing so documents land in the right engagement file
Duplicate detection to catch when the same document arrives twice
Missing-document flags that compare what's received against a standard intake checklist per client type

Time savings: 4–6 hours/week

Honest boundary: AI sorts well when documents are legible and standard. Handwritten notes, foreign-language documents, and heavily redacted files still need human eyes. Build a "needs review" queue for exceptions rather than forcing everything through.

Pattern 2: Client Communication Triage & Follow-Up

The problem:
The same questions arrive every season: "Did you get my documents?" "What's the status of my return?" "Can I deduct this?" Each reply takes 3–5 minutes, and firms see dozens per day. Meanwhile, the clients who still haven't sent their organizer get buried under the noise.

Time drain: 4–6 hours/week on status and intake follow-up

Automation approach:

An AI-assisted inbox that drafts status replies using the actual engagement progress (preparer name, documents received, expected completion window)
Automated chase sequences for missing documents — polite, timed reminders with a one-click upload link
FAQ handling for the 80% of questions that have the same answer every year (deadlines, extension procedures, standard deduction amounts)
Handoff rules that escalate tax-advice questions to the preparer while routing logistics questions automatically

Time savings: 3–4 hours/week

Honest boundary: Client trust is the asset. Automation should feel like a responsive assistant, not a chatbot deflection. Every automated reply should include a clear path to a human. Don't let AI answer "Can I deduct this?" without preparer review — the risk profile is wrong.

Pattern 3: Data Extraction from Standard Forms

The problem:
Keying numbers from W-2s and 1099s into tax software is precise, tedious, and error-prone. A transposed digit on a Social Security number or an employer EIN triggers rejections and delays.

Time drain: 3–5 hours/week per preparer during peak

Automation approach:

OCR extraction with field-level validation (box numbers matched to expected values)
Cross-check totals: does the sum of box 1 across all W-2s match the client's stated wage income?
Anomaly flags for unusual values (a 1099-NEC in the millions for a sole proprietor who reported $40K last year)
Confidence scoring: high-confidence extractions flow through; low-confidence extractions pause for human verification

Time savings: 2–3 hours/week

Honest boundary: AI extraction accuracy on clean, standard forms is typically above 95%. On damaged scans, non-standard layouts, or handwritten annotations, accuracy drops sharply. The safe practice is extraction-plus-review, not extraction-only.

Pattern 4: Scheduler Intelligence & Capacity Balancing

The problem:
Complex returns take longer. Simple returns can be batched. But appointment books often treat every slot the same, leading to either idle preparers or overloaded calendars — sometimes both in the same week.

Time drain: Not measured in most firms, but 2–3 hours/week of calendar shuffling and rebooking

Automation approach:

Return-type tagging that estimates complexity (single W-2 vs. multi-state business owner with K-1s)
Dynamic scheduling that reserves longer blocks for known complex engagements and shorter blocks for simple returns
Buffer management: automatic hold time before and after first-time clients or engagements flagged as high-complexity
Waitlist activation when cancellations open up, prioritized by deadline proximity

Time savings: 2 hours/week of administrative calendar work

Honest boundary: Client relationships don't fit a complexity score. A long-time client with a simple return this year might need the conversation time anyway. Use scheduler intelligence as a starting point, not a cage.

Pattern 5: Post-Filing Client Retention & Year-Round Touchpoints

The problem:
Most tax prep relationships go dark after April 15. The firm becomes a once-a-year vendor, not a year-round advisor. When the client gets a notice from the IRS in August, they've forgotten who prepared the return.

Time drain: None directly — this is lost revenue, not lost time

Automation approach:

Automated quarterly touchpoints (estimated tax reminders, mid-year withholding check-ins, deadline alerts for business owners)
Personalized outreach based on the prior year's return ("You had a significant capital gain last year — consider reviewing your cost basis records before year-end")
Notice-response triage: when an IRS notice arrives, an AI-assisted first pass identifies the notice type, the likely cause, and whether it's a standard information request or something requiring preparer attention

Time savings: Not about time reclaimed — about relationships maintained. Firms that stay in touch year-round see higher retention and higher lifetime value per client.

Honest boundary: Year-round automation only works if it adds value. Generic newsletters get ignored. The touchpoint must reference something specific to that client's situation, or it reads like spam.

What to automate first: a 15-minute priority sort

Before buying any tool, run this exercise with your team:

List your top 10 weekly time sinks during last season — the tasks that made people stay late.
Score each on two axes: repetition (does this happen the same way every time?) and consequence of error (is a mistake here expensive or embarrassing?).
Automate high-repetition, low-consequence tasks first. Document intake, status replies, and form extraction fit here. Complex tax-position judgments and client advising do not.
Leave a human checkpoint on everything that touches the return. Automation should feed the preparer, not replace the preparer.

The goal isn't a firm run by robots. It's a firm where the robots handle the paperwork so the people can handle the judgment.

The honest cost picture

Automation tools for tax prep firms typically run $50–200/month per seat for document extraction and workflow platforms. The break-even math is usually one reclaimed evening per week per preparer — which most firms hit within the first month of implementation.

The hidden cost is setup: training the system on your document types, building your intake checklists, and defining your escalation rules. Budget a day or two before season starts, not the week filings are due.

Next step: If you're prepping for the upcoming tax season, the Boring Automation Pack includes templates for document-intake workflows, client-follow-up sequences, and scheduler rules — built for service businesses with seasonal spikes, not SaaS companies with flat demand curves.

Your AI Stack Has Two Lanes Now: Move the Boring Work to the Cheap One

T.M. Gunderson — Sat, 31 Oct 2026 09:00:27 +0000

The cheap lane got real

DeepSeek's V4.1 Flash — a 552B-parameter model with a one-million-token context window and MIT-licensed weights — lists at $0.15 per million input tokens and $0.60 per million output off-peak. Those rates double during peak hours; cache hits cost fractions of a cent. Independent benchmarking by Artificial Analysis scores the current checkpoint in the mid-40s on its Intelligence Index, tied for second among open-weights models — a legitimate good-enough tier for most work, though about a dozen points behind the frontier leaders on the same index.

Run the transparent math on a genuinely heavy day of agent-assisted coding — say 10 million input tokens and half a million output. At the off-peak rates that costs about $1.80. The same tokens on a frontier flagship at list price ($4/$20 per million) cost about $50. That is not a rounding error. It's the difference between a line item you notice and one you don't.

The swap is three variables

DeepSeek publishes an Anthropic-compatible endpoint, and coding agents that speak Anthropic's format — Claude Code is the common one — point at it with environment variables:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API key>
export ANTHROPIC_MODEL=deepseek-flash

DeepSeek's own API docs describe this integration as working directly, no code required, for tools including Claude Code. Three cautions before you treat it as a five-minute job:

Set the model name explicitly. Third-party coverage of DeepSeek's endpoint behavior warns that a request carrying an unmapped Claude model name can get routed to the higher-priced Pro tier. The base URL alone doesn't protect you from the Pro rates — the model override is the other half of the swap.
Verify what's live before you trust it. Shell exports, settings files, and per-project configs override each other. Your agent has a status check — run it and confirm which endpoint and model are actually active before you bill a day against the wrong assumption.
Off-peak means off-peak. The advertised $0.15 doubles at peak. Still cheap at $0.30, but budget the doubled rate, not the headline one.

The two-lane rule

Not all work moves. Route by stakes, not by habit.

Cheap lane — high volume, low stakes:

Boilerplate, scaffolding, standard CRUD
Data cleanup and format conversion
First-draft scripts and one-off internal utilities
Summarization, classification, sorting, extraction

Frontier lane — keep paying for:

Edge-case-heavy logic where failure is expensive and silent
Anything reliability-critical: billing math, security, customer-facing production paths
Work you can't cheaply verify — the whole cheap-lane bet rests on checking the output being faster than writing it

The test is one question: if this output is wrong, do you find out cheaply or expensively? If the answer is cheap — the test suite passes, the diff reads fine, you eyeballed the spreadsheet — then price matters more than pedigree. If the answer is expensive — you find out when a customer does — you were never buying tokens; you were buying insurance.

When paying the premium still wins

The task is 5% of your volume and 95% of your risk. Aggregate savings come from moving the high-volume families, not the scary ones.
The cheap model fails on your inputs. Long, gnarly legacy code and unusual frameworks are where a dozen index points show up. Run the lane test on your actual work, not on demos.
The savings are smaller than the switching time. If the task runs once a month for twenty minutes, you don't have a cost problem — don't spend an evening optimizing a $0.30 problem.

That last bullet is the honest boundary for this whole idea. The two-lane rule pays when the boring lane is high-volume. A business with three hours a week of AI-assisted work should stop reading pricing posts and go do something else with the evening.

Tonight's version

Pick one task family — not the most important one, the most repetitive one.
Point a spare agent session at the cheap lane: three variables, one status check.
Run the next week of that task on it, checking outputs the way you already check them.
If nothing breaks, leave it there and stop reading pricing news. If something breaks, you've located your lane boundary — which is worth more than the savings.

Want the worksheets? The Boring Automation Pack includes the model-tier audit worksheet — the quarterly review that tells you which task families belong in which lane as prices keep moving.

Your AI Agent Forgets Your Rules After Hour One: The Context-Hygiene Check

T.M. Gunderson — Sat, 31 Oct 2026 05:05:15 +0000

You set up an AI agent with clear rules: "Always cc my assistant on external emails," "Never delete files without asking," "Flag invoices over $500 before paying." It follows them perfectly for the first hour. Then it starts drifting. By hour three, it's drafting replies without the cc line and moving files you didn't tell it to touch.

The agent isn't being disobedient. The rules just got buried.

Why agents forget

AI agents process information in a context window — essentially a rolling transcript of everything they've seen and done. Research from Chroma, a vector-database company, tested 18 frontier models and found a consistent pattern: as input length grows, performance degrades on even simple tasks. The assumption that models process context uniformly doesn't hold. Every tool result, every file opened, every intermediate step piles more tokens onto that transcript. Your standing rules — the ones you set at the start — get pushed deeper into the pile until the agent effectively can't see them anymore.

The symptom isn't dramatic failure. It's quieter: a rule that worked at 9 AM stops working at 11 AM. The agent still "knows" the rule in the sense that it learned it, but the rule has been displaced by newer information in the working window.

Four habits that keep the rules visible

1. Put away used results once they're consumed

When an agent reads a file, generates a draft, or checks your calendar, don't let that result sit in the active context forever. Summarize what it found into a compact note, then close the conversation about it. The full tool output — which might be hundreds or thousands of tokens — is the single biggest cause of rule burial. Keep the conclusion; discard the full transcript.

What this looks like: Instead of letting the agent hold ten files in working memory, ask it to summarize each one in two sentences, then move on. The summaries stay. The files don't.

2. Periodically summarize-and-continue instead of scrolling forever

Long-running tasks are where rules decay fastest. Every few turns, tell the agent: "Summarize where we are, then continue." This resets the active context to the current state, pushing the accumulated noise to the background. It's the equivalent of clearing your desk before the next task instead of stacking papers on top of papers.

What this looks like: On a two-hour project, ask for a one-paragraph status check every 30 minutes. The agent restates the goal, the decisions made so far, and the next step — in about 100 tokens instead of 10,000.

3. Keep standing rules in a file outside the chat

Don't rely on the initial prompt to carry your rules for the entire session. Put them in a file the agent can reference. When the rules are external, the agent can re-read them on demand instead of hoping they're still in active memory. Think of it as a shared handbook rather than something you said once at the beginning of a meeting.

What this looks like: A simple text file with five to ten bullet points: escalation rules, format preferences, off-limits actions, approval thresholds. Reference it explicitly when the agent starts a new phase: "Re-read the rules in agent-guidelines.txt before proceeding."

4. Send big subtasks to a fresh agent with clean context

The cleanest way to prevent rule burial is to not let the context grow in the first place. For large, distinct subtasks — writing a report, analyzing a dataset, drafting a batch of emails — spin up a fresh agent instance with just the task description and relevant rules. That agent starts with a blank slate, does the work, and hands the result back. The parent agent never gets clogged with the intermediate steps.

What this looks like: Agent A handles the overall workflow and orchestration. Agent B handles the report draft. Agent C handles the data cleanup. Each one stays within its lane, and none of them accumulate the full mess of the project.

The 60-second spot-check

Here's the simplest way to know if your agent's rules are still active: mid-task, ask it to recite your standing instructions.

"What are the rules I gave you at the start of this session?"

If it answers accurately, your rules are still in active context. If it summarizes them vaguely, misses key points, or asks you to remind it, the rules are buried and the agent is already drifting. This isn't a gotcha — it's a diagnostic. Check once per long session, or whenever the agent's output starts feeling "off."

When to expect this vs. when to worry

Context decay is normal and expected on long or complex tasks. It's not a bug in the agent software; it's a property of how these models process information. What matters is recognizing it and having habits that compensate.

Expected: A two-hour session where the agent gradually needs more explicit reminders. This is just the cost of doing multi-step work with AI.

Worrying: An agent that ignores a critical safety rule after a single short task. That's not context rot — that's either a badly written rule, a misaligned tool integration, or a task the agent shouldn't be doing at all.

Honest boundaries

These habits help, but they don't eliminate the problem. If your workflow involves dozens of sequential tool calls, complex branching logic, or real-time monitoring over many hours, no amount of summarizing fully solves context decay. In those cases, the honest answer is: you need a different architecture — multiple agents with handoffs, a workflow engine with checkpointing, or a more traditional automation tool that doesn't depend on a single model's working memory.

The four habits above are the low-hanging fruit. Most small-business use cases don't need more than this. Start here, check with the spot-test, and only escalate to heavier infrastructure when you've confirmed the problem is real and recurring.

Practical next step: If you're building workflows with AI agents, the biggest time-saver is having a ready-made playbook of automation recipes — the boring, repetitive tasks that actually eat your day. I put together a pack of 50+ verified automation blueprints for small businesses: https://smbscaleup.gumroad.com/l/boring-automation-pack

Open-Weights Day Is November 4: Redo Your Own-vs-Rent Math (Most of You Will Still Rent)

T.M. Gunderson — Fri, 30 Oct 2026 05:10:50 +0000

The first frontier-class download has a date

Moonshot AI says the full weights of Kimi K3.1 Pro — its flagship, a 3T-class model with a one-million-token context window — ship by November 4. Until then the model runs through the company's apps and API only; after that, in principle, you can run it on hardware you control.

The quality claim is not vendor hype alone. Artificial Analysis, which benchmarks models independently, scores K3.1 Pro at 55 on its Intelligence Index — fourth among all models it tracks, three points behind the leader (Claude Opus 5.5 at 58), one point behind Claude Sonnet 5.5, and nine points above the best open-weights model available today. It lists at $2 per million input tokens and $10 out — the same price as Sonnet 5.5, and a third cheaper than its own predecessor, Kimi K3.

That last detail is why this is a business story rather than a benchmark story. For the first time, "download a frontier-class model and run it yourself" is a question with a date on it. Which makes own-vs-rent a recheckable question — not a stance you adopted once and forgot.

A downloadable model is not a free model

Open weights mean you get the recipe, not the meal. Once the weights are in hand, every cost the vendor was absorbing becomes yours:

Hardware. A 2.8-trillion-parameter model does not run on the office laptop. It needs serious GPUs — bought, or rented by the hour.
Operations. Someone installs it, keeps it running, restarts it when it falls over, and upgrades it when a better version ships. That someone is you, or a contractor you pay.
Security and hygiene. The server is now your server — patching, access control, backups. A model with access to your business data is an attack surface like any other.

When you rent AI, you are borrowing a vendor's ops team amortized across thousands of customers. That is the actual product. The model weights were never the expensive part.

The honest checklist: when self-hosting still does not make sense

Run down this list before you get excited about November 4:

No compliance driver. If no regulator, contract, or client requires your data to stay on your infrastructure, you would be self-hosting to save money that probably is not there.
Unpredictable volume. Self-hosting pays for peak capacity whether or not the peak shows up. Bursty, occasional use is exactly what APIs were built for.
No ops person. Not "you are handy with computers" — an actual human accountable for uptime. If the honest answer is "it would be me, at night," that is the answer.

Most small businesses check all three boxes. That is not a failure — it is the correct result, and knowing it is the point of the exercise.

The three signals it is time to reprice

Self-hosting earns a fresh look when any of these is true:

Regulated or confidential data. Healthcare, legal, finance, or client contracts that demand data residency. Here the comparison is not price — it is what a vendor's breach clause costs you versus keeping the data inside in the first place.
Predictable high volume. Steady, heavy, round-the-clock token use flattens your unit cost and makes capacity planning real. Volume you can forecast is capacity you can buy efficiently.
Renegotiation leverage. A tested, working fallback is the only thing that makes a vendor's renewal desk polite. Even if you never switch, the credible ability to switch reprices the contract you are already in — which matters in a market where frontier AI is being financed on expectations and the bill eventually lands somewhere.

The 30-minute math

One sitting, three lines:

What you spend now. Annual AI spend: subscriptions plus API usage. The real number, from the card statement, not the guess.
What self-hosting would cost. Hourly GPU rental for a model of this class, sized to your actual volume for the hours you would actually use it — plus a monthly ops line even if the ops person is you, because your time is not free either.
The difference. If line two is not clearly smaller — not marginally, clearly — keep renting.

The output is a number, not a decision. Most businesses run this and land on rent. The value is that the next time a vendor raises prices or a contract comes up for renewal, you are arguing from a number instead of a feeling.

What to actually do on November 4

Three things, in order:

Read the license. It was not published at announcement, and "open weights" is a marketing phrase until it is. The license decides whether commercial use, hosting for clients, and productization are allowed — and those terms can change the math on their own.
Recheck your trigger list. No compliance driver, no predictable volume, no renewal on the table? Note the date and move on. The math does not expire, but it deserves a quarterly recheck — this category re-prices faster than most software you buy.
If a signal applies, price it. One evening, the three lines, done.

Honest boundaries

The weights are not in hand yet. "By November 4" is a promise, and ship dates in this industry have slipped before. Nothing here requires the download to exist for the math to be worth having.
Hold the benchmark numbers loosely. The Artificial Analysis index is independent; Moonshot's own comparison tables are vendor-run and should be read as direction, not measurement.
Renting is not the same as ignoring. The rent lane still needs its own discipline — a quarterly review of what you pay and which tier you are on. Cheap self-hosting math does not excuse a lazy rental bill.
This is a recheck, not a switch. If you finish this post more certain you should rent than when you started, it did its job.

Tonight's version

Write the three lines: what you spend, what self-hosting would cost, the difference. Thirty minutes. Mark November 4 on the calendar to read the license. That is the whole assignment — most of you will still rent, but now you will know why, and you will have the number to prove it.

Want the worksheets? The Boring Automation Pack includes the spend-audit and quarterly-review worksheets — the rent-lane discipline that keeps this math honest between rechecks.