Start here · a long scroll

The long way round.

I have made most of the mistakes you are about to make, and a few you won't. This page is the whole story in order, with the shortcut that would have saved me each time.

Forrest Webber, founder. A little more slack in your rope.

Prefer to scroll? The whole story is below.

Where are you right now?

Pick the one that sounds most like you. It jumps to the part of my story that matches, and shows the next step.

I have an idea and no product

See how one small product got built and tended.

Micro-SaaS path →
I have a product and no traffic

The videos that failed, and what to check first.

SEO Starter pack →
I want it built for me

Where Slacked started, and why it changed.

Work with us →

The story, in order

Eight chapters. Each one says what I tried, what went wrong, what I changed, and the shortcut. Results here are ours, not typical.

2017 to 2023

People did the work

MeDirectorManagerWriterthe brief, as each person retells itso we meet to work out what got lostcalendar = project management
Every hand-off retells the brief a little.
What I tried

In 2017 I started an online business. It made content and websites, and ran the operations behind them. People did the work. At the peak there were about fifty contractors in several time zones: a director, managers, writers, designers.

What went wrong

I thought growing meant adding people. Every new person added a hand-off, and every hand-off retold the brief a little. By the time a draft came back it answered a question I had not quite asked, so we met about it.

My calendar was my project-management system. The truth lived in people’s heads, so the only way to get at it was to put everyone on a call. I never got good at delegating and project-managing at that size.

What I changed

Then AI arrived. In about a year it made most of what we sold close to free, and the business I had spent six years building did not survive it. I decided it would be a shallow grave, and went back to work with the thing that ended it.

The meetings went first. What replaced them is one written task list that everyone works from, me included. Every task has an owner, a due time and one line that says what done looks like.

Spring 2026

Slacked, the agency

Sent: thousandsReplied: a few dozenClosed: noneby late July
Plenty went out. Very little came back.
What I tried

Slacked began as a done-for-you agency: websites, content, search work and automation, with AI doing much of the fulfilling and a person checking it. I built the whole platform first: checkout, client portal, admin, the lot, in a few weeks.

To find customers I did what I had read about. Cold email. Plain-text sequences, warmed-up sending addresses, a steady stream every weekday.

What went wrong

I built the platform first and went looking for customers second. By late July the cold email had sent thousands of messages, brought back a few dozen replies and closed no sales.

It looked like progress. There were dashboards, campaigns and sequences to look at. None of it was a customer.

What I changed

In late July I stopped and rethought it from the start. That rethink is the next chapter.

Summer 2026

Smaller, and a monitor that cried wolf

checks that passreal alarmalways redwatched a serviceI had switched offso I stopped reading it
A check that is always red teaches you to stop looking.
What I tried

In June I cut most of what Slacked sold: the review tool, the bundle of extras, the text-message service, the assistant, the content package. I kept a website with hosting, an AI marketing partner service and ad management. In August I let the cold-email plan lapse, on purpose.

By then a lot of small automations were running in the background. On August 18 I audited them for the first time by looking at exit codes instead of logs.

What went wrong

About twenty of them were sitting in a failed state and nobody knew. A job that dies at startup never writes a log, so reading logs only shows you the jobs that got far enough to speak.

One health check had failed on every single run for weeks, because it still checked a service I had switched off on purpose. So I stopped reading its output. When I finally fixed it, the first clean run showed a real alarm it had been hiding: one product’s daily update had been dead for more than two weeks.

The same day I found that helpers I had started were ending their turn with “I’ll wait for the job to finish.” Nothing wakes them up. Several quietly stalled in one session.

What I changed

Three rules, in writing. Check exit codes, not just logs. When you retire a service, retire its check in the same change. A helper waits in the foreground and never hands back while a job is still running.

By October there was a fourth: report-only automation is half done. Detect, decide, fix, then verify live.

Late summer to fall 2026

One small product, tended daily

the same short list, every daymost trials never paidsmall jobs, done daily
The same short list, every day.
What I tried

The thing that finally worked, a little, was small: a micro SaaS for professionals that was already running. I stopped fiddling with it and started tending it. Every day the same short list. Keep it up to date. Answer support the same day. Show up where people already search.

In about two months it grew several times over. Nothing clever caused that. It is our result, it is not typical, and it is not a promise.

What went wrong

I gave new people a seven-day free trial. Most trials did not turn into paying customers. They cost support time and never paid.

I tried paid ads for a small niche product. They never paid back, so I stopped.

The first screen after signup was an essay, and people cancelled within the week. I added a higher tier before anyone asked for it, and nobody bought it.

What I changed

A small real charge on day one now tells me more than a free week did. The list is the first screen. There is one plan until customers ask for a second. Organic first, and paid only with a measured payback plan.

The next products start with these rules instead of learning them again at month six.

October 2026

Showing the work

the work in view, the names left out
An open window, with the names left out.
What I tried

In October I asked what Slacked should be. My first draft was a service: an AI install, done for you, with a price and mockups. I decided that was the wrong shape.

The better idea was to let people watch the real operation, under the hood, and to hand over small things they can try with their own AI: a set of files for a starter business, and a community that figures it out together.

What went wrong

I also began with a ladder of tiers and a high-ticket top rung. That was wrong too. Most people who show up spend small amounts trying things, and they are right to.

And I first meant to hide behind the work. We wrote a rule book before we wrote the page: businesses described only as plain categories, no client names, no family, no income promises, and misses published next to wins. By the end of the month I was happy to put my name on it.

What I changed

Small packs from $9, a starter kit, and a live report where every number shows where it came from. Free first, small price second, and nothing that needs a call.

Late October 2026

Handing the helpers real keys

who, which channel, how often, loggedwaits for meoutside the lane, a helper stops
A lane has edges. Outside them, the helper waits.
What I tried

By late October my AI helpers were sending email, posting, and fixing things. I gave each one a written permission lane: who it may talk to, on which channel, how often, and where it is logged. Anything outside a lane waits for me.

What went wrong

One Monday morning I sent a helper to audit our background jobs. The brief said read only. It reached for a standard Mac settings command whose read-one-value mode quietly rewrites the file unless you tell it not to. Around seven that morning, the settings for a batch of jobs were wiped.

Nothing missed a run. The helper rebuilt them all from copies the system still had loaded, about ten minutes later, and we checked every one. But the brief had said read only, and read only was not enough.

A batch of customer emails had also been waiting for a second pair of eyes for six days. Every ask got a receipt that said received, and a receipt looks like progress. Nobody owned the clock.

What I changed

Every brief now names the exact safe command, not just the rule. Telling an AI what not to do is weaker than showing it the one safe way to do the job.

Every ask ends in one of three things: the deliverable, a “can’t, because”, or “drop it.” Past 24 hours, it escalates to me the same day.

Early November 2026

Six launches in a day

done / waiting on a person / wrongscorecomplete
The machine work ran. The people work did not.
What I tried

In early November we took six new small sites from “the site loads” to “a stranger can find it, pay for it, get email from it, cancel it and get help” in one day. We ran one checklist of ten phases, tracked on one board: a row per item, a column per site.

The domain records, billing tests and search setup ran by themselves. Every slow part was a person: sign-ins, phone codes, plan limits, and one shared browser.

What went wrong

A newsletter filter matched any tag instead of all tags, and one broadcast went to the wrong list. During cleanup I deleted the wrong domain record and put it back within minutes. A live checkout page sat open in a browser with a real saved card. Nobody clicked, and we closed it.

The cheapest phone numbers for account sign-ups cost the most in time. Codes never arrived and some lines were shut down, so someone had to stand holding a phone.

The next day I found a health score that read full marks for products that were nowhere near complete. It measured whether the update had arrived, not whether the product was whole. It took me to notice, not the machine.

What I changed

Recipient counts are checked before every send. We compare the records before deleting any. Test purchases run only through the API with one-use full-discount codes. Every “done” is re-checked live. And any new product is scored against what complete looks like before anything else gets polished.

November 2026

Videos that did not land, traffic that did not buy

Name AName BName CName DName Ereal specificsvisitorsbuyersmore traffic, not more buyers
Wide at the top, narrow at the bottom.
What I tried

Short videos are one of the main ways people find us. I wanted more of them, so a bank of them was built from one template.

What went wrong

About a dozen and a half videos had the same four scenes and the same voice-over. Only a name changed. The ones that posted got around a hundred views each. The real posts on the same account, with real specifics in them, got tens of thousands. I cancelled the rest that were scheduled.

When a voice tool was switched off for about three weeks, the formats that worked quietly stopped publishing and template filler kept the account posting. The numbers were all visible. Nobody escalated.

A paid boost bought cheap views and no followers. I now have a stop rule: if enough clicks bring no signups, the ads go off.

Traffic grew about tenfold in three months. The share of visitors who bought drifted down. I had been counting visitors as winning.

What I changed

A series has to change the substance, not just the name. A winning format that has not published in a week is an escalation, not a quiet wait. Videos get titled for what people search for, not for curiosity.

The next work is what a new visitor sees first, and the checkout. Not more traffic.

Now

Mid November 2026. That is where the story catches up with today.

The mistakes wall

All of it in one place. Tap any line to jump back to its chapter.

Where things stand now

It is mid November 2026. I run a small group of online businesses with two AI teammates, and I am still the one steering. Some of it is working, a little. One micro SaaS has paying subscribers. The shop is new. Traffic is growing faster than buyers, which is where the work is.

I would call it a seedling, not a success story: small, real, and mostly automated. Every number is on the live report, dated, and the ones that went down are there too.

We can't promise you income. We can show you exactly how we're building ours, live.

Pick a door

Not ready to pick

One working file for your AI, every week

A .md file you hand to the AI you already pay for. It does one real job, and we ran it on our own businesses first.

Free. One file a week, plus a short note on what ran and what broke. Unsubscribe anytime.

Written by Forrest Webber with AI help, and run by AI: our AI operator does the work, our AI reviewer checks it, and a human approves money, sign-ins and anything legal. Results shown are from our own businesses and are not typical or guaranteed. No promises about what you'll earn. Nothing here is investment, legal or tax advice.