Halo
September 2026 · For friends
· Nine months in

What's still working.

I run a company on my own. This is what I'm doubling down on, what I've let go, and why.

Amos Whitewolf · Halo

Halo shield
If you saw the April version, this is the update. If you didn't, start here.01 / 23
Halo
02 · The short version
· The short version

Five things.

1 · Less

Smaller instructions. Better output.

2 · Loops

Give it a clear end and it runs itself.

3 · Buy

Stop building your own tools.

4 · Judgement

AI multiplies what you bring.

5 · Setup

Two tools, one rule.

The rest of the deck is these five, in order02 / 23
Halo
03 · Part 1 · Less
· Part 1 · Less

Reduce everything.

Nine months ago, more instructions did mean better output. The models got smarter. Now it's the other way round. Smaller skills. Fewer words on every page. The model and the operator can put it together.

The one lesson everything else follows from03 / 23
Halo
04 · Part 1 · Less · Skills
· Part 1 · Less

Skills: still the core. Just smaller.

A skill is a reusable instruction file the AI reads before a job. Mine used to run to thousands of words with strict rules. The short ones do better. The model needs room to be right. Over-prescribe it and you stop it growing, and it's a pain to maintain.

4,563 → 874
Words in my own voice skill, before and after this week's cut. The output got sharper.
There's still a time for a detailed one. Next slide.04 / 23
Halo
05 · Part 1 · Less · When
· Part 1 · Less

Smaller doesn't mean loose everywhere.

Stay detailed when

It's an exact, repeatable task.

Steps in order, same every time. The blog pipeline. Long is fine here, because the job is long.

Go loose when

It's judgement or taste.

My voice skill. How I sound changes with my mood and who I'm talking to. Same with art, photography, video. Taste is the one thing you never over-prescribe.

Pick on purpose. Then let the things with a clear end run themselves. Part 2.05 / 23
Halo
06 · Part 2 · Loops
· Part 2 · Loops

A loop needs a clear end.

Loop it

Clear goal. Examples of right and wrong.

It gets there on its own, and keeps going while you sleep.

Don't loop it

Vague goal. No way to say what right is.

The loop wanders. This needs a human going back and forth, and that's fine.

Once the instructions are small, the question is what you let run without you06 / 23
Halo
07 · Part 2 · Loops · Iterate first
· Part 2 · Loops

Nothing starts as a loop. It earns it.

First, iterate

You and the AI, by hand.

New data every week. Rage clicks, funnel drop-off, why people don't pay. The human matters most here. The paywall is still at this stage.

Then, loop

When there's data on both ends.

A number going in to plan the work, a number coming out to judge it. Tie a bow on it and let it run. The blog got there.

Iterate until it's boring. Boring is the signal.07 / 23
Halo
08 · Part 2 · Loops · Live
· Part 2 · Loops

Three that have earned it.

Content

The blog.

Reads search data, picks the keywords it can win, writes, publishes, checks the result, plans next week. Every 2 hours, unattended.

Design

Pixel drift.

Screenshot the built screen. Screenshot the design. Overlay them. The agent sees the gap and fixes it. Loop until they match.

Product

The flow map.

Screenshot every screen of the site, the app and the watch app. Build one map. Every flow in one view. Used to be a full manual day. Sometimes longer.

Each one has a right answer the agent can check itself against08 / 23
Halo
09 · Part 2 · Loops · Result
· Part 2 · Loops

About 1,000 clicks a month. From zero.

Fresh domain, no paid traffic, 5 months. One post carries most of it, which is the point. I didn't pick that topic. The loop found the keyword.

1 → 40
Clicks a day, late April to mid September.
Search Console · 28 days to 19 Sep 2026 · 1,097 clicks09 / 23
Halo
10 · Part 2 · Loops · Why it matters
· Part 2 · Loops

Execution isn't the moat. Your data is.

Anyone can build the app now. Nobody can copy a loop that feeds on your numbers, because they don't have the numbers. It costs money to start one. Once it's running, it's hard to stop.

Which is exactly why the big software companies aren't going anywhere. Part 3.10 / 23
Halo
11 · Part 3 · Buy
· Part 3 · Buy, don't build

SaaS isn't dying. Slow SaaS is.

Same logic, applied to the companies selling you software. The ones with millions of users feeding them data now ship at startup speed. Their moats get stronger. And each one covers more ground every quarter, so the niche a new startup can own gets smaller.

Build on them. Don't try to be them.11 / 23
Halo
12 · Part 3 · Buy · Picking tools
· Part 3 · Buy, don't build

Picking the tool is the skill now.

PostHog is the analytics I use. This year it started reading my session replays and errors, writing the fix, and opening a pull request for me to review. I couldn't build that, and I shouldn't try. Knowing which tools to use matters more than it ever has. The tools you pick are part of your moat.

posthog.com/docs/self-driving, if you want the detail12 / 23
Halo
13 · Part 3 · Buy · Lesson
· Part 3 · Buy, don't build

I built my own attribution. Mistake.

I'm a developer, so I built it. Weeks I should have spent on product and marketing. Then the maintenance bill. And every bit of data I burned getting it right was data I could have used fixing the product. Time, money and data. Someone already sold the thing. Buy it on day one.

Your time is the most expensive thing in the company. Which raises the question of whose. Part 4.13 / 23
Halo
14 · Part 4 · Judgement
· Part 4 · Judgement

AI plus a beginner loses to AI alone.

Chess, 1997 to 2015

Human plus engine beat the engine. For about 10 years.

The 2005 winners were 2 amateurs with a better process. The edge was knowing what to check, not chess skill.

Harvard and BCG, 2023

Outside AI's competence, users did 19 points worse.

Consultants with AI, on a task it wasn't good at, were less often right than consultants without. They took what it said.

Then the flip side. Next slide.14 / 23
Halo
15 · Part 4 · Judgement · Flip side
· Part 4 · Judgement

AI plus an expert beats AI alone. By ten.

Same tools, opposite result. Hand AI to a video editor with 15 years of reps and they don't just make better work than you would. They make 10 times as much of it. Judgement is what switches the multiplier on.

Better work, and ten times the volume of it15 / 23
Halo
16 · Part 4 · Judgement · What I did
· Part 4 · Judgement

I hired the humans back.

6 months writing my own scripts and ads with AI. Didn't work. So: a script writer, an ads person, an AI video person. Same tools, better output, because they know what good looks like. The one hire I won't make is an AI developer who was never a developer. With video you see the problem straight away. With code you find out later.

Judgement is the thing the tool doesn't come with16 / 23
Halo
17 · Part 4 · Judgement · Research
· Part 4 · Judgement

Every skill starts as research. Check the source.

I built my paywall off research into apps people already trust. It gave me the wrong answers, because Halo isn't trusted yet and those apps don't have that problem. So I re-scoped it to products at my stage and what worked for them. Clearer answers, better results. When you turn research into a skill, be specific about who it came from.

AI will hand you the answer to the question you asked. Ask the right one.17 / 23
Halo
18 · Part 5 · Setup
· Part 5 · Setup

Hands-on in one. Always-on in the other.

Claude Code · desktop

Anything I'm doing with my hands.

Building, reviewing, thinking out loud. Every workflow starts here, by hand, until it's boring.

Hermes · always on

Anything that runs without me.

The blog loop, every 2 hours. Browsing on its own. Voice notes from my phone that land in Notion as scripts.

How the whole thing runs day to day. Codex next, once I can afford the week a setup change costs.18 / 23
Halo
19 · Where the time goes
· Where the time goes

Doubling down on.

Anything with a loop

The blog, and whatever earns it next.

Data on both ends means it compounds while I do something else.

Distribution over product

Getting Halo in front of parents.

The app works. It needs more people using it. 80 percent of my week now, up from 30.

Data moats

Finding numbers nobody else has.

Every loop starts with data you own. Working out where the good data is, and how to get it, is the job now.

Knowing the competition

Their ads, their accounts, their creative.

Research got deeper, not wider. It feeds the ads and the positioning directly.

Four things. Each one either compounds, brings in users, or gives me an edge nobody can copy.19 / 23
Halo
20 · Where the time goes
· Where the time goes

Letting go of.

Dropped

AI posting for me.

Autonomous social, AI-written ads. It's good with data and not good at taste, and I don't have enough data yet for it to learn. I write, AI helps.

Dropped

Selling to schools.

I was going to sell Halo through schools. Then I talked to teachers. Schools won't mandate it, parents will. That's domain knowledge you only get by asking the people in the room. Parent nights, not principals.

Parked

Building my own tools.

Attribution, gamification, anything a company already sells. One exception: where I know the domain deeply, like SEO and the blog, and no product does it as well as I understand it. That's a moat. Everything else, buy.

Everything here failed the same test: no data on both ends, or someone already does it better20 / 23
Halo
21 · If you want to try this
· If you want to try this

Claude Code is mislabelled.

The word "code" scares off the people who'd get the most out of it. It's not a coding tool. It's the agent that does everything. The chat window will do the thing. Go one level deeper anyway. That's where people are getting ahead.

Open the terminal. Type what you want. Watch.21 / 23
Halo
22 · The test
· The one question

What are AI's limitations?

If you can't answer with your own list, you haven't pushed it hard enough. AI tells everyone they're good at AI. That's why it's hard to prove you are. Go find the edges. Then you've got an opinion, and that's the part nobody can fake.

If you don't know where it breaks, you're not good at it yet22 / 23
Halo
September 2026
· Thanks for looking

Ask me anything.

If anything here sparked something, a question, a "wait, how did you", or just a riff over coffee, tell me.

Amos Whitewolf · September 2026

Halo shield
ai-for-friends · September 202623 / 23