Fredy Montero · Still playing

From the Field
to Code

22 years pro. English as a second language. I talk to a computer now, and I ship.

Deportivo Cali/ Atlético Huila/ Seattle Sounders/ Millonarios/ Sporting CP/ Tianjin Teda/ Vancouver Whitecaps/ Real Cartagena/ Still on the field/

Professional goals

Years pro

Countries played

Continents scored on

Goals on four continents. Tap one to watch.

The turn

Let me be honest with you. I am a soccer player, not an engineer.

Deportivo Cali. Seattle Sounders. Sporting CP. Vancouver Whitecaps. Tianjin Teda. Real Cartagena. Twenty two years on the field, and I am still playing. Code was never my world. I am a dad of three girls and a follower of Christ, and my hands were made for a ball, not a keyboard.

But something changed. The AI got good enough that you do not have to type code anymore. You talk. You plan. It builds. People call it agentic engineering. I just call it the first time in my life that software felt possible for me.

So I stopped making excuses and I set it up. One piece at a time. Free tools, no fancy degree, English that is not perfect. And I started shipping real apps that real people use. If I can do this, you can too. That is the whole reason for this page.

The road

Deportivo Cali

Atlético Huila

Deportivo Cali

Seattle Sounders

Millonarios

Sporting CP

Tianjin Teda

Vancouver Whitecaps

Sporting CP

Vancouver Whitecaps

Seattle Sounders

Deportivo Cali

Real Cartagena

Still on the field · 2026

Real talk about money

What you need, and what it costs

Nobody tells beginners this part. I will.

$20/mo

Starter plan · check current prices

The tool I use is Claude Code, made by a company called Anthropic. It is not free, but it is cheaper than most people think.

You get it with a Claude subscription. The starter plan is around 20 dollars a month, and that is enough to begin and build real things. There are bigger plans for when you go all in and build all day.

Prices change, so always check the current ones at claude.com/pricing. That is the honest cost. A computer, a few dollars a month, and your own effort. No degree, no team, no waiting for permission.

For beginners

Start here in 10 minutes

You have never done this? Good. Here is what I would tell a friend on the phone. Five steps.

Get a coding agent on your computer

You need one tool that can build for you. I use Claude Code. Install it, open it, and say hello. That is your teammate now. Do not worry about understanding it all yet.

Say your idea out loud

Turn on voice or just type it rough. Tell it what you want to make, like you are explaining it to a friend. It does not have to sound smart. Mine never does.

Ask it to make a plan first

Before it builds, say: make a plan first. This one habit changes everything. The plan makes the agent do the real work instead of rushing. You do not even have to read it.

Let it build, then look

Tell it to build the plan. Let it run. When it is done, look at what came back and tell it what to fix. That is your real job: taste and direction, not typing.

Ship something small today

Do not wait for perfect. Put something tiny out into the world. Nobody starts big. I did not. You learn ten times faster once something is real.

My setup

What I actually run

This is not theory. This is what is on my Mac right now. The simple version of each one, in my words.

Always start with a plan

Before I build anything, I make a plan first. I do not open code. I tell my agent the idea and it writes the plan. Then it builds from that plan. A plan makes the agent do the deep work instead of cutting corners. This is rule number one for me.

HowInstall Compound Engineering once: /plugin marketplace add EveryInc/compound-engineering-plugin. Then type /ce-plan to plan, and /ce-work to build it.

Research before I decide

Picking a tool? Starting a feature? I check what people are actually saying right now, this month, not old information. Then I plan with that. You make way better decisions when you are not guessing.

HowInstall last30days (open source), then run /last30days your topic before you plan.

I talk, I do not type

I do not type code, and I barely type English. I talk to my computer. The AI is smart enough to understand me even when my accent or my words are not perfect. For a guy whose first language is Spanish, this changed everything.

HowInstall a voice app like Wispr Flow (free). Then just talk into your agent instead of typing.

Lots of tabs at once

I keep a few sessions open at the same time. One is planning. One is building. One is fixing something. I bounce between them. It is like having a small team that never gets tired.

HowGet cmux when you start running more than one thing at a time (with one project you do not need it yet). Install with Homebrew: brew install --cask cmux, or download it from cmux.dev. Then keep four to six tabs open, one task in each.

Let it run

I let my agent work without asking me for permission on every little step. With a few sessions going, you cannot babysit each one. You point it, you trust it, you check the result. It is my computer. If something breaks, I fix it.

HowIn ~/.claude/settings.json set "defaultMode": "bypassPermissions" so it stops asking on every step.

A sound tells me it is done

When an agent finishes, my Mac makes a sound. So I can walk away, play with my daughters, and come back when I hear it is ready. I am not staring at a screen waiting.

HowAdd a Stop hook that plays a sound when it finishes: afplay /System/Library/Sounds/Blow.aiff.

One keystroke to start

I set my terminal so a new tab opens straight into my AI helper. No typing, no setup. I open it and it is already there. When starting costs nothing, you start a lot more. Small thing, big difference.

HowPoint your terminal to open a small script that runs claude on every new tab. One line in your terminal config.

Latest addition

Copy, paste, go

Steal these prompts

Anything in a dark box is meant to be pasted, not typed. These are mine. Tweak the parts in orange.

Here for the Replies — research the idea

/last30days what people say are the funniest / most-liked replies and comment culture on [topic] this month

Here for the Replies — plan the build

/ce-plan rank the most-liked comments on a topic as a table. Reddit + YouTube only for now; leave X and Instagram as a later upgrade, don't block on them.

Santo Coffee — weekly brief

/ce-plan a Monday brief that runs last30days on "Santo Coffee" and my local coffee scene and gives me the 5 things worth knowing, in plain language.

Soccer academy — announcement helper

/ce-plan a tool that turns a few bullet points into a clean weekly announcement for academy parents, in English and Spanish.

Turn a voice note into a plan

/ce-plan here's a rambly voice note about an idea. ignore the tangents and turn it into a real plan.

Make a launch video

/hyperframes a 30-second launch video for [tool]. here are screenshots. punchy, big logo, my voice on the script.

When you're stuck

[paste a screenshot] what do I do here? fix it for me.

Don't read the plan

eli5 this plan in one paragraph, then just /ce-work it.

These are the exact ones I lean on. Copy one, swap the orange part for your idea, and let the agent run.

The deep version

The full playbook

You do not need this to start. The prompts above are enough. But when you want the whole thing, it is all here, in Matt's own words.

The moment you have an idea, make a plan

Not "let me think about it," not "let me start coding." Every time. The thinking goes in the plan; the execution is mechanical.

Do not read the plan

Plans are for agents. Forcing a plan to exist makes the agent research and commit to an approach. Make the plan. Trust the plan.

Research before you plan

Before I plan, I check what people actually know right now, not six-month-old training data. Research, plan, build.

Every Agentic Engineering Hack I Know 21 hacks · by Matt Van Horn · tap to open

This is Matt's whole stack, written up. Lightly trimmed for this page.

Every Agentic Engineering Hack I Know (June 2026) — by Matt Van Horn.

The TL;DR hack: paste this whole thing to your agent and tell it to set up everything it can. That is the fast path.

The moment you have an idea, make a plan

The moment I have an idea, it is /ce-plan. Not "let me think about it," not "let me start coding." Every time. It takes images too. Screenshot an error, a design, a bug, drop it in. When the idea is still fuzzy, start with /ce-brainstorm to think it through, then /ce-plan once it is sharp. Under the hood it fans out research agents, then writes a structured plan. /ce-work builds it. The thinking goes in the plan; the execution is mechanical.

HackPaste a screenshot, bug, or error, then /ce-plan, then /ce-work. Fuzzy idea? /ce-brainstorm first.

Do not read the plan

I always make the plan. I almost never read it. Plans are for agents. Forcing a plan to exist makes the agent research, commit to an approach, and actually hit its own acceptance criteria. I skim the title and run it. If I have a question I ask inline: "wait, why this approach?" or "eli5 this plan." Make the plan. Trust the plan. Do not read the plan.

HackAsk inline: TLDR?, eli5 this plan, or "why this approach?"

Use the plan loop for your deepest non-engineering work

The biggest thing I have learned is that the plan loop is not just for code. Strategy docs, product specs, board updates, same loop. The trick is to make the first plan a plan for the plan: hand it all your context and a transcript and say "do not write the doc yet, just plan how you will produce it." It is the single best trick for making an LLM not lazy.

HackDeep work: "make a plan for the plan," hand it everything, then build.

Get voice-pilled

Voice-to-LLM is different because the listener understands context. It guesses what the mic got wrong. You can mumble, trail off, restart. On Mac I use a dictation app and a gooseneck mic. On the phone, the built-in dictation is plenty, because you are talking to an LLM, not a human.

HackTalk into your agent. Dictation app plus a mic on the desk; phone dictation on the go.

Lots and lots of tabs

Four to six sessions at once: one writing a plan, one building from another, one running research, one fixing a bug. While one researches, I switch to another. By the time I cycle back, the first is done and waiting.

HackKeep four to six sessions open, a different task in each.

Make your terminal open into the agent, not a shell

A new tab should drop you straight into an agent. When a new session costs one keystroke, you start a lot more of them.

Remote-control every window, and give the agent an email address

Turn on remote control so every session is reachable from your phone. Start at your desk, steer from a soccer field. You can even give the agent an email address so emailing it opens a fresh working session.

Dangerously skip permissions

With several sessions you cannot confirm every edit. I run in bypass mode. It is my computer, and version control is my undo button. Pair it with a sound that plays when a session finishes so you know when to come back. Newer "auto" mode is a safer middle ground if that makes you nervous.

HackAdd a sound that plays when a session is done, so you hear it finish.

Run work through a second engine

One agent plans, a second coding engine builds. Two engines side by side is a whole extra set of hands. Push big parallel builds to one, keep the other on planning and taste.

Research before you plan

Before I plan, I run last30days on it. Choosing between two libraries, picking a feature, prepping for a meeting. It searches Reddit, X, YouTube, TikTok, Hacker News, GitHub and the web in parallel and hands back what people actually know right now, not six-month-old training data. Research, plan, build.

HackBefore you plan, run /last30days on your topic.

Record everything, and paste the raw transcript

Record your meetings, then drop the raw transcript into the plan. Do not summarize first. The tangents do not matter; the agent extracts the signal against your codebase and prior plans. I have turned a lunch conversation into a product proposal the same night.

HackDrop the raw meeting transcript into the plan, do not clean it up first.

You are the human signal

When you run several agents, your job is not to do the work. Your job is to be the signal. The agents supply volume; you supply taste and direction. "Option two, but use the language from option one." "This paragraph is too long." The rare, valuable thing is your judgment, not your typing.

HackBe the taste. Let them be the hands.

Build video like everything else

Video used to be the thing I skipped. Now I make it the same way I make everything. I talk, an agent builds it as HTML, I react, it renders a video. Launch reels, demos, explainers. The cost of a video dropped to a conversation.

Your notes are your agent's knowledge base

The reason a plan gets better every time is the agent can read every prior plan I have written. So I pointed it at my whole brain, a notes app with a CLI, a synced knowledge store. Pick a notes tool with an API and let your own knowledge compound.

Work from anywhere

A remote machine plus tools that survive bad wifi means I have shipped features on a flight home. The work runs on the box, not the laptop. You reconnect and it is exactly where you left it.

Share a plan with a normal human

A plan is perfect for me and useless to hand a coworker who does not live in a terminal. Publishing it as a clean, commentable doc closed that gap. Non-terminal people read it, comment inline, and those comments flow back into the loop.

Write your own skills

The biggest level-up is teaching agents tricks that stick. Anything I do more than twice becomes a skill, a reusable command. You do not write them from scratch. Point the agent at a skill that already works and say "make me one like this for X." Write it once, every session after is faster.

HackDo it twice? Make it a skill. "Look at this skill and make one like it for [X]."

Contribute to the projects you love

The same loop that ships my projects ships everyone else's. Pick a tool you use every day, find one real thing it is missing, and ship it. The merged pull requests are not even the real prize. It is meeting the people who build the thing.

HackOne tool you use daily, one real missing thing, one pull request. Then show up in the community.

Real hardware, and CLIs that run real life

Running this all day is heavy, so carry power. And the frontier is agents doing errands, not just code. A fleet of small tools that wrap real services so an agent can preheat the car, reorder groceries, watch a game and ping you when it is close, or price out flights from a soccer field. Not "AI writes my code." Agentic engineering that runs real life while you do something else.

The honest part: do not disappear into it

Agents were supposed to do the work for us. Instead everyone I know is working the hardest they ever have, because the build loop is the best video game ever made. Take breaks. Talk to your loved ones. Ask if anyone actually wants the thing. And if it is just a tool for you, that is okay too. If you do want an audience, you start at zero and work up. Nobody starts at thousands.

HackTake breaks. Talk to the people you love. Build something people want, even if "people" is just you.

This page was made this way

Talk, plan, build, react. From a desk, a couch, a car, a soccer field. That is the whole thing, Fredy. Now go build.

All 21 hacks by Matt Van Horn. Follow Matt Van Horn on X →

Proof it works

What I have built

Talk is cheap. Here is the proof. Real apps I built by talking, planning, and letting the agent do the hands.

MiCharla Tac

A tactical board for coaches. Draw the play, animate it, export the video. Free, works offline, no account.

LiveOpen source

The MiCharla Tac board with a play drawn on it

Here for the Replies

A leaderboard for the funniest, most-liked replies. The comments are the show.

Live

First Touch

A Spanish app for a kids soccer school.

Web app

2Toques

A 2D soccer game for iPhone. Because of course I built a soccer game.

iOS

Writing Companion

A tool that helps my wife keep her novel consistent as she writes.

For family

Pills Taken

A medication reminder for one person to use and their family to set up. One pill, one button.

In review

x-this

A little tool I published that writes replies in my own voice.

On GitHub

None of these needed a computer science degree. They needed an idea, a plan, and the patience to react and redirect.

The honest part

Do not lose yourself in it

This stuff is powerful, and it can pull you in. Building with agents is the most fun I have had with a computer in my life. That is the danger too. It is easy to disappear into it.

So I will tell you what I tell myself. Take breaks. Touch grass. I am a dad of three girls and a follower of Christ, and none of that can come second to a screen. Build things people actually want, even if the only person is you. And do not lose the people around you while you build.

Start small. Post into the quiet. One person notices, then three, then more. Nobody starts at the top. I sure did not.

Yeah, that is my little builder avatar. We do not take ourselves too serious around here.

Your turn

If a soccer player can build this, so can you.

You do not need to be a genius. You need to start. Pick one idea. Make a plan for it. Let the agent build. That is the whole game.

Follow the journey on X →

Credit where it is due

Real talk. I did not invent any of this. My friend Matt Van Horn handed me the playbook, his hacks, his whole article. I just took the pieces that fit my life and started building. That is the secret nobody tells you. You do not have to figure it out alone. Find someone a few steps ahead and follow them. Matt did that for me. Now I want to do it for you. Come follow my journey and build with me.

Follow Matt Van Horn on X →