Tools

What You'll Build

A single-purpose talking assistant that gives a blind, elderly, or non-English-speaking family member their day back. He presses one big button, says what he wants in his own language, and the computer does it and tells him out loud that it did.

No screen. No menus. No passwords. No "swipe here" or "tap the third icon from the left." Just a button and a voice.

This is a blueprint, a "steal this idea" build. We are not going to pretend a specific family already has this humming away in a living room in Buenos Aires. We are going to show you the exact real problem, then show you how you would build the thing to solve it. The problem is extremely common, and the pieces to solve it finally exist and are cheap.

The Story

Two posts went up on Reddit from someone trying to help their 72-year-old father-in-law. He lives in Argentina, he speaks only Spanish, and he recently went completely blind.

Think about what that combination takes away. Overnight, the phone stops working for him. The TV remote is a grid of identical buttons he can no longer see. His email, his music, his ability to call his own daughter without help, all of it now runs through a screen he cannot use.

The family tried the obvious things first. Standard phone and computer accessibility tools, the built-in screen readers and voice assistants. They were too hard and too unreliable for him. The big-name voice assistants wander off script, mishear, answer the wrong question, or launch into something nobody asked for. For a sighted person that is annoying. For a blind 72-year-old who cannot see what went wrong, it is the difference between independence and giving up.

So the family asked a sharper question. Not "how do we teach him to use a computer," but "how do we build him a tiny, dependable machine that only does the handful of things he actually wants, and never surprises him?"

That question is the whole playbook. The magic here is not a smarter chatbot. It is a dumber, safer one, wrapped around a real person's actual daily life.

Why "Restricted" Is the Whole Point

Most people building with AI want it to do more. This build wants it to do less, on purpose, and that is exactly why it works.

A general chatbot is a bad fit for this job. It can go anywhere, which means it can go wrong anywhere. It might misunderstand "call my daughter" as "call Diana's Bakery" and dial a stranger. It might get confused and just fall silent, leaving a blind man talking to a machine that is not answering.

So we flip the design. Instead of a machine that can do anything, we build one that can do a short, fixed list of things and nothing else. In plain terms, an "allow-list" is a short menu of approved actions the assistant is permitted to run. If a request is not on the menu, it does not guess and it does not improvise. It gently tells him what it can do and waits.

That single decision, do less but do it perfectly, is what turns a flaky science project into something a family can actually trust with someone they love.

How It Works, In Plain English

Here is the whole thing, described the way you would explain it to a friend.

One button to talk. There is a big physical button on his table. He presses it and the microphone turns on. This is called "push-to-talk," which just means the computer only listens when he chooses to talk to it. Nothing is listening in the background, which is better for privacy and means the assistant never reacts to the TV or a passing conversation. The button can be a cheap USB "macro" button, the kind streamers use. It costs a few dollars and it is impossible to miss by touch.

He speaks in Spanish. He says something simple. "Poné las noticias." "Continuá mi audiolibro." "Llamá a mi hija." The computer turns his spoken Spanish into text it can act on. That step, the computer understanding what he said, is called speech-to-text.

The assistant confirms before it does anything that matters. Before it places a call or does anything he would not want done by accident, it asks out loud, in Spanish, "Do you want me to call your daughter? Say yes or no." He says yes. Then it acts. This confirmation step is small and it is the difference between a tool he trusts and a tool that scares him.

It does the thing and tells him it did. It starts the news, resumes the audiobook, reads the email, places the call. Then it speaks back to confirm what is happening. That speaking-back step, the computer talking to him in a natural Spanish voice, is called text-to-speech.

If it does not understand, it never fails silently. This is the part cheap voice gadgets get wrong. When the assistant is not sure what he wants, it does not sit there dead. It says, in Spanish, "I didn't catch that. You can say: play the news, continue your book, play tango, read your newest email, or call your daughter. What would you like?" It re-offers the menu. He is never stranded.

The daughter runs the settings from anywhere. The caregiver controls a simple config file, a plain settings sheet, that holds his favorites, his contacts, and which actions are turned on. She can add his cardiologist to the contact list, swap the news station, or turn an action off, all without touching the machine in his house.

The Command Menu

Start with a short list. These are the commands that are genuinely clean to build, meaning the software gives you a proper hook to control them:

Notice how small that is. Five or six things. That is not a limitation, that is the product. He will use these every day, and every one will work.

How You'd Actually Build It

The machine. An old laptop you already have, or a small mini PC, sitting in his home and always on. It plugs into power and stays put. No screen interaction ever required.

The ears and the voice. The agent runs OpenClaw or Claude Code as its brain. For understanding his Spanish, you use a speech-to-text tool like Whisper, which handles Spanish well. For speaking back, you use a text-to-speech voice like ElevenLabs, which sounds warm and human rather than robotic. That warmth matters more than you would think for someone who now experiences this machine entirely through its voice.

The button. A USB macro button wired so that pressing it starts the microphone listening. One press, one turn to speak.

The recipe. You write the agent a short set of instructions in plain language. In OpenClaw or Claude Code this is just a text file, not real programming in the scary sense. It says, in effect: "You are a Spanish-speaking assistant for one person. You may only do these five things. Before calling anyone, confirm out loud. If you are unsure, read the menu back. Never do anything not on this list." That file is the safety rail.

The settings sheet. A simple file the daughter controls with his favorites and contacts. Which YouTube channel is "the news." Which playlist is "tango." Who "my daughter" maps to. Change the file, change his world, from her kitchen.

Wire those five things together and you have it. A weekend to get the core three commands working (play media, read email, call a contact), and a second weekend to smooth the voice, tune the confirmations, and add the settings sheet.

Gotchas and Honest Limits

We are not going to tell you the whole thing is a Saturday of copy-paste. Three of the commands are genuinely clean. Two are harder, and you should know that going in.

Audiobooks are the messy one. "Continue my audiobook" sounds simple and is not. Audible, the big audiobook service, has no clean automation hook, meaning there is no tidy, supported way to tell it "resume playing" from a script. You have a few honest options. You can control it with the media keys, the same play and pause signals a keyboard sends, which works but is cruder. Or you can move his audiobooks to a service that plays nicely, like a podcast app or Apple Books, and point the "audiobook" command there instead. Do not promise "continue my Audible book" and quietly fail. Pick the path that actually works and build the command around that.

TV volume needs extra hardware. If you want him to say "turn the TV up," the computer has to physically reach the TV. That takes either HDMI-CEC, a feature where the TV listens to commands over the same cable that carries the picture, or a small infrared blaster, a cheap gadget that mimics the TV remote's signal. Neither is hard, but both are an extra part to buy and set up. Do not assume the laptop can control the TV out of the box, because it cannot.

Confirm the high-stakes actions, relax the low-stakes ones. Confirming "play tango" every time gets annoying fast. Confirming "call my daughter" every time is exactly right. Put the confirmation step only where a mistake actually costs something.

Test the failure cases on purpose. Mumble at it. Say something off-menu. Say half a command and stop. The most important thing this build does is fail gracefully, so make sure it always lands back on "here is what I can do."

The Business Angle: A New Service You Could Sell

The heart of this build is not money. It is dignity. A blind man calling his own daughter without asking anyone for help is the entire point, and it would be crass to bury that under a spreadsheet.

But there is a real business here, and it is a service someone could start selling next month.

Millions of families have a parent who is losing their sight, their hearing, their memory, or who never spoke the local language and now cannot navigate a screen. Those families do not want a chatbot. They want exactly this: a safe, restricted, spoken-word helper set up for their specific parent, in their specific language, with their specific contacts.

Almost nobody sells that setup as a service today. You could. "I will set up a voice assistant for your elderly or disabled parent" is a clean offer with real demand behind it. The people positioned to sell it:

Price it as a setup fee plus a small monthly check-in, because the settings will need updating as the parent's life changes. This is not replacing a $2,000-a-month vendor. It is creating something that mostly did not exist, and that families will happily pay for, because you are selling back a piece of a parent's independence.

Who Should Steal This Idea

Anyone with a blind, low-vision, elderly, or non-English-speaking family member who has been shut out of their own phone and TV. Build it for them first.

Then, anyone who helps those families for a living. Independent tech helpers, home-health agencies, senior-living operators, occupational and low-vision therapists. There is a paid service hiding inside this build, and it is wide open.

Keep Reading

Want the foundations? Read our Best OpenClaw Skills guide for the skill patterns this build is made of.