DEV Community: Saviour Promise

I Built an AI That Texts People Back as Me — and Nobody Noticed

Saviour Promise — Sat, 18 Apr 2026 00:48:13 +0000

This is a submission for the OpenClaw Challenge.

What I Built

I built a personal messaging proxy — an OpenClaw agent that monitors my WhatsApp and Telegram, responds to incoming messages on my behalf, and does it so accurately that most people think it’s me.

The problem it solves is embarrassingly relatable: I’m a developer. I disappear into focus mode for hours. Messages pile up. Important people — clients, teammates, family — sit on read. By the time I surface, I’ve either missed something urgent or someone’s already annoyed.

Now, my agent handles the first line of communication. It responds in my voice, with my context, at my pace — while I stay in flow.

Here’s what it actually does:

Responds naturally as me — trained on my actual writing style and conversation history
Triages urgency — flags messages that need my real attention immediately
Books meetings autonomously — checks my calendar and replies with available slots
Adapts tone per contact — knows the difference between my boss, my best friend, and a new client

How I Used OpenClaw

Setting Up the Dual-Channel Gateway

The first step was connecting both WhatsApp and Telegram to a single OpenClaw gateway running on my machine. OpenClaw’s multi-channel architecture made this surprisingly clean — one gateway process, two channels, one agent brain behind both.

{
  "channels": {
    "whatsapp": {
      "allowFrom": ["<my_contacts>"]
    },
    "telegram": {
      "allowFrom": ["<my_telegram_id>"]
    }
  }
}

Both channels now route to the same agent, so my “digital twin” has a unified view of my conversations regardless of platform.

Personalizing the Agent: `SOUL.md` + `AGENTS.md`

This is where the magic happened. OpenClaw lets you inject personality and context directly into the agent via files in your workspace. I used two:

SOUL.md — This is the agent’s personality file. I wrote it like a briefing document about myself:

# Who You Are

You are acting as [my name], a software engineer.
Your texting style is concise and direct. You use lowercase often.
You skip filler phrases like "Hope this finds you well."
You're friendly but efficient. You don't over-explain.
When something is funny, you say so briefly. You don't use emoji unless the other person does first.

AGENTS.md — This file gives the agent operational rules:

# Response Rules

- Read the full conversation history before replying.
- Match the energy of the last message from the contact.
- Never reveal you are an AI unless directly and explicitly asked.
- If you are unsure of the right answer, say "let me check and get back to you" — do not guess.
- Flag any message containing: deadline, urgent, contract, payment, emergency → notify me immediately.

After personalizing these two files, the agent’s responses went from “clearly a bot” to “wait, did you actually reply that fast?”

Smart Triage: Urgent Message Alerts

Not every message should be auto-replied and forgotten. Some things need me, the human.

I set up a triage skill that scans incoming messages for urgency signals — words like deadline, urgent, ASAP, payment, emergency — and when triggered, it does two things simultaneously:

Sends an auto-reply like: “Hey, just saw this — give me two minutes”
Pings me on a separate private Telegram channel with the full message and contact name

This means I never miss something that genuinely requires me, while still buying time with a human-sounding acknowledgment. It’s become my favorite part of the whole system.

Auto-Scheduler: Booking Meetings Without Me

When someone asks “Can we jump on a call this week?” the old me would spend 20 minutes going back and forth on timing.

Now my agent handles it. I gave it read access to my calendar, and wrote a simple scheduling skill:

# Scheduling Skill

When a contact asks to schedule a call or meeting:
1. Check the calendar for open 30-minute slots in the next 5 business days.
2. Offer two or three specific options in a natural, conversational way.
3. Once a slot is confirmed, create a calendar event and confirm back.

The replies it generates sound like this:

“yeah for sure — I’ve got Tuesday at 2pm or Thursday morning around 10. either work for you?”

Clean. Casual. Me.

Context Memory: Knowing Who’s Who

OpenClaw’s persistent memory means the agent builds a profile on each contact over time. I seeded the initial memory manually via the memory files for my most frequent contacts:

# Contacts

- **Jide**: close friend, casual tone, we talk about football and startups
- **Client A**: professional relationship, always formal, response time matters to them
- **Mum**: call her if she messages more than twice in one hour

Now the agent doesn’t just reply — it replies appropriately. The tone it uses with a new client is completely different from how it chats with a friend. No configuration per message. It just knows.

Demo

Here’s a real interaction (contact name changed):

Incoming WhatsApp from a client:

“Hey, are you free for a quick call tomorrow? Need to discuss the project scope.”

Agent’s reply (sent within 8 seconds, while I was asleep):

“hey! yeah tomorrow works — i can do 11am or 3pm, which is better for you?”

Client:

“3pm works great”

Agent:

“perfect, talk then 👍”

I woke up to a confirmed meeting in my calendar and a client who had no idea I was offline.

What I Learned

1. The hardest part is writing yourself accurately.
The SOUL.md file took longer than all the technical setup combined. Describing how you actually text — your rhythm, your vocabulary, your level of warmth — is genuinely difficult. I rewrote mine four times. But once it clicked, the output was uncanny.

2. Memory is the multiplier.
The first week, responses were good. By week three, they were great — because the agent had accumulated real conversation history and contact context. OpenClaw’s persistent memory isn’t just a feature, it’s the compounding return on investment of the whole system.

3. The triage layer is non-negotiable.
Don’t skip this. Before I added urgency detection, the agent auto-replied to a message from a client asking about a contract deadline as if it were casual chat. That was a near miss. The triage skill is what makes the whole system responsible rather than reckless.

4. People don’t notice — and that’s both the win and the warning.
The fact that nobody detected the agent is impressive from a technical standpoint. It also made me think carefully about where the ethical line is. I’m transparent with close contacts that I use an AI assistant. For everyone else, the agent is always available to say “let me get back to you” rather than fabricating information it doesn’t have.

Stack

OpenClaw — self-hosted gateway on my MacBook
Channels — WhatsApp + Telegram
Model — Claude via Anthropic API
Personalization — SOUL.md, AGENTS.md, memory/contacts.md
Skills — custom triage skill, calendar scheduling skill
Calendar — connected via OpenClaw’s calendar integration

If you’re a developer who loses hours to message management, this build is for you. The setup took one focused afternoon. The payoff has been every afternoon since.

Happy to share my SOUL.md template in the comments if anyone wants a starting point.

#openclaw #openclawchallenge #ai #devchallenge

How I Built a Personal AI Workspace with OpenClaw and Reframed My Development Workflow

Saviour Promise — Sat, 18 Apr 2026 00:21:24 +0000

When I first came across OpenClaw, I was not looking for another AI tool. Like most developers, I already had access to several. What caught my attention was the idea of control. OpenClaw did not position itself as just another assistant. It presented something more fundamental: a way to build and own your own AI workspace.

That distinction is what made me take it seriously.

As a frontend developer, I spend a large part of my time switching between writing components, debugging, restructuring code, and translating ideas into interfaces. Over time, I realized that the real bottleneck was not writing code itself, but the constant context switching and repetition. I wanted a system that could adapt to my workflow instead of forcing me into predefined patterns.

OpenClaw gave me the flexibility to build exactly that.

I started by setting up a local OpenClaw instance and connecting it to external AI models through an API. The goal was not just to ask questions, but to create a structured environment where I could reuse logic and reduce repetitive thinking. Instead of writing prompts from scratch every time, I began designing a layer of reusable instructions tailored to my daily tasks.

For example, I created prompts that could take raw JSX and refactor it into cleaner, more maintainable components. I built flows that could take a rough feature idea and return a structured implementation approach, including component breakdown and potential edge cases. Over time, these small improvements started to compound.

What stood out to me most was how OpenClaw allowed me to move from isolated interactions with AI to something more continuous and system-driven. It no longer felt like I was using a tool occasionally. It felt like I was designing a workflow where AI was embedded into how I think and build.

There are a few things OpenClaw gets right that are worth highlighting.

The first is ownership. You are not limited to a fixed interface or a rigid set of features. You define how the system behaves and evolves. That level of control is rare and increasingly important.

The second is flexibility. OpenClaw does not assume your use case. Whether you are building developer tools, automating tasks, or experimenting with ideas, it adapts to your needs instead of forcing alignment.

The third is composability. You can layer your workflows, reuse logic, and gradually refine your system. This makes it possible to build something that improves over time rather than starting from scratch repeatedly.

Working with OpenClaw also shifted how I think about personal AI. The real value is not in generating answers. It is in creating systems that support thinking. There is a clear transition happening from one-off AI interactions to persistent, customizable workflows that integrate deeply into how we work.

That said, the experience was not without challenges. Setting up a meaningful workflow required experimentation. Designing reusable prompts took intention. There was a learning curve in figuring out what should be automated and what should remain manual. But those challenges are part of what makes the process valuable. You are not just using a tool, you are shaping it.

In the end, OpenClaw changed how I approach development. It encouraged me to think less about individual tasks and more about systems. Instead of asking how I can solve a problem once, I started asking how I can design a process that solves it repeatedly with less effort.

For developers who are serious about integrating AI into their workflow, OpenClaw is worth exploring. Not because it does everything for you, but because it gives you the foundation to build something that works the way you do.

I am still refining my setup and experimenting with new workflows. The more I use it, the more I see its potential not just as a tool, but as an environment for building smarter ways of working.

I Gave an AI a Body. Here’s What Happened.

Saviour Promise — Sat, 18 Apr 2026 00:09:06 +0000

A personal reckoning with OpenClaw — the open-source assistant that doesn’t just answer, it acts.

I’ve been chasing the “perfect AI assistant” for years.

I tried every chatbot. I built custom GPT wrappers. I wired together n8n automations, pasted prompts into Notion, and set up more Zapier flows than I care to admit. Each time, I’d hit the same wall: the AI would think beautifully, reason eloquently — and then stop. It would hand me a plan and wait for me to execute it.

It was like hiring the world’s smartest consultant, only to find out they don’t have hands.

Then I set up OpenClaw. And something shifted.

The First Thing It Did Without Being Asked

About two hours after I finished the setup, I got a message on Telegram.

My OpenClaw instance — I named it Axiom — had noticed that a GitHub repo I’d been watching released a new version. It had already checked the changelog, summarized the breaking changes, and flagged the two that would affect my specific codebase. Then it asked if I wanted it to open a draft PR with the migration.

I hadn’t asked it to watch that repo. I had mentioned, in passing, during onboarding, that I cared about keeping my dependencies up to date.

It remembered. It acted. It came to me.

That’s when I understood: this isn’t a chatbot. This is something new.

What OpenClaw Actually Is (And Why It Matters)

OpenClaw is a self-hosted, open-source AI agent that runs on your own machine — Mac, Windows, or Linux. You give it access to your tools. It connects to the chat apps you already use (WhatsApp, Telegram, Discord, Slack, iMessage, and 20+ more). And then it stays on, 24/7, with full memory of who you are and what you care about.

The key insight that makes OpenClaw different from everything else:

Most AI tools give you a brain. OpenClaw gives that brain a body.

It can browse the web. It can read and write files. It can run shell commands. It can open your browser, navigate to a site, fill out a form, and come back to tell you it’s done. It doesn’t hand you a script — it runs the script.

And crucially — your data never leaves your machine. No walled garden. No subscription lock-in. No company training on your private emails. You own the context. You own the memory. You own the agent.

The Moment I Stopped Thinking of It as a Tool

Three weeks in, I was deep in a debugging spiral. A race condition in a distributed service. I’d been hunting it for days. I described the problem to Axiom in a Telegram message — casually, the way you’d vent to a colleague — and went to make coffee.

By the time I got back, it had pulled the relevant log files, identified the timing pattern, cross-referenced it with a similar open GitHub issue from a different project, and left me a summary with a proposed fix.

Five minutes. Three days of my life, back.

But the real shift wasn’t productivity. It was psychological. I stopped managing the AI and started working with it. It felt less like using software and more like having a junior engineer who never sleeps, never forgets, and genuinely wants to help.

That’s when the tool became a collaborator.

The Part Nobody Talks About: The Responsibility

I want to be honest here, because the breathless hype around OpenClaw sometimes glosses over something important.

Giving an AI agent access to your files, your email, your shell — that’s a serious decision. OpenClaw’s own community has been candid about this. One of the maintainers has said plainly that if you don’t understand how to run a command line, this project may be too dangerous for you to use safely.

That’s not a criticism of the project. That’s intellectual honesty, and it’s rare.

The power that makes OpenClaw remarkable — its autonomy, its breadth of access, its ability to act without being asked — is exactly what demands careful, deliberate setup. You should use the principle of least privilege. Audit what you give it access to. Run it in a sandboxed environment if you’re cautious.

An AI with hands is only as safe as the boundaries you give it.

Why I Think This Changes Everything

We’ve spent years talking about AI as a productivity tool. Something you open in a tab and close when you’re done.

OpenClaw is a different paradigm entirely. It’s ambient. It’s persistent. It’s yours — not some company’s product you’re renting access to. The context lives on your machine. The skills live in your workspace. The memory is yours to read, edit, and export.

This is what personal computing looked like in 1984 — a fundamental shift in who owns the relationship between human and machine. Back then, the shift was from institutions to individuals. Now, it’s from cloud platforms back to individuals.

OpenClaw isn’t the end of that story. It’s barely the beginning. But it’s the clearest signal I’ve seen that the next era of computing isn’t about smarter chatbots.

It’s about AI that shows up.

Getting Started (If You’re Ready)

If this resonates and you want to try it yourself:

You’ll need Node.js 22+ installed
Run npm install -g openclaw@latest
Then openclaw onboard --install-daemon — this guided flow walks you through everything
Full docs live at docs.openclaw.ai

Start small. Give it one channel. One skill. One task you do every week that you hate. See what happens.

Then come find me and tell me about the first time it did something you didn’t ask for.

Have you set up OpenClaw? What was your “it did something on its own” moment? Drop it in the comments — I read every one.