I Rewrote My Entire Workflow in Two Weeks

By Suraj Kumar · Mar 9, 2026 · 12 min read · 0 views · 0 likes

Week one: I removed the desk from coding. Week two: I removed the apps from everything else.

Two weeks ago I was a normal developer. Laptop on desk. Tabs open. Notifications pulling me between six apps. The usual.

Today my laptop is optional for coding, I talk to a single AI agent on Telegram for everything else, and my decisions get published as web pages on my own domain. Not gradually. Two weeks.

Here's what happened.

Week 1
Desk-bound coding
Code from phone
Agent Orchestrator
+ Mobile App
+
Week 2
6 apps, context-switching
1 Telegram chat
OpenClaw
+ Personal Agent
2Weeks
6+Apps replaced
0Desk required
1Telegram chat
· · ·
Week 1 — Remove the Desk

I Code From My Phone Now

I'm a contributor to Agent Orchestrator, an open-source system by Composio for managing 20+ parallel AI coding agents on your codebase. AO is powerful, but it had a bottleneck: me. Every time an agent needed input — an approval, a merge decision, a stuck session — I had to be at my laptop.

Multiple agents running in parallel, and a human at a desk as the chokepoint.

App closed, phone in pocket, walking somewhere — buzz: "Session needs input." I handle it from my phone. Agent continues. I keep walking.

The orchestrator removed the engineering bottleneck. Mobile removed the desk. What's left is judgment — from wherever you are.

I wrote about this in detail — the PWA dead end, the native app build, the architecture. After a week of using it, coding from my desk felt like going back to a flip phone.

But something nagged me. I'd removed the desk from coding. Everything else in my life still required six apps and constant context-switching.

· · ·
Week 2 — Remove the Apps

One Agent for Everything

I kept seeing people on Twitter offloading real tasks to personal AI agents. Not demos. Actual workflows: parsing bank statements, publishing research, writing in their own voice. They were getting tangibly more productive.

I set up OpenClaw. $13/month. 45 minutes to install. Connected it to Telegram, authenticated with Codex and Claude, enabled memory search, pointed a domain at it, and configured three tiers of web publishing.

The first day it was running, I told it: "Revamp my portfolio. Clean, minimal, dark theme."

Ten minutes later it had read my old repo, studied the design I wanted, built the entire page with inline CSS, and published it to my domain. Done.

My last portfolio took a full Saturday. Figma, CSS tweaks, deploy to Vercel, check mobile, redeploy. Eight hours.

This time: one Telegram message while making coffee.

What Actually Got Replaced

TaskBeforeAfter
NotesApple Notes / Google KeepCaptured to inbox.md
ResearchChatGPT & Claude (separate tabs)Living HTML pages on my domain
Small appsCursorAgent builds them
HostingVercelCaddy on same VPS
Deep researchGoogle + manual synthesisFull reports with data & sources
Social postsTwitter draft modeAgent drafts in my voice

The biggest shift isn't any single feature. It's having one place to dump ideas and explore every possibility. Before, a thought would either die in my head or cost me twenty minutes of context-switching across apps. Now I message my agent. It captures, researches, or builds — whatever the thought needs.

The Research Problem, Solved

When I ask ChatGPT to research something, I get a summary. Useful in the moment. I'll never find it again.

When I ask my agent, I get the research itself — a full HTML page with tables, comparisons, source links, analysis. On my domain. With a URL I can share. That I can revisit in six months. That the agent updates when things change.

The document IS the research. It doesn't describe findings — it contains them. I don't get a summary I'll forget. I get an artifact I'll use. This idea comes from Prateek Karnal's living documents setup.
🌐 Public anyone on the internet
Articles, comparisons, tech research, open-source guides
🔒 Shared password protected
Trip plans, group decisions, family logistics
🛡 Private tailscale only
Finances, health data, investment analysis, personal logs

Every significant decision or research task now produces a URL. Decisions stop being ephemeral. They become artifacts.

The Personality Layer

This surprised me the most.

I started using the agent to draft tweets and LinkedIn posts. At first — generic. The usual AI voice everyone scrolls past. But OpenClaw has SKILL.md files that teach the agent how to behave for specific tasks. I wrote one describing my style: punchy, data-forward, tables over prose, narrative-driven, minimal jargon.

How the agent learns my voice
SOUL.mdvalues + boundaries
SKILL.mdwriting style rules
Memoryevery conversation
Outputsounds like me
↻ feedback loop — every interaction improves the next

After a week, the agent isn't just following the style guide. It's learning my actual patterns — what I emphasize, how I structure arguments, which phrases I use, what I'd never say. The memory system builds context with every conversation.

Now when I say "draft a post about my workflow changes," it writes something that sounds like me. Not perfect. But close. And getting closer every day.

This is the real unlock. Not a smart chatbot. An assistant that's mine.

A Day in the New Workflow

Morning — walking
Phone buzzes: AO agent needs merge approval. I approve from the mobile app. Agent continues. I keep walking.
Mid-morning — coffee shop
"Remember to review the Codex integration PR today" → captured to inbox.md. Zero context switch.
Afternoon — couch
"Compare co-working spaces in Koramangala and publish it" → full comparison page live at surajmarkup.in within minutes.
Afternoon — still couch
AO session stuck on a failing test. Kill it from phone, reassign. Back to reading.
Evening
"Draft a tweet about this two-week workflow change" → agent writes it in my voice. I edit one line. Post.
Night
"Summarize today" → journal entry written. Tomorrow the agent has full context.

Blue dots = AO mobile app. Green dots = OpenClaw personal agent.

No laptop opened. No desk visited. Every task handled from my phone through two systems: AO for coding, OpenClaw for everything else.

· · ·

The Two Systems

Complete Architecture
📱
My Phone
AO Mobile App
Telegram
Coding
Agent Orchestrator
20+ Coding Agents
Git Worktrees + tmux
Everything Else
OpenClaw Gateway
Codex / Claude
Flat Files + Memory
PRs, Merges, Fixes code shipped without a desk
surajmarkup.in living documents, research, posts

Both run 24/7. One manages coding agents. The other manages me. Together they cover everything I used to need a laptop and six apps for.

ComponentCost
Hetzner VPS (OpenClaw)$13/mo
LLMs (Codex + Claude)Sponsored 😂
Embeddings~$0.01/mo
Domain~$3/yr
AO Mobile AppFree (I built it)

What's Next

The obvious next step: connect the two systems. I'm adding Agent Orchestrator as an OpenClaw plugin — so I can triage AO issues, monitor builds, and fix bugs through the same Telegram chat that handles everything else. One interface for coding and life.

Beyond that: Tailscale for zero-trust access. Sandboxed agents in WhatsApp groups. Cron jobs for daily briefings. Voice notes that get transcribed and acted on.

Two weeks ago I was a normal developer. Six apps, one desk, constant context-switching. Now I have two AI systems running my workflow — one for code, one for everything else — and I control both from my phone.

The foundation took two weeks. Everything else is just adding skills.

This personal agent setup is inspired by Prateek Karnal's agentic infrastructure — his living documents, 3-tier publishing model, and the idea that every decision should have a URL.

Want to build this? I wrote a complete step-by-step setup guide for the OpenClaw side — every command, every config, every gotcha. Read the full technical walkthrough →

ai-agents openclaw agent-orchestrator productivity self-hosting mobile-dev