I Ran OpenClaw for 30 Days — Lessons From a Retired Runtime

Status · August 1, 2026: Historical operating notes. OpenClaw is not my current runtime. Use the principles, verify current documentation, and do not treat old commands as a live setup guide.

When I first set up OpenClaw, I made the classic mistake: I installed it, opened Telegram, typed "hello," and expected magic.

I got a capable AI that knew nothing about me, had no memory of previous conversations, no access to my tools, and answered questions with the same generic helpfulness as any free chatbot.

The tool was powerful. I had set it up like a toy.

At the 30-day checkpoint — after wiring it to Todoist, adding persistent memory, and setting up scheduled jobs — I understood the useful part: an agent only becomes operational when context, tools, verification, and boundaries are designed deliberately.

Here's what I wish someone had told me before I started.

What Is OpenClaw, Actually?

Most people describe OpenClaw as a "Claude chatbot on Telegram." That's like describing a scalpel as "a very thin knife." Technically accurate. Completely misses the point.

OpenClaw is a gateway daemon — a process that runs on your machine 24/7, connects your AI model (Claude, Gemini, or others) to your messaging platforms, gives it access to tools (shell commands, web search, browser control, file I/O), and lets it operate autonomously with persistent memory and scheduled tasks. Once you've learned the 30-day lessons here, my full architecture writeup goes much deeper on cron jobs, model routing, and sub-agent patterns.

The Telegram interface is just how you talk to it. The real capability is everything running underneath.

Top 5 Things to Do When You First Set Up OpenClaw

#1: Write Your Identity Files Before You Send a Single Message

The single highest-leverage action in OpenClaw isn't installing a plugin or connecting an API. It's writing three text files that tell your agent who you are.

OpenClaw loads files from your workspace directory at the start of every session. These files are your agent's long-term memory. Without them, every conversation starts from scratch.

The three files you need immediately:

AGENTS.md — Identity and operating rules. Who the agent is, who you are, your active projects, what it's NEVER allowed to do autonomously, and what to prioritize.
USER.md — A profile of you. Location, profession, communication preferences, schedule, expertise areas, goals.
SESSION-STATE.md — The handover document. Updated before each context compaction or session end. This is how your agent picks up exactly where you left off.

Write these three files before your first real conversation. 30 minutes of setup unlocks weeks of coherent context.

#2: Set a Hard Autonomous Spend Limit

OpenClaw has access to the shell on your computer. It can run commands, access the web, write files, and — if you give it the keys — send messages, make purchases, and push code to production.

Before you start, be explicit about what it cannot do without your approval. In your AGENTS.md:

## Safety defaults
- $0 autonomous spend limit — never purchase without approval
- Never push to production or submit to app stores autonomously
- Never delete files or run destructive commands
- Confirm before sending any external message to a third party

This isn't paranoia — it's operating procedure. The rule of thumb: anything irreversible requires your explicit approval.

#3: Build a Cron Job That Works While You Sleep

OpenClaw's cron system lets you schedule jobs at recurring intervals without you initiating anything. Your agent wakes up, executes the task, writes output to a file, and sends you a summary.

The most valuable cron job in that OpenClaw period was a bounded morning review that selected tasks and sent a concise messaging summary.

Other useful cron ideas:

Daily journal writer — Logs what was discussed and decided each day to a dated memory file
Morning brief — Weather, top tasks, news, and a Bible verse (mine pulls live METAR data from FAA — useful when you're a commercial pilot tracking flying conditions)
Weekly review digest — Compiles progress across all active projects every Sunday night

Key to good cron jobs: write the system prompt carefully, include explicit safety bounds, and tell it exactly where to write output.

#4: Connect Real Tools — Not Just Chat

Fresh out of the box, OpenClaw is a chat interface. The real upgrade is connecting tools. Priority order:

Web search (Brave/Tavily) — Essential. Without it, your agent has no access to current information.
Todoist (or task manager) — Highest-ROI integration. Your agent can read tasks, add comments, create subtasks, and triage priorities.
File system — Already enabled. Make sure your agent knows its workspace structure.
Browser control — The Camoufox integration lets your agent browse real pages and scrape data that web_fetch can't handle.

#5: Build a Memory System on Day One

The agents that stay useful long-term are backed by persistent memory — not blind trust in chat history. My setup during that period was:

memory/
├── YYYY-MM-DD.md     ← daily journal (written every evening)
├── LEARNINGS.md      ← rules from mistakes (the most important file)
├── [project].md      ← one doc per active project
└── reference/        ← deep research, competitive analysis

The most valuable file is LEARNINGS.md — a running log of mistakes the agent made and the rules derived from them. Every time the agent does something wrong, I write the correction as a rule. Over 30 days, this file becomes the institutional knowledge that makes the agent dramatically more accurate.

Dos and Don'ts

✅ Do

Give your agent a name and character — it improves consistency
Write every important decision to a file — chat is ephemeral, files are permanent
Use sub-agents for long tasks — orchestrate, don't execute everything inline
Update SESSION-STATE.md before any significant break
Set your $0 autonomous spend rule in writing

❌ Don't

Ask it to "just handle everything" — vague instructions, vague results
Give it production credentials on day one
Skip the memory files — every unlogged conversation is lost context
Treat it like a chatbot — it's an agent, the model is "set context → assign task → review output"
Run cron jobs without clear safety bounds in the prompt

The Honest Assessment After 30 Days in 2026

At that checkpoint, OpenClaw behaved like infrastructure rather than a productivity hack.

The first week, I spent more time configuring it than I saved. The second week, it started doing useful things. By week three, it was running cron jobs every morning, managing my Todoist, drafting content, and keeping memory that made every conversation materially better than the last.

The ceiling was high, but the floor was low. The durable lesson was not “install this runtime.” It was to define identity, memory, safety gates, and one useful scheduled job before expanding.

Historical takeaway: start with one bounded workflow, verify every external action, and judge the system by useful output rather than agent theater.

2026 Setup Checklist, Preserved as Historical Reference

Write AGENTS.md (identity, projects, safety rules)
Write USER.md (who you are, schedule, preferences)
Write SESSION-STATE.md (current context, handover doc)
Add Brave or Tavily API key for web search
Connect Todoist (or preferred task manager)
Set up one cron job (Morning Taskmaster or daily journal)
Start LEARNINGS.md (even just two entries)
Set $0 autonomous spend limit explicitly in AGENTS.md

These eight items capture what mattered during that 2026 experiment. They are not a current OpenClaw installation guide. The AI side-income post explains the current Hermes-based stack and the paper-only status of the weather-market research.

Nick Rae writes about aviation, AI tools, and building software products as a one-person operation. He's the author of the Private Pilot License Workbook and the developer of FeedFare. Follow the build publicly at Flight Funded.