Put your own agent to work/
Over two live sessions, we’ll set up OpenClaw and give it a useful job. I’ll show you how I write instructions my agent can reuse, choose skills for the task, and check the result before relying on it.
Bring a recurring task from your own work. We’ll use it as a starting point, and work through what to do when the agent gets stuck.
- 2 × 2.5 h
- live online sessions
- Small group
- with time for questions
Hands-on workshop · Claude Pro required · Preparation and costs
The weekly report arrived, but the figures are missing. Find out why before changing anything.
The job ran, but its data request failed. I’ll check the connection and show you what I find.
Once we’ve tested the fix, update the handbook so we know what to check next time.
Start with one job worth repeating/
The aim is to leave with a job you’ve tested, not a long list of tools you’ve installed. We’ll build one scheduled task together and use a few examples to practise giving instructions and checking results.
A report you actually read
Decide what belongs in a recurring update, where its information comes from and where it should arrive. Check the received message, not just whether the schedule ran.
A procedure you don’t have to explain again
Write down a familiar task, then ask the agent to find and use those instructions in a fresh conversation.
A problem you can investigate
Work through a failed task: inspect the evidence, test a likely cause and check the fix before saving it as a procedure.
We’ll choose a small task that fits the session. Connecting all your business accounts or automating a complex workflow is outside the two-session scope.
Give your agent instructions it can reuse/
OpenClaw already has memory. The handbook adds something you can maintain and check yourself: how your setup works, why you made a decision, and the steps for jobs you do regularly.
We’ll write one useful procedure and show the agent where to find it. Then we’ll start a fresh conversation and see whether it consults the right instructions. If it doesn’t, we’ll improve the handbook rather than assume it remembered.
When a problem is fixed and verified, the relevant page gets updated. The next attempt starts with a tested procedure instead of another explanation from you.
Inside the handbook exercise
We’ll create a short index, a procedure with an expected result, and a concise pointer in AGENTS.md. Existing instructions stay intact. The agent checks whether a procedure still applies before following it.
Secrets stay out of the handbook. Record where credentials are managed, never their values, and keep private project details available only to the people who need them.
Know when the work is actually done/
A skill gives the agent a method for a particular kind of work. I’ll walk through examples from my own setup and explain why I keep them, when they help and what still needs checking.
We’ll use debugging to separate a symptom from its cause, verification to check the result rather than trust a completion message, and a backup procedure to distinguish “an archive exists” from “we can recover the data”.
The useful lesson is how to choose a method and judge its result. Installing more skills isn’t the goal.
The case studies and troubleshooting exercise
- Debugging: reproduce a failure, inspect the evidence and test one explanation at a time.
- Verification: compare an agent’s claim with the actual output or published result.
- Backups: inspect what was saved and distinguish an integrity check from a full recovery rehearsal.
We’ll also inspect where a skill comes from and what it asks the agent to do. Details specific to your setup belong in the handbook; a method useful across different jobs may belong in a skill. Untested guesses remain unresolved, not permanent instructions.
Set it up, then put it to work/
Session 1 · Get it running
Connect to your server, set up OpenClaw with your Claude account and send the agent a message through Telegram. We’ll explain the access and privacy choices as we go.
Checkpoint: you can reach your own agent and understand which parts run on your server and which use outside services.
Session 2 · Make it useful
Write and test your handbook, run a small scheduled job and work through a failure. I’ll use the skills and procedures from my own work to explain what to check.
Checkpoint: you’ve checked the job’s result, tested the handbook in a fresh chat and know how to begin investigating a problem.
Each session is 2.5 hours, including a break. Dates and times will be announced to the waitlist.
Detailed curriculum
Session 1
Server login and recovery access; private networking; OpenClaw installation and onboarding; Claude Code subscription login; Telegram pairing; a security checkpoint and preparation for the next session.
Session 2
Handbook structure and agent instructions; a fresh-chat test; skill walkthroughs; a troubleshooting exercise; a scheduled job with delivery checks; subscription limits, backup checks and next steps.
You’ll have copyable examples and procedures to refer back to. We’ll keep optional topics separate rather than try to fit every integration into the workshop.
Preparation and costs/
You don’t need to be a developer, but you should be comfortable copying terminal commands and editing a text file. This is a hands-on technical workshop, not an introduction to chatting with AI.
- A laptop and Telegram. You need to be able to install tools and connect to a server.
- Your own Claude Pro subscription. Have it active before Session 1 and be able to sign in. The free plan is not enough for this workshop. An existing Max plan is also accepted.
- A small VPS. Budget from around €9 a month; actual prices depend on the provider. We’ll send server preparation instructions before the workshop.
There are three separate costs: the workshop fee, your server and your Claude subscription. Workshop pricing and dates are not set yet. Joining the waitlist is not a booking.
Claude subscription limits apply. The standard workshop path uses Claude Code through OpenClaw’s Claude CLI connection, not separately billed API access. General Anthropic API usage is not included in Pro. We won’t ask you to share your login or buy API credits for this path.
Control of the setup, responsibility for it too/
I run my own agent because I want to choose the tools it connects to and keep the instructions and project files in a setup I maintain. A hosted assistant is easier to start with; self-hosting also gives me updates, access and backups to look after.
It doesn’t make everything local. Claude and Telegram process the information you send them, and model-provider limits still apply. We’ll make those boundaries clear before connecting anything sensitive.
How the pieces fit together
OpenClaw runs on your VPS. You message it in Telegram, and it uses Claude Code with your subscription for model access. Administrative access stays private, with Tailscale and a gateway listening on loopback rather than a publicly exposed dashboard.
The workshop uses this one path so we can work through it together. Other models, local inference and team arrangements are possible follow-up topics, not additional setups promised within these sessions.
Hi, I’m Ágúst/
I use agents for website work, document search, internal tools and running my own products. The workshop comes from that day-to-day work.
I’ll show you the instructions and skills behind the tasks, including the checks I use when an agent says it’s finished. The handbook matters because I don’t want to solve the same problem from scratch next week.
Before you join/
What if I get stuck?
We’ll work through checkpoints together and make time for questions. Follow the preparation instructions before Session 1 so the live time can focus on the agent rather than account setup. A complex or provider-specific issue may need follow-up outside the sessions.
What if I miss a session?
The plan is to record both sessions for participants. Recording arrangements and participant privacy will be confirmed before booking opens.
Can I use another model instead?
For the hands-on sessions, bring Claude Pro or an existing Max subscription. Other supported providers and local models are options to explore afterwards, but we won’t run several installation paths in parallel.
Can my team share an agent?
That is a possible next step, not part of the main workshop build. Shared access requires a shared trust boundary; different agent names alone do not securely isolate private work. Teams and schools can ask about a separate cohort.
Why not just follow the documentation?
You can. This workshop is for people who want to work through the setup with someone, bring their own questions and see how a handbook, skills and troubleshooting fit into actual work.
Bring a job you’d like help with/
Join the waitlist for workshop dates, pricing and preparation details. Tell me what you’d like your agent to do.
Join the waitlist →You’ll receive a confirmation email and can unsubscribe at any time. For a team or school, ask about a private cohort.