I use Claude Code all day, in the desktop app, and not only to write code. It has opened sessions in 131 project folders, it runs 21 scheduled tasks, it writes to my sites' CMSs through MCP servers, and it follows in-house skills for prospecting.
What makes this reliable is not the model: it is the written rules each project imposes on it.
01The workstation
The base is Claude Code in the desktop app, with a few plugins installed:
Most of the work does not go through these plugins, though, but through MCP servers and scheduled tasks.
02The MCP servers that do the real work
An MCP server gives the agent direct access to a tool, with no copy and paste. The ones I use most:
The Search Console and Bing server comes from the Klickbee Tracking repository: it connects a new site to both consoles in a single request.
03Scheduled tasks
SEO production for five sites, this one included, runs on its own:
Each task has strict instructions. For example, the agent never reads the backlog state directly: it goes through a script that only gives it what it needs. And it is only allowed to write in the drafts folder.
04Written rules matter more than the model
I have no global instructions file. Rules live in each project, where the agent reads them:
That last point is what lets me have an agent publish under my name. The model can be wrong, the script does not let it through. I listed the most common traps in my article on Claude Code mistakes.
05In-house skills
A skill is a written procedure the agent follows when the task fits. I have written several for Klickbee: preparing an action plan for a prospect, writing the brief for a discovery call, auditing the tracking of a list of online stores, building a Google Ads Search campaign, reviewing a cold-calling session, finding and summarising Reddit discussions.
Their value is repeatability: the same request produces the same deliverable, in the same format, with the same checks.
06What I do not hand over
The system is built to keep some things on my side:
For a wider view of what you can build with it, my notes on a SaaS coded in 24 hours show the other side: pure development. The starting point remains my guide to Claude Code.
07FAQ
Can you let Claude Code publish on its own?
Yes, provided every step is framed: a script that provides the state, limited write access, and a quality check that blocks anything breaking the rules. Without an automatic check, I would not do it.
What is the difference between an MCP server and a skill?
An MCP server gives access to a tool: a CMS, a console, a browser. A skill describes a procedure: which steps, in what order, with what deliverable. The two are often combined.
Do you need a global instructions file?
Not necessarily. I prefer rules per project, next to the code and content they apply to. They stay up to date more easily and the agent only reads what concerns it.
08Sources
For me, Claude Code does the work that repeats, framed by written rules and automatic checks. Lived experience, decisions and anything sent out stay on my side.
Talk about your project ↗