Architecture Proven for the First Time

OpenClaw monitors Pickaxe chatbot studios — pulling subscriber data, conversation history, feeding it to a model, producing daily audit reports. It was built with a studio loop from the start. Add a studio to the config, it runs for that studio too.

The problem: we had only ever had one studio.

This week we added a second. SketchScript — an AI sketchnoting tool — is the original. A second studio, serving a completely different audience, became studio number two.

The structural part was fast. OpenClaw’s config is driven by a list of studio objects, each with its own credentials, prompt file, and output path. Adding a new studio meant writing a config block, writing a prompt file, and setting one environment variable. No code changes. Twenty minutes.

Then the prompt engineering started.

SketchScript’s analysis prompt asks about image generation quality, style adherence, iteration count, and how often users hit their monthly limit. Those are the right questions for a visual creation tool.

They are completely wrong for a studio that routes users to specialist advisors across twelve categories. The second studio needed different questions: which advisors get traffic, which ones go unused, how effective is the document analysis feature, what patterns show up in conversations from month to month.

The system was built to handle multiple studios generically. What it means to analyse a studio, it turned out, isn’t generic at all.

The structural proof was cheap: config file, prompt file, environment variable. Done. But the intellectual work — deciding what questions to ask about a specific product — takes time regardless of how clean the architecture is.

OpenClaw can ask the same question about any studio. Whether that question is useful depends entirely on whether you asked the right one.

Architecture handles the how. You still have to figure out the what.