Are you using scripts enough in your AI harness?

Are you using scripts enough in your AI harnesses? Here is a recent example.

The need

A client needs to manage post creation in a git-based CMS (for the less technical: articles are stored directly in the codebase, so no external CMS and no database).

The goal: the human writes the article, and the AI takes care of adding it to the CMS and publishing it.

Basic AI harness

We create a skill for an AI agent, which explains how to do things properly from a text:

• fix typos;
• translate into a second language;
• fetch the existing articles to add internal links;
• create the files.

Read that last point again: create the files.

Of course, an LLM will do it right most of the time. Because there are 300 posts next to it in the same format, and it is dumb, repetitive work.

But it is still an intention dictated in a skill prompt, not an inviolable rule. Nothing prevents a nasty surprise on a day the model got up on the wrong side of the bed.

More advanced AI harness

We write a script that handles 100% of the file creation on disk. The LLM sends its texts to the script, which validates the data and creates the files deterministically and almost instantly.

It is both cheaper and more robust.

The LLM only handles its own part: translating, fixing text, spotting good links between articles. Then it delegates the rest to the script.

This strategy works very, very well as soon as you run into something the AI does repetitively, and that must be strictly identical every time.

In Böckeler’s terms, this is sensor feedforward in a good AI harness.

As always, it starts from an identified need, and must be customised for each project.

Scripts are only one part of the harness: the codebase also has to be ready for it.