Claude, Cookbook Audit

Does auditing a cookbook notebook mean Claude is exercising some general sense of editorial quality? No. This video walks through the cookbook-audit skill to show that it runs a written rubric against a notebook, the same way every time, rather than judging the notebook by feel.

What the skill folder contains

The cookbook-audit skill is a folder Claude reads before it works, not something it is trained on. It holds three files: SKILL.md, style_guide.md, and validate_notebook.py, together forming a roughly 12-kilobyte instruction set written in plain language with no hidden logic. Claude reads the file and then acts on it. As the video puts it, the file is the program.

Three steps, run in order

The instructions sit in a Steps section. Claude reads SKILL.md, executes each step in order, and returns a result, linearly, with no branching unless a step calls for it. The process the video walks through is: hand the skill a notebook, it reads the rubric, checks the notebook against that rubric, and returns a score. That's the same three steps every time, regardless of which notebook is handed in.

Same input, same output

The video demonstrates the consistency this produces: hand cookbook-audit the identical notebook twice, and it returns the identical rubric score both times. That is the payoff of the file being the actual program running the check, rather than a loose set of guidelines interpreted freshly each time.

What happens outside the rubric's scope

The reverse case matters just as much. Hand the skill a notebook with something outside the stated rubric, and it still runs the same steps, checking only what SKILL.md names. validate_notebook.py is the check for structure, rubric items, and scope, and all of it is caught the same way regardless of what else the notebook contains. Auditing a notebook this way is not Claude forming its own opinion of quality, it is SKILL.md's rubric being read, applied, and checked consistently against whatever notebook is handed over.

The prompt to try

The video's suggested prompt is to ask Claude to audit an Anthropic cookbook notebook against a rubric, to read the cookbook-audit skill first, and to walk through what it will do before doing it. The video emphasizes watching for that walk-through specifically, since explaining first is what surfaces which rubric items SKILL.md actually checks and where the stated scope stops, rather than only seeing the finished score at the end.

Key takeaways

• The cookbook-audit skill is a folder with SKILL.md, style_guide.md, and validate_notebook.py, not a training process.
• Claude reads the steps and executes them in order, then validate_notebook.py checks the result against a rubric.
• The same notebook run through the skill twice returns the identical score both times.
• A notebook with something outside the stated rubric is still checked only against what SKILL.md names, not judged more broadly.
• Auditing this way is not Claude forming an independent opinion of quality; it is a fixed rubric applied consistently.

Who this is for

This is for anyone building or using a Claude skill meant to check work against a rubric, and for anyone wondering whether a Claude audit reflects genuine judgment or a fixed, repeatable check.