The runbooks

6.14[BONUS]

A deeper version, not a bigger one

Go deeper is a skill that asks four questions about a task before it starts, then rewrites what you asked.

The questions are Felix Rieseberg’s, an engineering manager at Anthropic who works on Claude, and they have it ask for what a top professional would start from rather than what you thought to give it.

You“Go deeper on this”

Claude asks

What a top professional would do

What they would start from

How they would work it out

What would prove it right

ClaudeWrites your ask again, sharper

YouRead it and say go

You need it because telling Claude it is an expert changes how the answer sounds, not how good it is.

Better questions are what make it better, and it is worth it on a new offer, a hire, a price rise or a proposal, not on a quick email.

Paste this into a chat in your ecosystem folder and it writes the skill.

Then say “go deeper on this” before any task that matters, and read what it rewrites before you let it start.

Set up go deeper

Write me a skill called go-deeper, for when I want a deeper version of a task rather than a bigger one. Write its description as the sentences I actually say, like "go deeper on this" and "give me the deeper version", and say it is for work where the thinking matters, never a quick job. Put this prompt in it word for word, from Felix Rieseberg's essay: "I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I'm not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that." When it runs, it answers the four questions for my task in plain sentences, names any files in this folder it would want to read without reading them yet, writes the rewritten request in full, then asks me what to take out and waits for my go before it starts anything. Show me the file.

The prompt is Felix Rieseberg’s, from How to be more ambitious, a tutorial for humans with models, October 2026. The study is Hu, Rostami and Thomason, Expert Personas Improve LLM Alignment but Damage Accuracy, 2026, where accuracy fell from 71.6 percent with no expert description to 68.0 with a short one and 66.3 with a long one, on small open models rather than Claude.