Start here

deeper is a short set of instructions for an AI assistant. Type /deeper before a request, and the assistant rewrites the request the way an expert in that field would ask for it: what they'd do, what they'd start from, and how they'd check the result.

It doesn't do the task. It shows you the better request and waits for you to say go, so you can fix a wrong assumption before any work starts.

Pick the way you already talk to an AI

In a chat app

The skill is plain text you paste in, so it works on a phone too. Chat apps aren't tested yet, so treat the reply below as what should happen, and tell us how it went.

  1. Copy the skill. It's one text file of about 700 words.

    Read it before you paste it
    ---
name: deeper
description: Rewrite a task into its deeper version (same scope, more expert) before doing it. Use only when the user types /deeper, with or without a task, or asks for "the deeper version" of a task. Do not use it on your own initiative.
license: Apache-2.0
metadata:
  version: "1.1.0"
  source: "Felix Rieseberg, How to be more ambitious: A tutorial for humans with models (13 October 2026), Depth section"
---

# /deeper: a deeper version of the task, not a bigger one

Adapted from the Depth section of Felix Rieseberg, "How to be more ambitious: A tutorial for
humans with models", 13 October 2026:
https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/

The task is whatever the user wrote with /deeper. If they wrote nothing, use the task currently
under discussion. If there is none, ask for the task in one line and stop.

If the task is too small for depth to change anything (a status question, a one-line fix), say so
in one line and stop.

Keep the scope the same size. Deeper means more expert, not more features. Don't start the work.
First answer these four, briefly and specifically for this task:

1. **What would a top professional in this field do here that I haven't asked for?** Name the
   field. Give the one or two moves that separate expert work from competent work, not a list.
2. **What is the rawest material they would start from?** Not the summary, template or digest
   I'd reach for, but the primary source underneath it (the data, records, recordings, originals,
   the people). If you can read files, check the workspace first. Say what's here, where it is,
   and what's missing. If you can't, ask me for it. Name one thing we take as true that this
   material could disprove.
3. **Which methods and tools would they use?** Name the instruments themselves, for example a
   database and a statistical model rather than a better-looking table. Models tend to change
   their vocabulary without changing their method. Say which of these tools are already here.
   Don't swap in a weaker tool just because it's to hand.
   Then the skills and agents for the field, if your environment has them. List the installed
   skills that fit (for example `.agents/skills`, `~/.claude/skills`, `~/.codex/skills`,
   `~/.cursor/skills`, enabled plugins) and the subagent types on offer. If you can run shell
   commands, offer to search the open skills ecosystem with `npx skills find <field>`, which
   downloads and runs the skills CLI from npm, and run it only if I say yes. Then name each result
   as installed or only findable, with its source and install count. If you can't run commands,
   say so rather than guessing. Ask before downloading or installing anything.
4. **What, other than you or me, would show the result is right?** A test, a measurement, a
   tool, a real audience, a domain expert. Prefer something that can be run straight away. A
   model grading its own work doesn't count. Set the pass mark where the best in the field would
   set it, not where it's easy to reach, and say what result would count as a fail.

Then rewrite my task as a prompt that asks for exactly that, naming the skills and agents to
use, in a fenced block I can copy.

**Same deliverables.** Before writing the prompt, list what my task asks me to end up with (one
caption; a plan; a fixed bug). The prompt asks for those and nothing else. Depth changes how
each one is made and checked, never how many there are, so keep every method, source and check
from your four answers; don't thin them to stay short. A check stays in the prompt whenever it
tests a listed deliverable, however much work it takes to run: a measurement, a preflight, a
printed proof, a load test, an expert's review. It only counts as a new deliverable if it leaves
me something extra to keep or maintain: a second version, a new tool or script, a document
beyond the one I asked for, or a process that keeps running after the task ends. Anything else a
top professional would add goes after the fenced block, on one line starting "Worth knowing, not
in this task:", for me to choose. Leave that line out if there's nothing to add.

End by asking whether to run it.
  2. Paste it into a new chat. On the next line, type /deeper and your request, then send. For example: /deeper help me write a cover letter for a junior data analyst job
  3. Read the reply. It should give four short answers, then a rewritten request in a box, then ask whether to run it. Anything extra an expert would do goes on one line, "Worth knowing, not in this task", for you to choose. It can't see your files from a chat, so it asks you to paste what it needs, such as the job ad and your CV.
  4. Say yes, or edit the rewritten request first.

Keep it so you don't paste it every time

In a coding agent

You need:

  1. In a terminal, in your project folder, run:

    npx skills add EazyAccessEA/deeper

    It asks which agents to install into. Add -g to install it for every project, not just this one.

  2. Check it's there: npx skills list shows deeper. In Claude Code, type / and deeper is in the list of commands.
  3. Type /deeper and a real task you were about to ask anyway. If /deeper isn't recognised, ask for the deeper version of the task in words. The skill is written to load when you do.
  4. Read the four answers and the prompt, then say whether to run it.

Running deeper is tested in Claude Code, Cursor and Codex.

If /deeper doesn't appear

Still stuck? Open an issue or email hello@deeper.tools.

After the first run

Give it tasks where getting the method wrong would cost you: a migration, a purchase, a piece of writing that has to land. For a status question or a one-line fix it says depth won't help, and stops.

The docs cover updating and pinning a version for a team. Shallow patterns shows what it guards against, and Research shows how it was tested.