Nerd Mode for DeepSeek Harness

Learning in a chat usually ends when the answer arrives: nothing is checked, and nothing survives the conversation. Nerd Mode changes what dsh does with a question. It teaches in steps, it stops where the choice matters, and it can keep a dated record, so a subject lasts weeks instead of one sitting.

Requirements

• DeepSeek Harness with a profile.
• Node.js 22.19 or newer.
• Windows, macOS, or Linux.

Install

From npm:

dsh plugin --profile <profile> add dsh-nerd-mode

From GitHub:

dsh plugin --profile <profile> add github:bocchii-the-code/dsh-nerd-mode

From a local folder or tarball:

dsh plugin --profile <profile> add C:\path\to\dsh-nerd-mode

Restart dsh. Every mode gains the /nerd-mode command and a Nerd Mode pill in the composer bar. The mode picker gains no new entry: the plugin mounts at profile level, not as a preset.

Usage

Two skills, two sizes. Both teach one thing at a time and both stop to ask. They differ in what they keep.

Quick lesson: /study

What it does. Walks one concept or one task from the first step to the last, and stops once at the fork: the point where the choice changes the result. For a concept it explains in layers, from intuition to terms to formal detail. For a task it produces the artifact and writes one line of reason beside each step. It reads no file and writes no file, so it works in any folder and leaves nothing behind.

When to use it. A question you want to understand now and will probably not return to.

• Reading a paper and a term stops you.
• A homework problem you have to finish tonight.
• A file whose logic you need to follow before you change it.
• A small decision where you want the trade-off spelled out before you pick.

How to use it.

/study why does Adam divide by 1 - beta^t
/study walk me through the preconditioner in this file        # with code
/study help me finish problem 3, and explain each step        # with homework

Type the command, or turn the switch on and just ask. With the switch on, a learning-shaped request gets this shape by default.

What you get. A title, the central idea, numbered steps, and a symbol table when the symbols get dense. One worked example, the common mistake, and the single stop at the fork. For a task, the finished artifact with its reasoning.

What it does not do. No file appears, and nothing is scored. None of it survives the conversation.

Long lesson: /study-plus

What it does. Builds a concept graph for one topic, then teaches one concept per session against a record of what you have already demonstrated. The record is four markdown files in the workspace. Every session reads them before the first line, updates them at the end, and dates every claim about what you know.

When to use it. A subject you will still be learning next month.

• You are working through a textbook, a course, or a paper properly.
• You want to know what to study next, and a chat only answers what you ask.
• You want evidence of progress, not an impression of it.
• You keep looking things up and forgetting them.

How to use it. First session, later sessions, and the target.

/study-plus Shampoo and Adam optimizers              # first session builds the graph
/study-plus                                           # later sessions continue the topic
/study-plus I want to reach the matrix square root    # set a target

First session: it proposes 15 to 30 concepts with prerequisite links and waits. You approve or edit one row, then name what you already know well enough to skip. Then it teaches one concept in five steps: recall one earlier concept cold, teach to completion, three unseen questions, a teach-back, and a dated record.

Later sessions read the four files first, so the first line names the next concept, the reason, and the date of the evidence behind it.

What you get. A concept order, one concept taught to completion, a graded check, and a record that survives restarts.

What it does not do. It schedules no reviews and never guesses that you know something. Only an answer you produced changes the record.

What each one leaves behind

	/study, quick lesson	/study-plus, long lesson
Writes files	none	study/<topic>/: graph.md, state.md, log.md, plan.md
Reads files	none	those four, before the first line
Remembered next session	no	yes, with dates
Graded	no	pass and miss, rubric written first
Typical span	one sitting, one idea	15 to 30 concepts over weeks

If a study/ folder appears in your workspace, that was /study-plus. /study never creates one.

Which one

• A question you will not return to: /study.
• A subject you will return to for weeks: /study-plus.
• Started with /study and keep coming back? Move up. Build the graph once, and the earlier conversation still informs it.

An example. You are reading the Shampoo paper and hit the preconditioner. /study what does the Shampoo preconditioner actually do settles it in one sitting. If you decide to read the paper properly over a month, open /study-plus Shampoo and Adam optimizers and let it order the concepts first.

Commands

/nerd-mode on           teaching posture for this conversation
/nerd-mode off
/study <concept or task>
/study-plus <topic>     recorded session with the four study files

The pill in the composer bar sends /nerd-mode. It sits left of Full Throttle, stays visible in both states, and goes dim when off.

What the switch does

• It changes text, not tools. The tool set and the prompt cache stay stable.
• Teaching applies to learning-shaped requests only: explain a concept, tutor homework, walk through code, prepare for an exam. A feature request, a search, a config edit, or a command runs as usual.
• The switch is per conversation. Its state lives in the session log, so compaction and resume keep it.
• If the command fails, a toast at the top of the window names the reason, and the pill keeps its state.
• Subagents run their own session, so each starts with the switch off.
• It cannot delete a base mode's persona text. It outranks that text by declaration.

Skills included

• study: the guide. No file, no grade.
• study-plus: the recorded session. Five steps, concept graph, evidence log.
• diagram: one SVG figure per idea. Written for this package.
• answer-me-with-html: one-page HTML explainers, by QingYunA.
• teach: a dated teaching workspace, by Matt Pocock.
• writing-beats: beat-by-beat article structure, by Matt Pocock.
• humanizer: rewrite AI-sounding prose, by Siqi Chen.
• convert-documents-to-markdown: Word, PDF, and ebook conversion, by firecrawl.

Notes

• The plugin loads its own skills/ folder from cordis.patch.yml with a !!js expression. Review patch changes as host-process code.
• If you rename the package, update the resolve('dsh-nerd-mode/package.json') string in cordis.patch.yml.
• Bundled third-party skills keep their own licenses. See NOTICE.md.
• Design notes: SPEC.md, GLOSSARY.md, docs/adr/.
• Tested with @deepseek-ai/dsh 0.2.1-rc.3 on Node 22.

Uninstall

dsh plugin --profile <profile> remove dsh-nerd-mode

License

MIT. Bundled skills keep their own licenses; see NOTICE.md.