DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

summersummer-liu /

summersummer-liu/dsh-socratic-mode

Verified

Socratic learning mode for DeepSeek Harness — a tutoring agent preset that gives one piece of real knowledge then asks one invitational question per turn. 先给再问,不讲课不审问。

★ 0 Stars0 Forks0 IssuesN/A Community rating0 Confirmed installs
View on GitHubProject homepage
READMESource: main@562fb22f

dsh-socratic-mode

A Socratic learning mode for DeepSeek Harness — it gives you one small piece of real knowledge, then asks one invitational question. No lectures, no interrogation.

License: MIT

中文说明 · MIT License


Why this exists

Most “AI tutor” prompts have the same failure mode: they either lecture at you, or they quiz you forever — question after question with nothing given back. Both feel like pressure, and after ten minutes you stop wanting to chat.

This mode flips the order: give first, then ask (先给后问). Every turn hands you one concrete piece of knowledge — a key distinction, a counter-intuitive fact, a number, a one-line takeaway, or a common misconception — and then invites you to think one step further. Questions are invitations, not interrogations. You can always say “just tell me”, “just talk”, or “I'm tired”, and the mode adapts.

What you get

  • A new agent preset — 苏格拉底学习模式 (Socratic Learning Mode) — on the New Session picker.
  • A knowledge block every turn, so each reply is worth reading even when you don't answer.
  • Exactly one question per turn, always optional.
  • Pressure controls: a question-free turn every 4–6 turns, hidden mastery checks instead of “exams”, and a graceful exit when you are tired.
  • Warm register: specific appreciation (quoting your own words), safety when you're stuck, invitations (“what do you think?”) instead of commands (“please answer”).
  • A study note in your workspace at the end: knowledge map, key distinctions, your own best sentences, misconceptions, and three self-test questions.

Requirements

  • DeepSeek Harness desktop / web, dsh >= 0.2.1-rc.3
  • Node.js >= 20 (only for the repo's check script and development)

Install

  1. Clone the repository somewhere local:

    git clone https://github.com/summersummer-liu/dsh-socratic-mode.git
    
  2. Ask your DSH agent to install the bundle — say:

    用 plugin_manager 安装 dsh-socratic-mode 这个目录的 bundle。

    The agent calls plugin_manager with action: install_bundle and target set to the absolute path of the cloned directory. The bundle resolves the shipped Harness packages from your installation; it declares no dependencies of its own.

  3. (Optional) install the companion skill, which is useful even outside this preset:

    mkdir -p ~/.dsh/skills
    cp -r skills/socratic-tutor ~/.dsh/skills/
    
  4. Open Settings → Agent Presets. The mode appears under Custom and should show a healthy card (title, description, icon). Click the card to make it the default for new sessions, or just pick it per session.

Use

Create a new session, pick 苏格拉底学习模式, and say what you want to learn:

Goal Say
learn a concept from scratch 「教我通货膨胀,我基础很差」
study a document 「用苏格拉底法帮我读这份讲义」 (drop the file into the chat)
verify what you know 「考我文艺复兴,我看过两遍书了」
untangle a confusion 「我总觉得货币乘数讲不通,你来问问我」

During the session:

You want Say
the answer now 「直接告诉我」
a hint instead 「卡住了 / 给点提示」
a lecture with no questions 「讲讲吧」
to stop 「累了 / 先别考我」
to wrap up 「总结一下」 → writes a study note

What's inside

Part File Purpose
Preset declaration cordis.patch.yml Inserts preset-socratic into the agent-preset roster
Teaching method (always on) cordis.patch.yml → persona.prefix 20 hard rules: give-then-ask, tone, pressure control, hidden mastery checks
Teaching method (on demand) skills/socratic-tutor/SKILL.md Knowledge-block library, six-question ladder, four hint levels, tone manual, note template, anti-patterns
Display resources locale/en.json, locale/zh.json, icon.svg Bundle card title, description and icon

The preset mounts 13 child rows: persona, time-context, tool-fs, tool-fs-search, tool-bash/tool-pwsh, tool-web, skill-filesystem, tool-skill, tool-ask-user, tool-todo, present, and a compaction group. It deliberately omits plan mode, goals, schedules, background jobs, subagents, workflows and the plugin manager: a learning session should not carry system-modification tools.

Design

Two layers keep the constraints strong without spending context on them every turn:

  1. Persona (~1.4 KB, visible every turn) — the rules that must beat the model's default habits.
  2. Skill (~6 KB, loaded once) — the full playbook: knowledge-block templates, question ladder, hint ladder, tone manual, study-note structure.

The core loop:

接住他一句 (optional)  +  一块真知识 (required)  +  一个邀请式问题 (exactly one)
  • Give: one of five knowledge-block types — key distinction, counter-intuitive fact, concrete example/number, one-line takeaway, common misconception.
  • Ask: invitation only (“你会怎么想?” “想不想试试更刁的?”), never commands (“请回答”).
  • Never: two questions in one turn, two give-less turns in a row, judging “right/wrong”, broadcasting progress, or pushing questions when the learner is tired.

Mastery is checked silently — explain it, give your own example, transfer to a new case — and never announced as a test.

Customize

Want Edit
teacher persona and rules cordis.patch.yml → persona.prefix, then reinstall the bundle
question ladder, hint ladder, tone manual, note template skills/socratic-tutor/SKILL.md (re-read on every load, no reinstall needed)
tool set cordis.patch.yml → plugins
roster position config.order (built-ins use 1–4)

Validate before reinstalling:

npm install
npm run check

Reinstalling an installed bundle needs two steps — plugin_manager → remove_bundle, then install_bundle (installing over an existing bundle reports already-installed).

Troubleshooting

The preset card shows 加载失败 (failed to load) and new sessions refuse to start. Child rows are validated when the preset mounts; one invalid child fails the whole preset. The toast or card diagnostic names the row. Run npm run check. A known trap: @deepseek-ai/dsh-tool-todo requires allowParallelInProgress with no default — omitting it fails the preset. The check script asserts it.

The mode doesn't appear in the picker. Refresh the page. Confirm the bundle is listed in the profile, and that the preset row is active (plugin_manager → list_plugins → preset-socratic).

The tutor slides back into quizzing. Reload the skill and re-read rule 1 and rule 15 in persona.prefix: give-then-ask, and “直接讲吧” means deliver a complete explanation, not another fill-in-the-blank.

Uninstall

  1. plugin_manager → action: remove_bundle → target: dsh-socratic-mode
  2. Remove the skill: rm -r ~/.dsh/skills/socratic-tutor
  3. Delete the cloned directory

License

MIT — see LICENSE.

—/ 5

No ratings yet

Verified DSH bundle

Commit 562fb22fa41b

Community comments

No comments yet. Be the first to write one.

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout