dsh-learning-mode

English | 中文

An agent preset for DeepSeek Harness (DSH) that teaches while coding, modeled on Claude Code's official Learning output style: explain concretely with usage scenarios, guide your thinking with questions, and leave explicit practice blanks for you to do by hand.

Learning output style (Claude Code): "Collaborative, learn-by-doing mode where Claude will not only share 'Insights' while coding, but also ask you to contribute small, strategic pieces of code yourself."

Three Pillars

| Pillar | Behavior | |---|---| | A · Concrete, scenario-grounded explanations | Flexible, task-tied explanations: everyday analogy (boundaries marked when applicable) + scenario grounding (when/which/why results differ), used as needed; simple concepts get a sentence or two, complex ones get expanded. Depth layered (surface → medium → deep), deepen on demand; no re-teaching within a session | | B · Guided thinking | Question first: at conclusions you can reach yourself, ask one precise predict-then-verify question. When stuck, climb the hint ladder: L1 point at what to look at → L2 point at the principle → L3 reveal with explanation | | C · Practice blanks | Leave small, strategic pieces to you, marked TODO(你) (Claude Code's TODO(human)). Small, strategic, tied to what you're actually doing, self-verifiable; never blank safety-critical, irreversible, or correctness-critical steps |

Interaction protocol: Teaching-first by default; "just do it / no time / asap" switches to direct mode; asks your familiarity level (beginner/intermediate/advanced) once at the start; ≥2 failed attempts downgrades to a guided reveal. Output language follows your input: Chinese in, Chinese out; English in, English out.

Examples

Teaching DeepSeek Harness's principles — the verbatim output of a real learning-mode session — a full teaching turn showing how concrete explanations, guided thinking, a context-tied TODO(你), and the one-time opening calibration work together. This file is human documentation and is never loaded by any skill: the learning-mode skill teaches only principles and forms (placeholder templates); concrete examples are invented at runtime from your current task, so fixed examples cannot degrade generalization. If you want to keep a permanent example, put it here — not in learning-mode/skills/.

Install

Requires DSH 0.1.0-rc.x (a deployment with profiles/ under ~/.dsh).

# Option 1: clone and copy
git clone https://github.com/CHplus0/dsh-learning-mode.git
cp -r dsh-learning-mode/learning-mode ~/.dsh/.agent-presets/

# Option 2: run the installer
bash dsh-learning-mode/install.sh

# Option 3: install the npm bundle (auto-installs the preset)
dsh plugin --profile web add dsh-learning-mode

Then open the DSH web UI, start a new session and pick 学习模式 (Learning Mode) — no restart needed.

Customization

• Tone & identity: edit learning-mode/agent.cordis.yml → persona.text.
• Style details & phrasing templates: edit learning-mode/skills/learning-mode/SKILL.md.
• Rename: edit only name in learning-mode/preset.yml (the directory name is the preset id, must match [a-z0-9][a-z0-9-]*; renaming requires renaming the directory too).

How it works

• agent.cordis.yml is a full copy of the standard preset with two changes: the persona is replaced with the teaching identity (the three pillars, always in the system prompt), and skill-filesystem gains customSkillDirs pointing at this preset's bundled skills/ directory (the full guide loads on demand, not in the standing prompt). The toolset is identical to the standard coding agent (Shell, files, search, Skills, planning, goals, subagents, workflows).

License

MIT © 2026 CHplus0. The preset composition is adapted from the standard agent preset of deepseek-ai/deepseek-harness (MIT © 2026 DeepSeek); see LICENSE.