dsh-project-based-learning

项目制教学教练：学员动手，教练跑闭环
面向 DeepSeek Harness（DSH）的教学技能与组合包 —— 教学法与学科内容分离，领域导航可整体替换

简体中文 · English

This project constrains "AI tutoring" into an executable teaching protocol: teaching is organized around the work the learner actually wants to finish, explicit instruction is a formal stage, and beginners first get an understanding sufficient to act, then practice in a real project as soon as possible.

• Teach first, then practice: for unfamiliar knowledge, explain the minimal necessary mechanism first instead of asking the learner to guess; theory covers only what the current step needs, then returns immediately to one real operation.
• Explain in place: new concepts, new built-in functions, and non-obvious syntax are explained where they first appear; do not translate an entire file line by line, and do not unfold a whole subject at once for the sake of "completeness".
• Non-graded evidence: only five descriptive concept states record the difference between "explained" and "already demonstrated through practice"; no numeric ability scores and no fixed-dimension grades are used.
• Subject decoupling: the teaching protocol lives in SKILL.md and references/*.md, and subject knowledge lives in domain navigations that can be replaced wholesale; changing subject does not require changing the teaching protocol.

Naming: the repository name, the npm package name, and the registered skill name are identical, all being dsh-project-based-learning (the model invokes it through skill("dsh-project-based-learning")).

📐 Design stance

🚫 What it is not

• Not a question bank or a drill tool: a domain navigation gives only rough dependencies and threshold concepts; it contains no diagnostic question bank, no fixed lesson hours, and no standard answers.
• Not a ghost-writing tool: a review request does not automatically authorize modification; a teaching request does not automatically widen project write permission.
• Not an official DeepSeek plugin: this project is a third-party implementation.
• Not limited to any subject: two navigations ship with the package, Unity Shader and Unity C#; both are optional add-ons you can delete or swap without affecting the teaching protocol.

🚀 Quick start

Option 1: let DSH install it itself (recommended)

Copy the whole block below and paste it into any DSH session you are using — it installs the plugin itself and checks each item:

Please install the DSH plugin dsh-project-based-learning (project-based teaching coach). Steps:

1. Run: dsh plugin --profile web add dsh-project-based-learning
   Change the profile name to the one you actually use; the desktop default is web.
   If you get "dsh is not recognized", dsh is not on PATH. On the desktop build the entry point is:
   %LOCALAPPDATA%\Programs\DSH Desktop\resources\app.asar.unpacked\node_modules\@deepseek-ai\dsh\lib\bin.js
   Call it with node and the full path (note app.asar.unpacked; the older resources\app path no longer exists).
   If the installed version is below 3.0.0, use instead:
   dsh plugin --profile web add github:Kirisame1969/dsh-project-based-learning
2. Run: dsh --profile web --dump-config and confirm that a dsh-project-based-learning layer appears in the output
3. Restart the Harness, then call skill("dsh-project-based-learning") and confirm the first line of the body is "# 项目制教学教练 2.0"
   Two common misreads — do not treat either as a failure:
   - this plugin copies no files into any skills directory; the skill is registered at runtime, so "it is not in my skills directory" is expected;
   - the skill body is read once when the plugin loads, so the version will not change until the Harness restarts.
4. Report back to me: the installed version, whether that layer exists, and what the first line of the skill body is

If you hit an error, first read the "Installation details" section of the README at https://github.com/Kirisame1969/dsh-project-based-learning and docs/releasing.zh.md.

Option 2: type one command yourself

dsh plugin --profile web add dsh-project-based-learning

Once it is installed, just say in the session what you want to learn; there are no command words to memorize:

Teaching mode: I have never studied at all, please assess my level

The coach first decides which teaching lane to take, then teaches the minimal mechanism the current step needs, and then takes you to an observable result as soon as possible — instead of opening with a set of diagnostic questions.

The domain navigations shipped with the package are demonstrated in skills/dsh-project-based-learning/references/domains/unity-shader.md and unity-csharp.md.

📦 Installation details

Bundle: install from npm or from GitHub

dsh plugin --profile web add dsh-project-based-learning     # or --profile headless, or your own profile
dsh --profile web --dump-config                             # the dsh-project-based-learning layer should appear

Without npm, install straight from this repository:

dsh plugin --profile web add github:Kirisame1969/dsh-project-based-learning

The bundle layer (cordis.patch.yml) registers the skill distributed with the package through ctx.skills.register(). The plugin only consumes the skills service — it imports nothing from the harness and brings no second copy of Cordis.

Uninstall:

dsh plugin --profile web remove dsh-project-based-learning

Learning data lives in .learning/ inside your workspace; uninstalling the plugin does not delete it.

No install at all

Point your agent at skills/dsh-project-based-learning/SKILL.md and ask it to follow that file. The teaching protocol, the domain navigations and the scripts are plain files.

🛣️ Teaching lanes

The skill automatically picks the lightest lane for the request; you do not need to memorize the categories:

When the intent is clear, start directly, without an entrance knowledge test. Ask only when the missing information would clearly change the technical approach, the project boundary, or the teaching method, and ask at most three questions per round.

🧩 How it works

flowchart LR
    A["Locate this round's result"] --> B["Teach the necessary mechanism"] --> C["Demonstrate (as needed)"]
    C --> D["Learner practices"] --> E["Observe the phenomenon"] --> F["Causal feedback"]
    F --> G["Nearby variation"] --> H{"Adjust the scaffolding"}
    H -->|success| I["Reduce scaffolding / next result"]
    H -->|blocked| D
    I --> J["Record meaningful changes"]

skills/dsh-project-based-learning/
├── SKILL.md                         # Entry: lanes, inviolable principles, default loop, output format, routing
├── references/
│   ├── teaching-loop.md             # Attention focus, teaching granularity, demonstration strength, practice design, questions, error correction, preventing theory drift
│   ├── planning-and-state.md        # Route granularity (now/next/later) and concept-state semantics
│   ├── review-and-adapt.md          # Artifact review, debugging as teaching, scaffolding adjustment, milestone review
│   ├── permissions.md               # Distinguishing request types, project write boundaries, evidence and privacy
│   ├── domain-guidance.md           # The contract for what a domain navigation "should contain / must not contain"
│   └── domains/
│       ├── index.md                 # Choose at most one relevant navigation
│       ├── unity-shader.md          # Unity Shader dependency graph, threshold concepts, risks, version check
│       └── unity-csharp.md          # Unity C# dependency graph, threshold concepts, risks, version check
├── assets/
│   ├── state.template.json          # 2.0 state template
│   └── lesson-note.md               # Optional single-lesson note template
└── scripts/
    ├── validate-learning-state.mjs  # Zero-dependency state validator
    └── migrate-v1-state.mjs         # One-time migration from the old 2.x state

• Protocol loaded on demand: SKILL.md references the detail files listed above only when needed, so the methodology of a whole course is not packed into every round's context.
• The subject affects only the route: a domain navigation provides rough dependencies and threshold concepts; the concrete explanations, examples, exercises and questions are generated on the spot around the current work.

🎛️ Domain navigations

The pedagogy is bound to no subject. A domain navigation is a single-file, coarse-grained dependency graph used to keep the route from losing order; it stores no course content — even with no navigation at all, the general teaching loop remains fully usable.

The two navigations shipped with the package (click to expand)

File	Coverage
references/domains/unity-shader.md	Shaders, materials, Sprite visual treatment, screen filters. Threshold concepts such as "a fragment / a screen pixel / a texel are not the same concept" and "UV is a coordinate, not a loop variable"; includes risks such as transparency render state, coordinate ratios and overdraw
references/domains/unity-csharp.md	C# game logic, the component model, scene organization, runtime debugging, project structure. Threshold concepts such as "a C# object / a component / a scene instance are not the same layer" and "the lifecycle is driven by Unity calls"

Both write only "coverage, rough route, key threshold concepts, common misconceptions and risks, teaching stance, version check", and both require explicitly: do not force the learner through every prerequisite lesson just because the navigation lists a stage.

Adding a domain navigation: put a single file at references/domains/<id>.md, then add one line in references/domains/index.md pointing to it. The contract is in references/domain-guidance.md and CONTRIBUTING.md.

Three boundaries (stated plainly):

• A domain navigation must not contain a diagnostic question bank, fixed lesson hours, a prefabricated exercise list, an acceptance rubric or an exhaustive glossary — that would turn it into a course package.
• A domain navigation provides direction and dependencies only; skipping, merging, rolling back or reordering is decided by the current work and the existing evidence.
• A domain navigation must be located under references/domains/ in the skill directory.

💾 State and migration

Only an ongoing course creates a state file. A quick question, a one-off term explanation and a single small exercise create no state.

The state path is .learning/state.json, and concepts use only five descriptive states (increasing in strength):

尚未接触 (not yet encountered) → 已讲授待实践 (taught, awaiting practice) → 带练中 (guided practice) → 可在熟悉任务中独立使用 (usable independently in familiar tasks) → 已迁移到新任务 (transferred to a new task)

The last three states must have practice evidence; instruction or a demonstration alone does not count. The validator enforces this mechanically:

node skills/dsh-project-based-learning/scripts/validate-learning-state.mjs --state .learning/state.json

(The bundled CLI also works: coach-validate --state .learning/state.json.)

Migrating from 2.x

3.0 is a breaking change: the state path, the state schema and the domain package structure are all incompatible with 2.x. The old .coach/state.json is handled by a one-time migration script:

node skills/dsh-project-based-learning/scripts/migrate-v1-state.mjs \
  --input .coach/state.json --output .learning/state.json

(Or coach-migrate --input .coach/state.json --output .learning/state.json.)

The migration trade-offs are deliberate:

• Keeps the goal, completion criteria, environment, non-goals, current task and route text, compressed into now / next / later.
• Does not migrate numeric ability levels, and does not automatically map the old "verified" status to any concept state; concept states must be re-established from actual course evidence.
• Records the migration source in decisions as a reminder to verify it manually in the next round; the migration result must be reviewed by a human.
• Migration is a one-time operation: a new course maintains only .learning/state.json and does not keep two sets of state in sync.

⚠️ The migration script is lossy. The old strategy.assumptions (which may contain format conventions issued by the learner), open (unresolved items), authorizations (authorized scope) and routeChanges are not carried into the 2.0 structure. If those are still valid, add them by hand to preferences, decisions and context.exclusions after migrating.

📋 Requirements

• Node.js ^22.19.0 || >=24.0.0 (the validation and migration scripts need a Node with ESM support; teaching itself does not).
• No npm dependencies, no build step: lib/index.js is hand-written source, not a build artifact.

📚 Documentation

• CHANGELOG.md — version changes
• CONTRIBUTING.md — how to add a domain navigation
• docs/releasing.zh.md — release checklist and locally verified channel findings (Simplified Chinese)
• references/domain-guidance.md — the domain navigation contract

🤝 Contributing

The most valuable contribution is adding a subject domain navigation; start with CONTRIBUTING.md.

📄 License

MIT.