DeepTutor CLI Skill
Teach any AI agent to configure, manage, and operate the DeepTutor learning platform entirely from the command line — chat, knowledge bases, memory, sessions, and more.
The skill is documentation-only SKILL.md with no auto-executing code; destructive commands (kb delete, memory clear, session delete) explicitly require confirmation with an opt-out --force; data flows (settings under data/user/settings, skill installs from ClawHub behind a default security gate with --allow-unverified) are disclosed. Deducted for: skill install can introduce third-party code whose gate semantics need execution to verify, and the blast radius of auto-mounted context-gated tools (exec, etc.) is not explained at the SKILL.md level.
Command surface is complete and internally consistent, cross-corroborated by AGENTS.md, so the happy path is plausible; but this is a static read with no executed reproduction, no committed test evidence covering the skill's key paths, and thin failure-mode/error-feedback description — capped at 10 by anchor.
'When to Use' lists scenarios clearly; capability boundaries (kb/partner/book/memory) are reasonably explicit; --language flag and zh prompt files (per AGENTS.md) support Chinese use. Deducted for: undeclared reachability/failure behavior of external LLM/search services, and limited evidence on trigger precision and environment fit.
Well-structured docs, complete install/dependency notes, Apache-2.0 license, release history (v1.6.6) and named maintainer in CONTRIBUTING. Deducted for: SKILL.md itself carries no version/changelog, no known-limitation or troubleshooting disclosure, and maintenance responsibility for this specific skill file is unclear.
Workflow examples are concrete and self-consistent with the command tables, giving real marginal value as a CLI-teaching skill; but static review cannot verify outputs are directly usable, and marketing release notes are not execution evidence — capped at 7, deducted for lack of direct-usability evidence.
Secondary corroboration exists (CI references, pre-commit config, tests directory) but no committed tests or third-party execution evidence cover this skill's key paths; conclusions rest mainly on file self-consistency — capped at 5 for static review.
- This is a static source review; no commands were executed and all conclusions carry low confidence.
- deeptutor skill install can pull third-party skill packages from the external ClawHub hub; the --allow-unverified gate semantics should be risk-assessed by the user.
- The skill depends on external LLM/embedding/search services; reachability from mainland-China networks is not declared — verify before use.
- Boundaries of context-gated auto-mounted tools (e.g. sandboxed exec) are not explained at the SKILL.md level; review isolation settings when deploying.
- Destructive commands (kb delete / session delete) expose --force paths that skip confirmation and may cause data loss if misused.
What does this skill do, and when should you use it?
This is the root SKILL.md of the HKUDS/DeepTutor repository, a handover document that teaches any tool-using LLM to operate the deeptutor command line. DeepTutor itself is an open-source intelligent learning platform (Apache-2.0, Python 3.11+) led by the HKUDS lab; this skill packages its CLI surface into roughly 150 readable lines. Once read, an agent can run chat and seven capabilities (deep solve, quiz generation, deep research, visualize, math animation, mastery path), manage knowledge bases, IM-connected Partners, skills, memory, sessions and notebooks, and start the API server or full Web app. The project was released in late December 2025 and iterates openly with an active community.
The skill is a reference document: it walks an agent through every deeptutor command and flag — deeptutor init for guided setup, chat/run for seven capabilities (chat, deep_solve, deep_question, deep_research, visualize, math_animator, mastery_path), --tool mounting for tools like brainstorm, web_search, rag and reason, the kb subcommand family for creating and searching RAG knowledge bases, partner start/stop for IM-connected companions, skill search/install from hubs like ClawHub, book/memory/session/notebook/provider/config/serve/start management commands, plus REPL slash commands. --format streams NDJSON events for machine parsing, and --session chains stateful multi-turn runs.
- A terminal-first user who wants to chat, solve problems, and query documents with DeepTutor without opening the browser
- A developer who wants Claude Code or Codex to drive DeepTutor, wrapping deeptutor run as a tool inside a LangChain/AutoGen loop
- A learner who builds a knowledge base from PDF textbook chapters, then asks questions with rag or generates a 5-question thermodynamics quiz
- An operator who needs to launch the FastAPI backend headlessly (deeptutor serve) or the full backend+frontend Web app (deeptutor start) on a server
- A researcher who wants cited deep_research reports grounded in their own paper library
- An educator who creates persistent IM-connected learning companions via the partner commands
What are this skill's strengths and limitations?
- One ~150-line SKILL.md covers the whole CLI surface so any shell-capable agent can learn it in a single read
- run --format streams NDJSON events; headless-safe — ask_user pauses auto-resolve with an empty reply instead of hanging
- CLI and Web app share the same data/user/settings layout, so a CLI-only install can later grow into the full Web app in the same workspace
- Actively maintained (v1.5.2, July 2026), Apache-2.0 licensed, with a docs site (deeptutor.info) including a dedicated agent-handoff page
- The skill's value depends entirely on DeepTutor itself being installed and configured; LLM/embedding/search providers must be set up via init first
- The CLI-only deeptutor-cli package is not on PyPI — source checkout required, and the checkout must be kept around for the editable install
- CLI-only setup defaults embeddings to off; you must opt in during deeptutor init --cli to use kb/RAG features
- The source does not document test-suite coverage or platform-specific stability evidence, so verify on your own stack
How do you install this skill?
Prerequisite: Python 3.11+. Full Web app: pip install deeptutor (needs Node.js 20+). CLI-only must be installed from a source checkout: git clone https://github.com/HKUDS/DeepTutor.git, then python -m pip install -e ./packaging/deeptutor-cli (note: deeptutor-cli is not yet on PyPI). Then run deeptutor init for the guided wizard (add --cli to skip port setup); it writes settings under data/user/settings. The skill itself is the repository's root SKILL.md, which Claude Code, Codex, or OpenCode pick up automatically.
How do you use this skill?
Hand the repo to your agent or use the terminal directly. Typical examples: deeptutor chat --kb textbook --tool rag --tool web_search for the interactive REPL; deeptutor run deep_solve "Solve x^2 = 4" --tool rag --kb textbook for a one-shot turn; deeptutor run deep_question "Linear algebra" --config num_questions=5 for quizzes; deeptutor kb create physics --doc ch1.pdf to build a knowledge base; add --format for line-by-line NDJSON output and reuse --session to chain stateful turns.
How does this skill compare with similar options?
The README explicitly credits agent-native coding CLIs — OpenAI Codex and Anthropic Claude Code — as inspirations for the DeepTutor CLI workflow. Unlike those tools, however, this skill targets the DeepTutor learning platform: it exists so any agent can drive deeptutor as a tool, not to write code.