SWE Mentor Kit — v0.4.0
Learn to reason, build, debug, test, review, and ship software with AI — without outsourcing the skill you want to develop.
An installable, portable Socratic learning plugin for Claude Code and Codex, with a separate prompt-based guide for ChatGPT, inspired by Manware's AI Learning Toolkit, adapted for the Software Engineering & Computer Science learning path.
Pedagogical protocol
MAP → DISCOVER → UNDERSTAND → DO → TEST → LOCK
Three modes:
- TUTOR: new concepts and focused labs; the learner writes the target code, AI offers graduated hints.
- ASSISTANT: debugging, code reading, design discussions and reviews; inspect hypotheses and evidence.
- AGENTIC / SHIP: explicit delegation for production, with diff, tests, and risk review. This does not establish that the learner mastered AI-generated code.
The goal is Software Engineer competency, not a particular framework, badge, or technology stack. The kit does not replace any existing roadmap.
Install as a plugin (recommended)
SWE Mentor is now a distributable plugin for Claude Code and Codex. You don't need to clone the repository or execute an installer for normal use.
Claude Code
Run these inside a Claude Code session, one at a time:
/plugin marketplace add libsss01/swe-mentor-kit
/plugin install swe-mentor@swe-mentor
Restart Claude Code and try /swe-mentor:learn, /swe-mentor:hint 1, /swe-mentor:debug, or /swe-mentor:retrieve. The plugin ships 15 namespaced commands.
Optional local validation before installation:
claude plugin validate .
Codex
Register the GitHub marketplace, then install the plugin:
codex plugin marketplace add libsss01/swe-mentor-kit
codex plugin add swe-mentor@swe-mentor
In Codex, select or invoke the bundled skill with $swe-mentor and describe your workflow, for example: $swe-mentor /debug My deletion test fails after a hash collision.
Availability of marketplace installation depends on your local Codex version. If
codex plugin addis unsupported, consult the Codex plugin support for your client version. Do not assume marketplace installation is supported. This repository is not yet listed in any universal public plugin directory.
What gets installed?
plugin.json: portable Agent Plugins manifest, including OpenAI-specific interface metadata inextensions.com.openai..claude-plugin/plugin.jsonand.claude-plugin/marketplace.json: Claude Code package identity and GitHub marketplace..agents/plugins/marketplace.json: Codex local marketplace. A redundant.codex-plugin/plugin.jsonis no longer needed.skills/swe-mentor/SKILL.md: self-contained Socratic tutoring skill with references.commands/*.md: 15 Claude Code plugin commands.
This is a skills-only plugin. It has no MCP server, credential requirements, telemetry, persistent cloud service, or automatic file mutation. Learning logs remain optional.
Learn by building a plugin
A plugin is an installable bundle, a skill defines reusable instructions, and a marketplace indexes bundles. Start by reading skills/swe-mentor/SKILL.md, then .claude-plugin/plugin.json, plugin.json, and the two marketplace catalogs. To contribute: fork, edit a workflow, run python3 verify_plugin.py, validate the Claude plugin locally, then submit a pull request.
Usage
Claude Code: after installing the plugin, try /swe-mentor:learn, /swe-mentor:hint 2, /swe-mentor:debug, /swe-mentor:code-review, /swe-mentor:retrieve, or /swe-mentor:ship. The 15 commands live in commands/ and are namespaced by the plugin.
Codex: invoke the included skill using $swe-mentor /debug My Java deletion test fails after a collision, or select it in a supported skill picker.
ChatGPT: this repository does not install a native ChatGPT plugin. Follow the prompt-based instructions below.
Local validation
python3 verify_plugin.py
claude plugin validate .
The Python script checks the repository structure and bundled learning-log layout; the second command needs Claude Code. Actual installation and agent behavior still require local tests.
Complete command set
| Command | Purpose |
|---|---|
/learn | Select target competency and first exercise |
/hint | Graduated help 0–6 |
/debug | Reproduce, hypothesize, inspect, fix and test |
/autopsy | Postmortem a bug and learn its transferable lesson |
/read | Understand existing code and predict behavior |
/code-review | Severity-ranked review without immediate rewrite |
/test | Derive meaningful tests from contracts |
/explore | Compare tradeoffs and alternatives |
/arch | Challenge minimal architecture and failure modes |
/explain | Explain-back and check understanding |
/retrieve | Active recall from inspectable past work |
/api | Understand an API's contract, errors and constraints |
/map | Locate a prerequisite or next micro-lab |
/lock | Consolidate a lesson and link actual evidence |
/ship | Delegate an explicit, bounded production task |
Manware-inspired commands are adapted, not reproduced verbatim. /map, /lock, /ship extend the workflow.
Repository structure
.claude-plugin/
plugin.json # Claude Code plugin metadata
marketplace.json # Claude Code marketplace catalog
.codex-plugin/
plugin.json # Codex compatibility metadata
.agents/plugins/
marketplace.json # Codex marketplace catalog
plugin.json # Portable manifest + OpenAI interface metadata
skills/swe-mentor/
SKILL.md # Main pedagogical behavior
references/
commands.md # Workflows for all 15 commands
assessment.md # Evidence standards
learning-logs.md # Persistent learning log protocol
commands/ # 15 Claude Code plugin commands
templates/learning/ # Four log templates, not student data
scripts/learning_logs.py # Optional local init/append tool
tests/ # Local filesystem + multi-session persistence tests
.github/workflows/test.yml # CI checks on push and pull request
verify_plugin.py # Plugin structure validation
README.md
LICENSE
Using the method in ChatGPT (prompt-based, not a native plugin)
A locally installed Claude/Codex skill is not automatically loaded into ChatGPT. Copy this brief guidance into a ChatGPT conversation and attach the skill files if the conversation does not already contain the instructions:
Apply SWE Mentor: MAP → DISCOVER → UNDERSTAND → DO → TEST → LOCK. Tutor new concepts and labs; assist reviews and debugging; use agentic implementation only on explicit delegation. During /learn, dynamically use hints 0–6, debugging, testing, code review and retrieval without requiring slash commands. Treat '/hint', '/debug', '/retrieve', and '/ship' as textual intents, not built-in ChatGPT commands. Protect the learner's independent practice. Do not claim access to unavailable logs or tests. Store learning notes only in authorized project-local files and with consent.
Example: /learn I need to save and load my Java book catalog, and want to understand serialization while implementing it.
Serialization learning scenario: three-session acceptance test
- Session 1: A learner building a Java library needs books to survive application restarts.
/learnconnects persistence requirements to serialization, lets the learner implement a small save/load round-trip, and adapts hints whenNotSerializableExceptionappears. It should not require separate/debugor/hintcalls. - After practice: The learner approves meaningful notes in four project-local logs. An actual demonstrated save/load test is recorded separately from an explanation merely given by AI.
- Session 2: A fresh session reads
context, recalls the student's incorrect assumption about nested fields, and asks about class version compatibility before giving the saved answer. - Session 3: A new settings-persistence problem tests transfer. Only inspected independent work supports a scoped mastery claim.
Automated tests simulate project-local note creation, retrieval across separate sessions, record preservation and missing-history behavior. They cannot prove the language model will adapt its teaching; that requires a real-client multi-session evaluation.
Manual behavior checks (also supported by automated structural and filesystem tests)
| Scenario | Expected |
|---|---|
| /learn: Java project needs serialization | A project-specific path from concept to micro-lab, no demand to call /hint |
| Student hits NotSerializableException | Hypothesis, discriminating observation, no unrequested rewrite |
| /hint 0 | Question to elicit a prediction, not complete code |
| Explicit /hint 6 | Explained solution and transfer challenge |
| Session starts with approved existing logs | Retrieval from one real prior difficulty, not invented memory |
| No project logs / declined logging | Session still works; no files created |
| /lock after meaningful learning | Suggest concise notes; write only with consent |
| /ship explicitly delegated production task | Authorized implementation with tests/diff/risks; no mastery claim |
Automated tests: run python3 verify_plugin.py and python3 -m unittest discover -s tests -v. GitHub Actions runs both on pushes and pull requests. These tests do not test model responses or prove Claude/Codex compatibility; those require real-client evaluation.
Safety, evaluation and limits
- No implicit repository edits in TUTOR mode, and no automatic escalation to SHIP after a blockage.
- Verification = observed tests and inspectable artifacts; AI-generated changes alone do not demonstrate mastery.
- The assistant should first invite a learner's own verdict and evidence, without blocking a requested review.
- No mandatory 30-minute timer, no compulsory journal after every message, no needless quizzes.
- Keep capability progress separate from what was discussed, attempted, or planned.
- Do not update learning governance documents automatically.
- Syntax and static checks are provided; behavior must be tested locally with your Claude Code and Codex versions.
Original inspiration: Manware. Platform documentation: Claude Code skills · Codex skills. This is an independent pedagogical adaptation, not an official plugin.