How do I get started?

Create a new assignment from the dashboard by clicking 'New Assignment'. Describe the assignment topic, select a language and difficulty, then let the AI generate everything for you.

What AI providers are supported?

CodeTeach supports OpenAI (GPT-4o), Anthropic (Claude), Google Gemini, Mistral, Groq, and DeepSeek. You bring your own API key -- add it in Settings before generating assignments.

BYOK assignment generation costs 1 credit; Platform AI costs 2 credits per generation (we cover the AI inference cost). New accounts get 10 free credits. Buy more from the Credits page or see the full pricing breakdown at /pricing. Credits never expire.

What does the generated assignment include?

Phase 1 produces starter code, a complete solution, automated tests, and student-facing instructions. After you review and validate that package, Deploy builds and verifies the GitHub Actions autograding workflow.

How do I deploy to GitHub?

First install the CodeTeach GitHub App on your account or organization in Settings. Then after reviewing the generated assignment, click Deploy to create a private template repository. Keep that template instructor-only and create separate private repos for students.

What languages are supported?

Currently supported: Python, JavaScript, TypeScript, Java, C, C++, C#, Go, Rust, Ruby, SQL, and Jupyter Notebooks. More languages can be added by an admin.

Why did my generation fail?

Common causes: expired or rate-limited API key, insufficient API plan quota (free tiers often hit rate limits), or a temporary provider outage. Check that your API key is valid and you have a paid API plan for best reliability.

Can I edit the generated code?

Yes. After generation, the Review step lets you edit starter code, solution code, tests, and instructions before deploying. Editing a graded artifact invalidates the previous QA verdict so the current complete package can be checked again.

Help

Quick answers below, or browse the full Help Center for in-depth articles.

Browse the full Help Center

Step-by-step guides for every part of the workflow — picking a generation mode, choosing a difficulty level, the autograder check, deploying to GitHub, and troubleshooting.

Search and browse 20+ articles organized by collection.

How CodeTeach builds and verifies an assignment

AI creates candidate artifacts. Trust comes from the checks that follow: observed program behavior, isolated test runs, complete-package review, and an autograder proven in GitHub.

1
Build the assignment package
CodeTeach uses the generation mode to decide what the instructor owns and what AI may create.
- Creates only the missing starter, solution, tests, and instructions; uploaded or instructor-approved artifacts are not silently rewritten.
- Checks that generated starter code leaves meaningful student work unfinished and is not simply a copy of the solution.
- Preserves uploaded support files exactly. Small text fixtures may be constructed only when the source document supplies enough detail.
2
Derive tests from observed behavior
The app does not trust an AI model to guess the exact output a program should produce.
- Classifies the assignment by behavior, such as standard input, interactive loop, function or class, files, database, notebook, or nondeterministic output.
- Uses AI to propose scenarios and coverage goals, then runs the solution in an isolated sandbox to capture real output, errors, exit codes, and file effects.
- Builds behavioral expectations from successful observations and re-runs the resulting tests in isolation. Broken candidates cannot become a false green by being silently discarded.
3
Check package readiness
Before Review, deterministic checks confirm that the artifact set is complete enough to grade.
- Requires the correct starter and solution entry files, nonempty instructions, and at least two executable tests.
- Requires valid scoring weights that total 100 and meaningful expected results for behavioral tests.
- Rejects weak patterns such as using only an exit code to grade visible behavior or supplying only part of an interactive conversation.
4
Run complete-package quality validation
After the instructor reviews and approves Instructions.md, CodeTeach validates the current package as one unit.
- Scans the solution for known grading hazards, then compiles or runs it with the approved dependencies and executes every test in an isolated sandbox.
- Shows each test's pass or fail result, actual output, expected output, errors, and exit status instead of presenting an unexplained score.
- Runs a read-only AI quality review across the instructions, starter, solution, tests, and source document to flag semantic drift such as different input order, required APIs, or output format.
- If an AI Wizard fix changes one artifact, the complete sandbox and semantic QA process runs again. Earlier findings remain until a fresh verdict replaces them.
5
Prove the GitHub autograder
Deployment has a separate technical gate because a sound assignment can still be packaged into a broken workflow.
- Builds the GitHub Actions workflow and scoring files deterministically from the approved tests.
- Runs an internal replica of the exact autograder bundle and requires a 100% answer-key score before publication.
- After the push, watches the real GitHub Actions run on the answer-key branch and reports validation as pending, passed, or failed.

Why this is stronger than a one-pass manual build

CodeTeach does not replace instructor judgment. It automates the repetitive, cross-file, and environment-specific verification that is difficult to perform consistently by hand.

Execution evidence, not model confidence

A fluent AI response is not treated as proof. Code, tests, and the autograder must run.

Repeatable checks

The same isolated validation can be rerun after edits, avoiding the easy-to-miss differences of an informal manual check.

Whole-package consistency

CodeTeach compares artifacts that are often reviewed separately: what students are told, what the solution does, what tests enforce, and what the rubric scores.

Protected instructor intent

Instructor-provided work is not silently changed to make a check pass. When intent is unclear, the issue returns to Review.

Two deployment environments

The autograder is checked internally before publication and again in the real GitHub Actions environment afterward.

Visible limits and findings

Skipped AI reviews, warnings, failed tests, and post-push failures remain visible. CodeTeach does not turn missing evidence into a clean result.

What still requires the instructor

You remain the authority on learning objectives, difficulty, accessibility, institutional policy, and whether the assignment is appropriate for your students. Read the instructions and inspect the artifacts before approval; CodeTeach supplies evidence and catches inconsistencies, but it should not make the pedagogical decision for you.

Quick FAQ

Resources

GitHub Template Repos

Learn how to create repositories from templates.

Autograding Action

How GitHub Actions workflows run on repository pushes.

Feedback

Have feedback or found a bug? We would love to hear from you.

Or email admin@codeteach.ai.

Loading…

Help

Browse the full Help Center

How CodeTeach builds and verifies an assignment

Build the assignment package

Derive tests from observed behavior

Check package readiness

Run complete-package quality validation

Prove the GitHub autograder

Why this is stronger than a one-pass manual build

Execution evidence, not model confidence

Repeatable checks

Whole-package consistency

Protected instructor intent

Two deployment environments

Visible limits and findings

What still requires the instructor

Quick FAQ

Resources

GitHub Template Repos

Autograding Action

Feedback

Help

Browse the full Help Center

How CodeTeach builds and verifies an assignment

Build the assignment package

Derive tests from observed behavior

Check package readiness

Run complete-package quality validation

Prove the GitHub autograder

Why this is stronger than a one-pass manual build

Execution evidence, not model confidence

Repeatable checks

Whole-package consistency

Protected instructor intent

Two deployment environments

Visible limits and findings

What still requires the instructor

Quick FAQ

Resources

GitHub Template Repos

Autograding Action

Feedback