Most engineering teams tried AI coding tools early, and many came away with a firm opinion: unreliable in a real codebase, too much effort to review. Since then, Claude Code, OpenAI Codex, Cursor and GitHub Copilot have become agents that read code, edit files, run tests and prepare pull requests. The old opinion is often out of date.
The new problem is different. Every engineer tries a different tool, there are no shared rules, review load grows, and nobody is quite sure what the agent is allowed to do. Training is worth it when it sorts out exactly that.
Claude Code training for engineering teams
Claude Code training from Specialty Tokens is a hands-on workshop on your own codebase, not a demo project. Beforehand we agree which tools are approved, which repositories may be used and which systems the agent should reach. In the session, the team works on real tickets, with its own tests and its own build. We are not tied to any vendor and train in Claude Code, Codex, Cursor or GitHub Copilot, whichever your company has approved.
Claude Code vs Codex vs Cursor vs GitHub Copilot
All four tools are now agents that change files and run commands. They differ mainly in where they run and how they fit your workflow. Details as stated in each vendor's documentation (as of October 2026).
| Where it runs | How it works | Strengths | Fits when | |
|---|---|---|---|---|
| Claude Code (Anthropic) | Terminal, VS Code, JetBrains, desktop app, browser, Slack, GitHub Actions and GitLab CI/CD | Agent reads the codebase, edits files, runs commands, creates commits and pull requests; instructions in CLAUDE.md, also reads AGENTS.md | Hooks, skills, subagents and MCP; scriptable in the terminal and in CI | the team lives in the terminal and wants the agent built into its own workflows and CI |
| Codex (OpenAI) | CLI, extension for VS Code, Cursor and Windsurf, Xcode, JetBrains, desktop app, Codex Cloud | Works locally in the repository or as a cloud task; instructions in AGENTS.md | Included in ChatGPT plans from Free to Enterprise | the company already uses ChatGPT Business or Enterprise |
| Cursor | Its own AI editor, CLI, cloud agents | Agent in the editor with a plan mode; rules in .cursor/rules or AGENTS.md | Choice of models from OpenAI, Anthropic, Google and Cursor | engineers work in the editor rather than the terminal and want to switch models |
| GitHub Copilot | VS Code, Visual Studio, JetBrains, Eclipse, Xcode, Copilot CLI, GitHub.com | Completions, chat and agent mode in the IDE; the cloud agent works issues in a GitHub Actions environment and opens a pull request; instructions in copilot-instructions.md, AGENTS.md or CLAUDE.md | Built into issues, pull requests and code review on GitHub; broad IDE coverage | code and work already live on GitHub and Copilot Business or Enterprise is licensed |
Honestly, the existing license often decides. If you run ChatGPT Business, you already have Codex. If you have Copilot Enterprise and manage everything on GitHub, you rarely need a second tool. Claude Code shows its strengths when a team wants to build the agent deep into its own workflows, scripts and CI. Cursor is the better choice for engineers who want to stay in the editor all day.
The tools are not mutually exclusive. Codex and Claude Code also run as extensions inside Cursor, and all four read an AGENTS.md in the repository. Keep your project instructions clean and you can switch tools later without starting over.
How to roll out agentic coding without hurting code quality
This is the sequence we practice in the training, on your repository:
- Set the tool and the rules: which repositories the agent may see, which commands it may run without asking, which MCP servers are approved, and how credentials and customer data stay protected. Written into settings, not just discussed.
- Write project instructions: CLAUDE.md or AGENTS.md with architecture, conventions, build and test commands, and clear don'ts.
- Plan first, then build: the agent proposes a plan, the engineer corrects it before any code is written.
- Small steps with tests: the agent writes and runs tests; one reviewable diff per step.
- Review stays with humans: every diff is read before it is merged. AI code review is an extra check, not a replacement.
- Automate the recurring work: hooks for formatting and linting, skills for workflows such as release notes or dependency updates, MCP servers for tickets and documentation.
- Tighten the rules: wherever the agent repeats a mistake, a rule goes into the project instructions.
Vibe coding versus agentic coding
Andrej Karpathy coined vibe coding in a post on X on February 2, 2025. He described letting an AI write code, accepting every change without reading the diffs and pasting error messages straight back in. His own verdict: fine for throwaway weekend projects.
For prototypes, click-through mockups and personal tools, that is legitimate. For production code a team maintains for years, it lacks exactly what protects quality: someone who reads the diff. We teach agentic coding with discipline: the agent takes on well-scoped, verifiable work, and the engineer decides what ships.
If you want to learn to build software with AI without knowing how to code, this training is the wrong fit. A beginner vibe coding course, or our AI Champions Program, where business teams build their own automations and agents, will serve you better.
Online course, public class or team workshop
Specialty Tokens offers the third format and wrote this page. Each format has its place:
| Format | Fits when | Strength | Doesn't fit when |
|---|---|---|---|
| Free vendor courses, such as Claude Code 101 in Anthropic's Claude Academy | one developer wants to learn the features | free, self-paced | the whole team needs shared rules |
| Public classes from training companies | one or two engineers want structured instruction | fixed dates, exchange with other companies | the team needs to practice on its own codebase |
| Private workshop on your own codebase (Specialty Tokens) | a whole engineering team is adopting agents | real tickets, shared project instructions, agreed rules | only one person needs training |
We work on site in Vienna and across Europe, the UK and the US, in English or German.
What this looks like in practice
At Nyra Health, which builds digital therapy software for neurological rehabilitation, we ran a four-hour hands-on workshop on Cursor and Claude Code for an experienced, skeptical engineering team. The goal was to move faster without lowering the product bar. By the end, the team was trying tools and workflows it had not used before.
Teams without engineers have other formats: Claude training for everyday work, ChatGPT training, Microsoft Copilot training, Power Automate training for automations in Microsoft 365, and the AI Champions Program. To open your own systems to agents through MCP, talk to us. This training is part of our AI training for employees.
