Rolling out Claude Code in the team: training, configuration, governance

Rolling out Claude Code in a team is not an installation topic, it is a configuration and governance topic. The agent can take on a large share of routine work, but only if it gets the right context (CLAUDE.md), the right capabilities (Skills) and the right boundaries (Permissions). This is exactly where most rollouts fail: the tool is there, but nobody configured it, and without configuration Claude Code stays a clever generalist instead of a specialist for your codebase. A Claude Code training that delivers is about exactly these three layers plus handling several agents in parallel.

I (Sebastian) see the same scene in our workshops almost every week: Claude Code runs in the terminal, but there is no CLAUDE.md, no project-specific skills and no thought-through permissions policy. The agent then delivers a fraction of what is possible, and the team concludes the tool is "quite nice, but not productive". That gap is what the training closes.

Why most Claude Code rollouts stall

Three patterns show up in almost every stalled rollout. None of them is a technical problem.

No context. Without CLAUDE.md the agent knows nothing about your architecture, your conventions, your build and test commands. It guesses generically. The result is mediocre, and the team gives up before it has seen the real lever.

No skills. The same recurring tasks (a specific code-review pattern, a migration routine, a doc format) get re-explained every time instead of being stored once as a skill. That leaves the biggest repetition lever untouched.

No permissions strategy. Either the agent may do too little, then it is unproductive, or too much, then nobody trusts it. Without a thought-through permissions setup (which tools allowed, which require confirmation, which forbidden), no trust forms, and without trust no adoption.

The 3-phase approach that works

We build every Claude Code training in three phases, because adoption happens in this order: foundation first, then habit, then scale.

Phase 1: Foundation

Setup, permissions policy and data governance before anyone works productively. This is where you decide which licence tier is used, which tools the agent may run and which require confirmation or are blocked. Short, but decisive for trust in the team.

Phase 2: Guided adoption

The actual training, on real backlog tickets. Four Claude Code-specific building blocks sit at the centre:

• CLAUDE.md. The context file that gives the agent your architecture, conventions, commands and no-gos. The biggest quality lever and the most frequently forgotten one. How to layer it is in our three-layer architecture.
• Skills. Storing recurring tasks as reusable capabilities instead of describing them anew each time. That is the repetition lever.
• Permissions. A deliberate allow, confirm and deny logic for the agent's tools. That is the trust foundation.
• Multi-Agent. Several agents in parallel for independent sub-tasks, with a clear split instead of uncontrolled sprawl.

Each participant leaves the phase with a finished CLAUDE.md for their own project and three to five concrete usage patterns.

Phase 3: Scale and governance

Individual trained developers become a trained organisation. A shared CLAUDE.md and skills repository, champions per team, the measurement system (see KPIs below) and a review cadence in which context, skills and permissions keep evolving. Without this phase, what was learned evaporates after the workshop day.

This logic carries through any tool rollout. The full path from use case to production is in our AI roadmap for engineering teams.

In-house training or open course

An in-house training pays off from roughly five to six developers. We work on your real codebase, the CLAUDE.md and the first skills are built for you, and the permissions policy matches your security needs. That is the fastest route to real workflow change.

An open course fits individual developers, a first look or distributed teams. Same content, more general context. You can find our open formats under our courses.

Rule of thumb: individuals and a first test go into the open course, a whole team with a productivity goal gets in-house. How we set up in-house projects is under how we work.

Funding through the Qualifizierungschancengesetz

A Claude Code training can be funded through the German Qualifizierungschancengesetz (Section 82 SGB III) when it teaches skills that go beyond short-term, purely job-specific adjustment. The funding rate for course costs is staggered by company size: for companies with fewer than 50 employees the employer cost contribution can be waived, so up to 100 percent of the course costs are funded. For 50 to 499 employees up to 50 percent are funded, and for 500 employees and more up to 25 percent.

Since 01.01.2026 new administrative directives from the Federal Employment Agency apply to Section 82 SGB III, tying the wage subsidy more strictly to the actual training-related work absence. The application runs through the local employment agency and must be filed before the measure starts. (As of May 2026) The full overview is in AI funding for the Mittelstand 2026.

The 4 KPIs you measure success against

Without these four metrics, any Claude Code training is gut feeling. Baseline first, then train, then measure again after eight to twelve weeks.

1. Lead time to merge (cycle time). From "started work on a ticket" to "merged into the main branch". The most honest productivity number because it includes all steps.

2. Share of agent-assisted tasks. How many real, multi-step tasks per week run through the agent, not just single questions in the terminal. That is the actual lever.

3. Review load and defect rate. Speed is worthless if quality drops. Measure review rounds per pull request and post-merge defect rate. If both stay stable or fall, the gain is real.

4. Onboarding time into unfamiliar code. How fast a developer becomes productive in an unknown module. With a good CLAUDE.md, this is often where the lever is largest.

How to set these metrics up and hold them against the 1.5x productivity expectation is in the KPI framework for AI productivity.

Data governance: which licence belongs in the rollout

Depending on the setup, Claude Code runs through a Claude subscription (Pro, Max), through Claude for Work (Team with premium seats, Enterprise) or through the Claude API. For company use the data policy is decisive, and it is tier-dependent: with the Claude API and with Claude for Work (Team and Enterprise), your inputs and outputs are not used for model training by default. The consumer plans have different default settings, so there the setting has to be checked actively. For a rollout that means: the API or Claude for Work belong in the fleet, not the personal consumer plan. On price, Team premium seats (which include Claude Code) sit at around 100 to 125 US dollars per seat per month, Enterprise on request plus API tokens by usage, and the API itself by token pricing (as of May 2026).

If you need the data policies of the coding-agent vendors side by side, they are in the coding-agent comparison of Cursor, Copilot, Claude Code and in the 5 security questions for coding-agent vendors.

FAQ

How long is a meaningful Claude Code training? The guided-adoption phase is one to two days. Impact comes with the scaling phase: a shared CLAUDE.md and skills repository, champions and a measurement cadence over eight to twelve weeks. A single workshop day without follow-up structure evaporates.

What is the single most important lever? The CLAUDE.md. It decides whether the agent understands your codebase or guesses generically. A well-layered CLAUDE.md is the difference between "quite nice" and "productive".

How do we stop the agent from being allowed too much? Through a deliberate permissions policy: reading and preparatory tools allowed, modifying or external actions requiring confirmation, critical operations blocked. That is part of the foundation phase and the basis for trust.

Is the training eligible for funding? Through the Qualifizierungschancengesetz (Section 82 SGB III) often yes, with a rate staggered by company size. The application must run through the employment agency before the start. (As of May 2026)

Read more

• GitHub Copilot training for engineering teams
• Cursor training: from tool access to a productive agentic workflow
• CLAUDE.md, Skills and the three-layer architecture
• Cursor vs GitHub Copilot vs Claude Code compared
• Headless coding agents in the CI/CD pipeline
• KPI framework for AI productivity (1.5x)
• AI roadmap for engineering teams (5 phases)
• AI funding for the Mittelstand 2026

Sources

• Claude Plans and Pricing, claude.com; Claude Help Center, Enterprise plan and data usage (as of May 2026)
• Section 82 SGB III, gesetze-im-internet.de; Federal Employment Agency, administrative directive on Section 82 SGB III, effective 01.01.2026
• Sentient Dynamics workshop aggregate (DACH Mittelstand clients, 2025-2026)

How Sentient Dynamics can help

We roll Claude Code out in the team so that it becomes genuinely productive: foundation with a permissions policy, guided adoption with CLAUDE.md, skills and multi-agent on your real codebase, a shared repository and the measurement system for the 4 KPIs.

Book a demo