Your people leave running Claude, not briefed on it.
We enable trusted AI. We build what we teach.
Hands-on training on Claude Code, Claude Cowork and agents, run on your own repos and workflows, with the controls set before day one.
Three formats- Half-day intensive4 hours · one capability, one team
- Full-day workshopMost teams8 hours · all four modules
- Multi-week bootcamp2 to 4 weeks · a real project
Azure
The gap isn't awareness. It's muscle memory.
Most teams have seen a demo. Few have shipped a change, run a recurring task or tuned a permission rule with Claude. That's what a bootcamp is for.
- 01On your own workEngineers work in their repos from the first session. Operators bring the documents and tasks they do every week. Nothing is a sample project.
- 02Controls before day onePermissions, sandboxing, audit export and spend limits are configured with IT before the first session, so the team learns inside the rules it will keep.
- 03Led by people who build thisOn-site or virtual, led by engineers who run Claude Code and agents in production. No recorded modules. Office hours after, so habits hold.
- Engineering teamsClaude Code, agents, MCP
- Operators and analystsClaude desktop and Cowork
- LeadershipControls, visibility, budget
- Teams that build agentsAgent SDK, Managed Agents
The tool alone doesn't pay. Setup, habits and a scorecard do.
Published results split cleanly: teams that changed how they work gained, teams that just turned it on didn't.
- 67%more merged pull requests per engineer per day after Claude Code rolled out across Anthropic's own engineering teamAnthropic, Aug 2025 survey of 132 engineers, self-reported
- 25 to 30%productivity when AI tools come with end-to-end process change, against 10 to 15% from the tools aloneBain Technology Report, 2025
- 21 to 40%gains for less tenured developers in field experiments at Microsoft, Accenture and a Fortune 100 manufacturerCui et al., field experiments
- 19%slower: experienced developers new to the tooling took longer on their own repos, while believing they were 20% fasterMETR randomized trial, July 2025
- 7.2%drop in delivery stability as AI adoption rose, in Google's DevOps Research and Assessment survey of software teamsDORA, 2024
- 30 to 40%of developers given a coding assistant never tried itCui et al., Microsoft field study
What this means for a rollout. Expect a dip before the gain, train on the team's own code, pair every throughput number with a stability number, and treat adoption as something to manage, not assume. The same Google research names a clear AI policy as one of seven things that make AI a net positive.
Four modules, each on the team's own work.
- Module 1EveryoneClaude fundamentals
- How the model works in practice, and when not to trust the output
- Prompting that holds up: context, examples, constraints
- The context window, projects and memory
- What to never paste in, and why
- Module 2Operators, analysts, leadershipClaude desktop and Cowork
- What Claude takes on, and what stays with you
- Projects, skills and connectors into files, email and systems, with approvals on
- Multi-step work: research, analysis, documents, recurring reports
- A shared prompt and skills library the team keeps
- Module 3Engineering teamsClaude Code for engineers
- Setup, permissions, sandbox and CLAUDE.md for each repo
- Explore, plan, code, commit; plan mode first; verification Claude can run
- Skills, subagents and hooks for team rules; headless runs and CI
- Code review on pull requests; model and effort choices
- Module 4Engineers and buildersAgents and MCP
- MCP: connecting Claude to databases, APIs and internal tools
- Claude Agent SDK and Managed Agents: when to use which
- Sessions, memory, evals and cost per task
- Enterprise administration: managed settings, spend limits, audit
Three formats. Pick by how deep the team needs to go.
All three are live, on-site or virtual, and run on the company's own repos and workflows. Hybrid works for distributed teams.
- Half-day intensive4 hours · one capability, one teamClaude Code for an engineering team, or Claude desktop and Cowork for an operations team. Best for first-time adopters who want to leave with one working habit.
- One module from the curriculum
- Pre-work: accounts, environment, a prep guide a week ahead
- Everyone ships one real change or runs one real task
- Shared prompt or skills library started
- Most teamsFull-day workshop8 hours · all four modulesMorning on fundamentals and controls, afternoon in hands-on labs on the team's own repos and documents. Claude Code, Cowork, MCP basics and prompt engineering.
- All four modules, labs in the afternoon
- Controls configured with IT before the day
- Repo CLAUDE.md files and a starter skills library checked in
- Leadership briefing at the end of the day
- Multi-week bootcamp2 to 4 weeks · a real projectWeekly sessions, homework between them, and a real workflow or agent built by the team with one-on-one coaching. Adds Managed Agents, custom skills and enterprise administration.
- Everything in the full-day workshop
- Weekly office hours and one-on-one coaching
- One workflow or agent built by the team, in production
- Champion team and the scorecard handed over
Before, during and after: what happens and who does it.
The work before the first session is what makes the sessions land. The work after is what makes them stick.
- BeforeOne to two weeks aheadWe doBaseline usage and task times from Git and the team's tooling. Collect three to five repeated weekly tasks and the documents and repos behind them. Configure controls with IT: permissions, sandbox, audit export, spend limits. Send a prep guide a week ahead.You doName a sponsor, an IT contact and two or three champions. Provision company Claude seats (we can help). Engineers confirm a working dev environment.
- DuringThe sessionsWe doTeach on the team's own repos, documents and tasks, never samples. Check in repo CLAUDE.md files and a starter skills library. Run the leadership briefing on controls, visibility and budget. Tune permission rules as real usage shows what's needed.You doShow up with real work. Every engineer ships one reviewed change; every operator runs one real task.
- AfterTwo to four weeksWe doWeekly office hours and, in the multi-week format, one-on-one coaching. Measure adoption, time saved, output quality and confidence against the baseline. Hand over the scorecard method and a ranked list of agentic opportunities we saw.You doChampions onboard the next cohort and new hires. Leadership reads the four scorecard numbers monthly.
Five layers of control, each set by IT and enforced on every machine.
- 01Permission modeHow much the agent does before askingStart in the default mode: reads run, everything else asks. Plan mode for exploration. Bypass mode disabled for everyone.
- 02Allow, ask and deny rulesWhich commands, files, domains and toolsDeny reads of .env and credentials, deny destructive git and rm. Allow the test runner and linter. A deny can't be overridden by a lower setting.
- 03SandboxWhere commands can write and reachWrites limited to the working directory. Network through an allowlist that starts empty. SSH and cloud credentials masked.
- 04HooksRules that run as code, not adviceBlock a command before it runs, scan an edit for secrets, log every tool call. Managed hooks can't be removed by a user.
- 05Managed settingsThe file that holds all of the aboveDeployed by MDM (Jamf, Intune, Group Policy) or from the admin console. Locks the MCP server list, the model list and the effort cap.
We write the managed settings file with IT before the first session, test it against the team's real repos, and tune the allow list in office hours as usage shows what's needed.
Specific habits, and the misuse each one prevents.
These come from the published studies and from our own delivery. Each is a rule a team can check in a code review.
| The habit | What it prevents |
|---|---|
| Plan before code on anything touching more than a few files | Letting the agent rewrite half the repo to fix one bug |
| Give Claude a test, build or screenshot it can check itself | Accepting a diff because it looks right |
| Keep pull requests the size they were before | PR size up 51% and review time up 5x on high-adoption teams |
| Read every line before you approve, same as a junior's PR | Stability drops while throughput looks great |
| Sonnet for daily work, Opus for architecture, Haiku for bulk jobs | Running the largest model on everything and blowing the budget |
| One session writes, a second session reviews | The author grading its own homework |
| Deny .env, credentials and destructive commands in settings, not in CLAUDE.md | Relying on advice the agent can ignore |
| Log what worked in CLAUDE.md and skills; prune it monthly | A 600-line CLAUDE.md nobody reads, and the agent doesn't either |
Seven things the team has when we leave.
Each one is checked before we close the engagement. If one isn't there, we stay until it is.
- 01Everyone productive on Claude desktop and Cowork within a weekEach person has one recurring task running, with approvals on and the right connectors.
- 02Engineers on Claude Code with shared configurationRepo CLAUDE.md files, a skills library and permission rules checked in. One workflow running headless with a person reviewing.
- 03Controls configured and documentedManaged settings, deny rules, sandbox, audit export and spend limits, owned by IT, with the compliance answer set filed.
- 04A prompt and skills library the team keepsWhat worked, in a form the next hire can use on day one.
- 05A scorecard and the method to run itBaseline and week-four read on adoption, throughput, stability and cost per developer. Leadership reads it monthly without us.
- 06A ranked list of agentic opportunitiesWhere an agent could run a step of the work on its own, scored on impact and cost, from what we saw in the team's workflows.
- 07A champion team that onboards the next cohortTwo or three people per team who ran the sessions with us and own the library and the rules after.
Four numbers the company can read itself, each with a source.
We set the baseline before the first session from Git and the existing tooling, and hand over the method. Leadership reads the same four numbers monthly.
- AdoptionWhatWeekly active share of licensed seats; sessions per engineerSourceClaude Code analytics dashboard; OpenTelemetry session countsAnthropic's own target: 70% weekly active
- ThroughputWhatLead time for changes; PR cycle time; merged PRs per engineerSourceGit and the PR system, the same four keys Google's DORA research usesExpect a dip in the first two weeks, then a gain
- StabilityWhatChange failure rate; incidents per PR; median review time; PR sizeSourceIncident tracker, CI and the PR systemMust hold flat. If it moves, throughput doesn't count
- CostWhatCost per developer per active day; cost per merged PRSourceOpenTelemetry cost metrics or the spend report, per userAnthropic's published average: about $13 per developer per active day
Numbers not worth reporting: lines of code, suggestion acceptance rate, and share of code written by AI. Agents inflate all three. The one decision leadership makes up front is what freed capacity is for (more features, lower spend, faster releases), or the gain never reaches the P&L.
How it's capped, and the levers that move it.
Claude seats and usage are billed by Anthropic to the company directly. These are the controls that keep a team inside its budget.
- Spend limits by group and personSet at the org, group and individual level. Start with per-group caps, define light, standard and power tiers, alert at 75% and 90%.
- Model routingSonnet for daily work, Opus for architecture, Haiku for bulk runs. Managed settings fix the available models.
- Effort capsManaged settings set a maximum effort per deployment. Most work runs at medium.
- Context hygieneCLAUDE.md under 200 lines and workflow detail in skills that load on demand. Shorter context is cheaper on every call.
- Per-user visibilityThe analytics dashboard and spend report show cost per user, so engineers see real numbers.
Three ways to start.
Each starts with a short conversation and ends with a dated agenda. No fee until the scope is agreed.
- 01 · Send us the workThree to five weekly tasks
- We bring
- A draft agenda within two business days, matched to the format that fits, with the pre-work list.
- We'd ask
- The tasks, the repos or documents behind them, and who does them today.
- What follows
- A dated bootcamp with the controls work scheduled ahead of it.
- 02 · Compliance firstClose the open questions
- We bring
- Anthropic's documentation on training, retention, deployment options and audit, mapped to your review.
- We'd ask
- Which review is in progress and who signs off.
- What follows
- A short written answer set the compliance team can file, then the bootcamp.
- 03 · A discovery callThirty minutes
- We bring
- A readiness check: seats, environments, repos, CI, review process and who the sponsor is.
- We'd ask
- Which team or company has the clearest need and a willing sponsor.
- What follows
- A format recommendation and a proposal for that team.
Common questions.
Is our code used to train Claude models?
No, on Team, Enterprise, API and cloud-provider plans, unless the company opts in, per Anthropic's published terms. Individual Pro and Max accounts are the gap: a developer on a personal plan is outside the policy, so every seat goes through the company.
Where does the data go, and how long is it kept?
Commercial plans keep data for 30 days, and zero data retention is available for qualified Enterprise accounts. Claude Code can also run through Amazon Bedrock, Google Cloud or Microsoft Foundry, inside the company's own cloud account.
Can we audit what Claude did?
Every session and tool call can be exported over OpenTelemetry to the company's SIEM, with per-user tokens and cost, and Enterprise adds an admin audit log. We set up the export with IT before the first session.
Who signs off on code Claude writes?
An engineer, every time. Pull requests stay the gate, and median review time is one of the four numbers on the scorecard we hand over, so faster output never quietly replaces review.
Which security standards does Anthropic meet?
SOC 2 Type II, ISO 27001:2022 and ISO/IEC 42001:2023 for AI management. The reports are on trust.anthropic.com, and in the compliance-first start we map them to the review your team is running.
What can't a developer change?
Anything in managed settings: the deny list, the sandbox, the MCP server list, the model list, the effort cap, and whether bypass mode exists. IT deploys the file by MDM or the admin console, and changes are picked up within the hour.