A day of Claude Code training on your repo and your tickets
One day in your code, on the tickets of the current sprint. Your developers at the keyboard, me next to them. We first bet on what the agent will manage on its own, look at where it succeeds and where it fails, and check whether your CLAUDE.md files and hooks hold when it switches tools. What gets written during the day is committed that evening.
Why in your repo and not in a classroom
From what I see, what separates a team that gets something out of Claude Code from a team that has licences is rarely knowledge of the tool. It is what is written around it, in the repo: CLAUDE.md (the instruction file the agent reads), skills (procedures it loads on demand), hooks (scripts run before or after an action), permissions, and above all whether the agent actually follows it. One example that has moved: until version 2.1.287 of October 1, 2026, a subfolder’s CLAUDE.md only loaded when the agent opened a file in that folder with its read tool. Since then writing and editing trigger it too, but a public issue still open on October 4 shows it still does not through the shell (cat, grep, find, heredoc). I expect the next fix to target the shell, and that is why it gets retested at every version. Another: since version 2.1.277 of September 18, Claude Code reads AGENTS.md as a fallback when there is no CLAUDE.md, a setting (claude-md-and-agents-md) makes it read both, and a personal CLAUDE.local.md is enough to cut the fallback, and with it every shared instruction, without a single warning.
None of that shows on a sample project, which has no debt, no odd convention inherited from 2019, and no flaky test everyone works around.
A catalog Claude Code course is the right format when a team is discovering the tool. Across the forty or so French catalogs my agents went through in late August 2026, only two put hooks and skills on the program, and none covers secrets handling when an agent is the one reading the repo. The day I sell starts from your tickets, whether the team is at its first try, adoption is stalling or the agent returns less than it should.
What happens during the day
In the morning we take three or four tickets from the current sprint, real ones, and push them through the agent together. Bare first, the --safe-mode option of Claude Code, which its documentation presents as a troubleshooting tool, starts it with no CLAUDE.md, skills, hooks or MCP (the connectors to your other tools), and we look at where it fails on its own. That usually says more than rereading the CLAUDE.md written a year ago, it shows what is missing, and it is what I redo at every new model. We also compare machines: Anthropic varies some settings from one install to another from its servers, and two developers on the same team can be running under different instructions without knowing it.
Before each try, each developer writes down a bet, success, partial or failure. We then note whether the result is mergeable as is, mergeable with edits or thrown away, the time, the tokens, and how many times the human stepped in, then discuss what made the difference: no way to check, context left in people’s heads, a tool not plugged in. The gaps between your bets and the results are often the best moment of the day, and they draw the map of what the agent gets right in your repo.
For every guard your team thinks it has, a nested CLAUDE.md, a hook, a rule that denies a sensitive file, we check whether it is scoped to a tool or to an effect. A rule declared on one tool (file editing) does not follow the agent when it reaches for another (the shell) to get to the same place, whereas a rule on the effect, like the sandbox, an enclosure the system imposes on the agent’s commands, holds whatever the tool. That is what gets checked first, before writing anything new.
In the afternoon your developers write the repo’s CLAUDE.md files, not me, during the session. We wire up the hooks that are useful at your company and test them for real, a hook is a script, and when it exits with code 1 it reports an error but lets the action through, only code 2 blocks it. We finish on cost, with the numbers from your own account. The tool estimates tokens at list price rather than your contract rate, and no metric tracks review time, so we rebuild it from your own timestamps.
In the evening what was written is committed to your repo. There is no slide deck to take home, the material is your code.
The price, and what it comes to per person
€2,500 excl. VAT for the day, billed by the day and never per head. With eight developers in the room that is a little over €300 a person, with twelve it drops to around €210. Past fifteen or so it stops making sense, nobody writes anything together any more and it turns back into a talk.
For scale, I went through the public prices of Claude Code courses in France in August 2026, they run from €900 to about €1,900 excl. VAT per person for one or two days in a classroom. The full list, with what it proves and what it does not, is in a separate piece.
I am not a Qualiopi-certified training body, so in France the day cannot be funded through an OPCO. It comes out of the team’s budget, like a day of consulting.
Questions I get before a booking
Does the team need to be using Claude Code already?
No. A team that is starting gets its first tickets done with the agent, a team that has used it for months measures where it returns less than it should. If you only want to learn the tool step by step, a course built for beginners, Zenika’s for instance, is sold with no prerequisites and costs less than my day.
We are on Cursor, Codex or Copilot, does it still work?
Yes, for the most part. I work mostly with Claude Code, but the text of the instructions carries over from one tool to another, not the hooks or the permissions, which live in a single tool. Shared instructions can live in AGENTS.md, an open format now stewarded by the Linux Foundation, which Cursor, Codex and Copilot read, and which Claude Code reads since September 2026 as a fallback when there is no CLAUDE.md.
On site or remote?
Both. Remote anywhere, screen shared and repo open on both sides, on site anywhere in France, and in Europe or further for the bigger engagements, with travel costs on top. I ask that the team sits in the same room on your side, a day across twelve Meet windows writes nothing. Very short interventions work too, see the office hours.
How many developers?
Between four and fifteen. Below four, take an hour of office hours instead. Above fifteen, we split it into two days.
What do you need beforehand?
Read access to the repo a week ahead, and the list of tickets for the sprint. I read the repo before I come, and we start from what I saw there rather than from your assumptions.
What about the NDA and security?
I sign yours. Nothing leaves the repo, I keep no copy of the code, and I already work under NDA with large companies. On the tool side, WebFetch sends Anthropic the domain name of every page it is about to read, for a preliminary check. A setting turns it off, and we set it together if your policy requires it.
What is left a month later?
The committed files, written by your developers, and a habit, reviewing the CLAUDE.md like code. The gain fades if nobody looks after it, which is why the office hours format exists, to come back one hour a month without redoing a full day.
Is this an official Claude Code training?
No. The official course exists, Claude Code 101, free on Anthropic’s learning platform. I have no financial tie to Anthropic, Codex or Cursor, no certification and no partnership, and my advice holds for all three. What I sell comes from nine years of ML in production and daily use of coding agents in real company repos.
Book a day
An email with two lines of context, how long agents have been in the repo and what makes you think they are not returning what they should.