Claude Code training and agent audits, on your repo
Four formats. Three tie nobody up, the async audit, the one-hour office hours and the build I code alone. The one that takes your team off the floor for a whole day is the workshop.
Prices in euros, excluding VAT. French clients pay 20% VAT on top. I take no staff augmentation and no long engagements, only these four formats, and a fixed price that is signed does not move.
Agentic setup audit
€1,500 excl. VAT per day, fixed price, two to four days
I take your repo and look at what actually happens when an agent works in it: which CLAUDE.md (the instruction file the agent reads) loads, which hook (a script run before or after an agent action) fires, and what a security rule is worth when the agent reaches the same result by a different path than the one it watches. The first move is to strip everything out, CLAUDE.md, skills, hooks, and see where the agent fails on its own before deciding what to put back. That is the principle of hobbling: remove first, modify next, add last, and what protects (read-only access, permissions) stays, only what compensated for the model’s limits goes. Most of the work is async, with two or three short interviews, so it does not tie up the team.
- The CLAUDE.md files, the skills, the hooks, the permissions, and who owns them
- The settings set at team level that do not do what everyone assumes, with nothing to flag the gap
- What gets checked before a merge, and review when it comes from an agent, which sometimes gives a verdict on a whole PR after checking only part of it
- Token cost, where it goes, and the spread between your developers on the same work
- What all of it changes, or does not, in what you ship
At the end you have a written playbook, specific to your repo, with a 90-day plan and an hour to walk through it. The document stays with you and is built to be run without me.
Workshop in your own repo
€2,500 excl. VAT per day, billed by the day and never per head
One day on site or remote, on your own code, with your real tickets. No sample project and no sandbox, what we write during the day is committed to your repo that evening. The hour by hour format and the questions I get before a booking are on the Claude Code training page.
- We take three or four tickets from the current sprint and push them through the agent together
- We write the repo’s CLAUDE.md files during the session, and your developers write them, not me, because a file written by somebody else has nobody to keep it up to date
- We wire up the hooks and skills that are useful at your company and throw out the ones that are not
- We look at what it costs, in tokens and in review time
With eight developers in the room the day lands a little over €300 a head, and that price does not move if you put twelve in.
Office hours
€300 excl. VAT per hour, or €1,400 excl. VAT a month for four sessions
An hour with your leads, screen shared, on whatever is blocking right now: a CLAUDE.md that has doubled in size and nobody rereads, a hook that blocks the whole team, a bill climbing with nobody sure why, a PR written by an agent that nobody quite dares to review.
- Per session when there is a precise question and an hour is enough
- On a monthly subscription (retainer), four sessions a month, when there is a piece of work that runs for a while, and a slot held every month costs more than an hour picked out of a gap in the calendar
The recurring format exists because models change fast. Each wave (Opus 4.5 in November 2025, Opus 5 in July 2026, Opus 5.5 in September 2026) changes what the agent can do on its own, an instruction that helped six months ago can hold it back today, and the lag of a team that does not rerun the test grows faster than it used to. It is not a guarantee, it is what I see at my clients, and I have no independent measure of the gap.
Building an agent system
€7,500 excl. VAT per week, two weeks minimum
For when what you need is not an opinion but shipped code. It is my core trade, ML and then LLMs in production since 2017, and it sells in full weeks rather than in hours scraped together between meetings.
- An agent wired into your product, an eval harness that tells you whether the answers are any good, a RAG (the agent looks things up in your documents before answering) that holds up under load, an inference platform that does not fall over
- I write the code, your developers review it, and all of it lives in your repo from day one
- At the end there is a system that runs, tests, and a handover so the team carries on without me
What this format is not is staff augmentation, one more developer on the team for six months, that is neither what you are buying nor what interests me.
Who this is for, and who it is not for
It is for any development team, and also for product managers and designers: from the first try to adoption that stalls, by way of the agent that returns less than it should.
When I am not the right choice
- You are looking for OPCO funding (French company training funds) or a Qualiopi-certified training body (the certification of French training bodies). I am not one, the engagement comes out of the team’s budget, like consulting.
- You are looking for one more developer on the team, as staff augmentation, for several months. The build format exists for shipped code on a written scope, not to fill a seat.
- You want a usage dashboard to show a steering committee: I can build one, but it will not tell you whether the team is gaining, see below.
How I read cost
Cost reads as time to value: time to merge on code that passes review, and up to the business when the team measures it. Claude Code already exports its line and token counters through OpenTelemetry (a standard for exporting metrics), and nothing in those counters tells an agent looping on nothing apart from real output. I look at the most expensive session, to find out whether it was wanted, calibrated or duplicated, rather than trying to cap it: a low cap mostly signals under-use, and self-serve guardrails that make people think do better. An agent bill compares to a salary, not to a licence.
How this differs from a catalog Claude Code course
Serious Claude Code courses exist in France, and when a team is discovering the tool they are the right format. The difference with what I do is one of kind: I turn up in your repo, on your tickets, and what comes out of the day stays there. I went through their public prices and programs in August 2026 in a separate piece, along with what that method is worth and what it does not prove.
None of the formats has a tool in its name, and that is deliberate. I work mostly with Claude Code, but what I maintain is the harness, and the text of the instructions carries over from one tool to another, not the hooks or the permissions, which live in a single tool. At a large company the choice of model vendor is made very high up and can change mid-year for reasons that have nothing to do with engineering, and what was written into the repo has to survive that.
A first conversation
Twenty minutes, no deck. You tell me what you have deployed, since when, and what makes you think it is not returning what it should. I tell you which of the four formats makes sense, and if it is none of them I say that too.
Tell me your situation, what is not working, and since when, I reply if that gives me enough to judge. If you give me read access to the repo before the call, I look at it before we talk, and we start from what I found there instead of your guesses.