The machine underneath
How a model behaves: stateless, non-deterministic, working from a fixed budget of context you pay for on every request. Plus who trains these models, what motivates them, and which organisations to follow.
In-house training · Effective use of coding harnesses
You approved the tools.Nobody audited how they're used.
Two days, on your premises, for 5–10 of your engineers and architects. They leave able to read what any coding harness is doing: the context budget, the loop, the tool surface, the review gate. They make their own calls without waiting for a vendor to tell them.
Format
2 days on site
Group
5–10 people
Price
€2,000 / person
Lead time
3 weeks
Recorded
No
01 · The cost you are already paying
Unmanaged adoption is not free.
It is just unbudgeted.
None of the numbers below are ours. They are the best public data on what happens when a team adopts coding agents without a shared operating model. Run them against your own last quarter.
+200%
Reviewer time spent on AI-assisted pull requests
LeadDev, 2025
+54%
Bugs per developer
LeadDev, 2025
66%
of developers name “almost-right” AI code their number-one frustration
Stack Overflow, 2025
1 in 5
security breaches now tied to AI-generated code
LeadDev, 2026
3%
of developers highly trust AI output, while 84% use it daily
Stack Overflow, 2025
60–65%
of engineering leaders report no team productivity gain from AI
LeadDev, 2025The pattern is consistent: adoption is near-universal, trust is not, and the review queue absorbs the difference. That gap is a training problem, not a tooling problem. No additional licence closes it.
02 · What your people can do afterwards
Six things they cannot do today.
Stated as capabilities, not topics, so you can check them at the two-week mark and know whether you got what you paid for.
- Read what a harness sends on every request, and say what it cost.
- Decide what to expose to an agent, what to add, and what must stay away.
- Write a prompt that carries its own way to verify the answer.
- Diagnose why output quality degraded mid-session, and recover the context.
- Evaluate a new tool against the layer underneath, not against its marketing.
- Argue against a configuration file or shared skill that should not exist.
03 · The mental model
Tools change. The layer doesn't.
Three layers, and every tool on the market has them. Copilot, Claude Code, Codex, OpenCode, Cursor. The branding differs, the layers do not. Teach the layers and your team can pick up whatever you adopt next without repeating the learning curve.
HARNESS
Day 1 · Module 03dispatch · permission gate · context-window management · file access
LOOP
Day 1 · Module 02plan → execute → verify · task decomposition · four kinds of context
MODEL
Day 1 · Module 01stateless · non-deterministic · fixed context budget, paid per request
Day 2 works outward from there: configuration as a shared asset, the tool surface and what it costs, and where the review gate sits once a human can no longer read everything that arrives.
Working the loop
Plan → execute → verify, task decomposition, and the four kinds of context an agent needs: system prompt, conversation history, tool outputs, and retrieved context.
The harness
The loop around the model: how work is dispatched, how a turn ends, where your permission gate sits, and what the tool does silently on your behalf. This is the layer shared by every product on the market.
Configuration as a shared asset
Instruction files, shared skills, and, just as important, the discipline of knowing when not to add another one. Concrete examples such as .cursorrules, CLAUDE.md, and similar patterns.
Tool surface and cost
What is currently exposed to the agent, what to add, and what needs to stay away, and how each of those decisions moves your token costs.
Review gates via verification
In the age of LLMs the human becomes the bottleneck in review. This is where tooling, automation, and advanced testing practices come into play.
04 · Evidence
Delivered, not theorised.
8x
more code deployed to production, with no loss of quality
A team at ASML, after this training and hands-on coaching from Stas. Individual feedback from team members is on LinkedIn.
Over the course of three workshops, Lucian helped us implement an entirely new testing paradigm that has already saved us countless hours of manual effort while also improving the quality of our codebase. The workshops themselves were engaging, interactive and well-structured, and the impact they've had on our products, and business, is hard to overstate.
At the moment, LLMs might go unchecked and suggest tests that are either redundant or that miss important corner cases. During the workshop, Lucian introduced useful techniques to help developers and AI reason about our tests and the problem space.
Lucian Ghinda
Ruby and Rails developer and product engineer, 20+ years in tech. Has been hands-on implementing LLM features at Indeed Flex, part of Indeed. Ran corporate training for more than five years with METRO, and has continued teaching solo since. His "Good Enough Testing with LLMs" workshop has been delivered repeatedly, online and on-premise. Creator of the Short Ruby Newsletter.
Stanislav Katkov
Software contractor, open-source maintainer, and startup founder. He has been working closely with LLMs for two years and training other engineers to use them. His research and day-to-day work centre around LLMs and developer tooling (CLI and MCP), and he builds his own project, Posh TUI, alongside other open-source projects.
05 · What this costs you
The invoice is the smaller number.
Most training pages quote a fee and stop. Your people's calendar costs more than the invoice does. Both are below, so the decision can be made once instead of discovered halfway through.
Training fee
€10,000 – €20,000
€2,000 per person, 5 minimum, 10 maximum. VAT not included.
Your team's time
10 – 20 person-days
Two days off the board for everyone in the room. Usually the larger line item.
Your time before
1 hour
The discovery call, with one team member and one team lead.
Our time before
3 weeks
Preparation and tailoring. Yours to schedule around, not to staff.
Afterwards
Optional
Office hours at €300/hour ad-hoc, or €250/hour on retainer from 10 hours a month.
Which budget this comes from
This sits in training or L&D, not in tooling. It is a per-engagement professional-services fee invoiced by KROONI B.V. (Netherlands, KVK 96156538). This is not a per-seat subscription, and it does not renew.
Against the alternative
Trial and error has no invoice, which is why it never gets compared. It is paid in tokens nobody reconciles, in review hours nobody logs, and in the defects that reach production before anyone notices the pattern. That cost is recurring, and already on your books.
06 · Is this you
Match your situation to the row.
We would rather you disqualify yourself here than three weeks into a preparation cycle.
Tools are adopted. Usage is inconsistent and nobody has written down why.
yesThis is the core case. Both days apply.
Review is the bottleneck. More arrives than anyone can properly read.
yesDay 2, Module 06 is built on exactly this.
Token and tooling spend is climbing and nobody can explain the shape of it.
yesDay 2, Module 05 is the tool surface and its cost.
Architects have to choose a platform and cannot evaluate the claims.
yesThe three layers are the evaluation framework.
You need a written internal policy for AI-assisted code.
partlyModule 04 gives the shape and the trade-offs. Writing and enforcing it stays internal work.
Nobody on the team has used a coding agent yet.
noStart them on the tools first. We are not the on-ramp.
You want to build an agentic product: LangGraph, CrewAI, MCP servers.
noDifferent discipline, different market. We teach using these tools, not building them.
You want something recorded and self-serve for the whole org.
noThis is not recorded. Ten people is a hard ceiling.
07 · Where this will not help
The limits, before you find them yourself.
Every one of these has happened or is a known failure mode. We would rather plan around them in the discovery call than discover them in the room.
Two days seeds behaviour change. It does not complete it.
Shared judgment holds when someone keeps it alive afterwards: a follow-up rhythm, an internal owner, or office hours. A single session with no follow-through decays like any other training.
If corporate IT blocks the tools, the hands-on half does not run.
Attendees need working tool access on the day. We confirm this in the discovery call and keep a recorded fallback, but it is a genuinely worse session.
We are not a substitute for your testing and observability.
We teach where the review gate sits and what has to hold it. If the safety net underneath does not exist yet, the training tells you what to build. It does not build it.
A very mixed room dilutes both ends.
Seniors get bored, juniors get lost. The discovery call captures current usage so we can pitch it correctly, but a group spanning first-week to expert is the hardest version of this.
We will not hand you a recipe.
No prescriptive “do it this way.” If what you want is a standardised checklist to roll out unchanged, this is the wrong purchase. And in a field moving this fast, that checklist would be wrong by spring anyway.
08 · Making the case internally
The paragraph you can forward.
You will have to justify this to someone. Take this, edit it, send it.
We have coding agents in daily use across the team, but no shared standard for how they are used or how their output is verified. Industry data puts reviewer time on AI-assisted pull requests up around 200% and roughly one in five security breaches now tied to AI-generated code. We currently have no way to tell whether that pattern applies to us. I want to bring in two days of in-house training for [N] engineers and architects (€2,000 per person, delivered on site, tailored to our stack after a discovery call). The outcome is a shared operating model for these tools and a verification discipline for what they produce, taught by practitioners rather than a tool vendor, so it stays useful when we change tools. Comparable delivery has produced measurable results elsewhere: a team at ASML deployed 8x more code to production with no loss of quality after the same training and coaching.
Replace [N]. Everything else is accurate as written.
09 · Next step
One hour, then you decide.
The discovery call is a working conversation, not a sales call. We go through your stack, your current usage, and the problems you have. If the fit is wrong we will say so on that call rather than sell you two days you do not need.
- No slides, no proposal deck
- One team member and one team lead
- We tell you if it is a bad fit