---
title: "A controlled development environment, ready for coding with AI"
description: "For Silico, a private registry of development environments described as code, and an AI usage framework for developers: allowed tools, approval gates, indicators defined before any measurement."
url: "https://mc-stud.io/en/work/coder-developer-environments"
author: "Matisse Chenavas"
date: "2026-04-08"
lang: "en"
ai: "true"
---

# A controlled development environment, ready for coding with AI

> Silico’s developers have standardised environments, described as code and hosted on controlled infrastructure, the foundation of a governed use of AI.

- **Client**: Silico
- **Sector**: IT services
- **Period**: 2026
- **My role**: Registry designer, contributor to the AI usage framework
- **Setting**: Direct
- **Team**: Solo · 1 person
- **Technologies**: Coder, Docker, Git, Terraform, Devcontainers
- **Reading time**: 3 min

## At a glance

- **Stake**: Before opening code to AI assistants, an IT services firm must know where they run, which repositories they reach and what trace they leave.
- **Mechanism**: Coder templates build on reusable Terraform modules, and written design rules set how to add new ones.
- **Proof**: The registry and its documentation, which I wrote alone in one week in early April 2026.
- **Impact**: A new workstation or project starts from a shared, versioned base, and the question "where may the assistant work?" gets a technical answer.
- **Limits**: Team adoption of the registry is not measured, and no AI-related gain is claimed.

## The starting point

In a team that moves from one client project to the next, every developer ends up with their own environment, slightly different from their colleagues'. The gaps stay invisible until the day they cost half a day of troubleshooting.

When AI assistants enter daily work, the question gets more serious: which machine does the assistant run on, which repositories can it reach, and what trace is left? A written rule is not enough if the environment does not carry it.

## What I built

I designed a private registry of templates for Coder, the tool that creates developers' workspaces. Each template, devcontainer or envbuilder, builds on reusable Terraform modules: the project database, the application environment, Git submodule handling. Written design rules explain how to add a template without duplicating what exists.

The structuring choice was to describe environments as code. An environment becomes a versioned, reviewed and reproducible artefact; fixing it once fixes it for everyone.

### A framework for coding with AI

On top of that foundation, I contributed to defining practices for a controlled use of AI in development[^iso]: which tools are allowed, which repositories need an extra approval before merge, how to evaluate an assistant on real tasks, a feature, a fix, a refactor, before adopting it.

## What the evidence shows

The registry and its design rules are the direct evidence of this work[^git]. For the AI usage framework, the evidence is of another kind: indicators defined with their method, ready to be collected[^baseline].

> [!NOTE] Indicators defined before they are measured
> Code review throughput, post-merge defect rate, test coverage, review cost: each indicator has its measurement method and a baseline to collect before rollout. I claim no gain until that measurement exists.


## What it changes

A new workstation or project starts from a shared, versioned base; gaps between environments shrink by construction. For AI adoption, the environment the assistant works in becomes something you can review and control.

My takeaway: the safety of AI use in code is decided upstream, in that environment. It is a short piece of work, but it conditions what follows: developers' trust, management's trust, and the ability to measure.

## Limits

- The number of developers and projects using the registry is not measured to date.
- The framework's indicators are defined but not yet measured: no before-and-after gain is claimed.
- The framework is part of an alignment effort with ISO/IEC 27001; it is not a certification.

[^iso]: Within Silico's governance effort, aligned with ISO/IEC 27001. The registry itself contains nothing specific to the standard.

[^git]: Sole author of the registry: 18 commits, from 1 to 8 April 2026.

[^baseline]: Baselines are collected during a discovery phase, and targets are set with the engineering and security leads.

---

Matisse Chenavas, Founder of MC Studio. https://mc-stud.io/en/work/coder-developer-environments