Alain Rivelon · Development workshop

Building Scutaria with AI

I design the roadmap with Claude and ChatGPT, then organise development with VS Code, Cline and DeepSeek for the software, Codex and Claude Code for the website.

I build Scutaria as a solo project, starting from questions in my GRC work. My role connects business requirements, architecture choices, coding instructions and verification.

01 · Before coding

The roadmap comes before coding instructions

With Claude and ChatGPT, I explore options, ask for explanations and challenge proposals. I decide scope, simplifications and deferred work. Assistant roles overlap: a second response alone is not independent validation.

  1. Explore the need

    Business questions, specifications, concepts and architecture alternatives. Understand what a technology choice enables before adopting it.

  2. Connect two roadmaps

    The product roadmap describes the expected result. The technical roadmap identifies prerequisites, dependencies and checks. The backlog preserves deferred options.

  3. Define executable stages

    Turn a plan into bounded instructions: allowed files, a precise change, expected results and stopping conditions.

The plan was not complete at the outset. Prototypes, reviews and incidents changed priorities. Orchestration components were implemented and later frozen while a manual workflow remained usable.

02 · My environment

Two projects, tools with distinct roles

Claude · ChatGPT / CodexDesign and review
Specifications, roadmaps, architecture, criticism and coding instructions. Codex and Claude Code have also contributed to implementation.
VS Code · Cline · DeepSeekSoftware implementation
Cline reads and edits scoped files and runs authorised commands. DeepSeek supplies model responses; Flash/Pro routing remains manual.
Python · Pydantic · Typer · JSONSoftware foundations
Typed models, CLI, persisted data and outputs. Project execution uses a .venv environment.
pytest · Ruff · mypyPython checks
Behavioural tests, linting, formatting and type analysis. Pre-commit hooks configure Ruff and mypy; pytest is a separate acceptance check.
Codex · Claude Code · Astro · TypeScriptPublic website
Shared FR/EN components and content, responsive CSS, strict TypeScript and static generation for o2switch. Antigravity was used for the early versions.
ObsidianProject management
Task notes, dependencies, decisions, roadmaps and reviews support handovers between sessions.
LM Studio · Qwen · Snapdragon X EliteLocal experiments
Experiments on my XPS 13 with 64 GB RAM: loaded context, latency and task adaptation.

VS Code is the editor, Cline the agent and DeepSeek the model in this configuration. These are project usages, not a tool ranking. No client data is sent to these tools: the public demonstrations rely on fictional cases and official texts.

03 · My coding workflow

One stage, one report, one decision

The roadmap prepares the work. The instructions define its scope. Files, checks and acceptance observations establish the result and can lead to a revised plan.

  1. Define

    Specify the need, inspect relevant project state and prepare acceptance criteria with an assistant.

  2. Hand over

    Give Cline bounded instructions through a task file or pasted prompt, with the context it needs.

  3. Execute

    Commission edits and specified commands. Stop when the repository contradicts a premise of the task.

  4. Verify

    Compare the report with files and results. Use a browser to check layout and interactions.

  5. Correct

    Prepare a focused follow-up preserving the defect, changes and unresolved points.

  6. Close

    Keep the final report with the prompt. Give separate approval before committing; pushing and deployment remain distinct.

Open configuration details

.clinerules defines stable scope, execution and reporting rules. .clineignore excludes dependencies, caches and large outputs from ordinary context loading; it does not prove absolute isolation across all commands. Design documents live in the vault; the agent receives relevant context in its instructions.

Example verification commands

.\.venv\Scripts\python.exe -m pytest
.\.venv\Scripts\python.exe -m ruff check .
.\.venv\Scripts\python.exe -m mypy src tests

npm run typecheck
npm run build

The report must specify the tests and commands actually run. A configured tool is not evidence of execution. This cycle still relies on my handovers and decisions; it is not a fully connected autonomous agent system.

04 · Three technical cases

From an observed problem to verification

These examples describe historical checks and experiments. They illustrate my workflow, not an audit of the current release or a model benchmark.

XML: checking text attachment

In the Akoma Ntoso pilot, sixteen sub-passages lacked their substantive text while article-level content was preserved.

Intervention and verification

An agent ran checks under my instructions. Diagnostics distinguished XML/index consistency, version comparison and XSD validity. Two outputs could agree while reproducing a shared defect; sub-passage comparisons across versions reported differences.

Documented result — The reported correction restored sixteen sub-passages and made four documents schema-valid, with additional content and attachment guards.

Scope — XSD validation establishes neither complete source fidelity nor a legally correct interpretation. Results remain scoped to the pilot.

LM Studio: reducing Cline context overhead

A 12,891-token request exceeded the 8,192 tokens loaded for Qwen3-Coder-30B-A3B. A recursive scan then produced over 22,000 lines for a task involving only a few files.

Intervention and verification

I narrowed instructions to explicit paths, prohibited general scans and split tasks. A subsequent attempt used a 16,384-token context. I later continued with another provider when latency became excessive.

Documented result — Instructions became focused: read necessary files, edit, give a short report and stop. The execution environment was adapted to the task.

Scope — These observations concern my machine and configuration at that time. They do not demonstrate general provider superiority.

Astro: measuring layout after a correction

A card grid overflowed at 768 px. An initial fix removed overflow but retained the tablet layout at 1440 px.

Intervention and verification

I arranged a component-scoped correction in Antigravity: a desktop breakpoint and an automatic placement reset for the last card. Browser acceptance measured card positions and compared scrollWidth with clientWidth.

@media (min-width: 1100px) {
  .maturity-grid {
    grid-template-columns: repeat(3, minmax(0, 1fr));
  }
  .maturity-card:last-child { grid-column: auto; }
}

Documented result — One column at 390 px, two plus a full-width final card at 768 px, and three at 1440 px, without horizontal overflow in that check. Type checking and the build passed.

Scope — This was a historical local check; publishing and later changes require their own acceptance checks.

05 · Preserving context

Preparing the next session

I keep software, website and documentation separate. Obsidian task notes hold status, the decision register preserves choices, roadmaps explain direction, and dated reports preserve historical results.

Recent rules limit work in progress to one workstream and distinguish verified facts, agent reports and declarations. These are practices to make observable, not a claim of flawless application since the start.

This organisation supports handovers, focused corrections and deferring ideas without an established need. Scutaria’s assessment scenarios remain fictitious demonstrations, distinct from validation in client engagements.

Discuss this experience

I can present a development stage and its checks using excerpts prepared for sharing. For projects connecting business expertise, roadmap design and AI-assisted development.