I work with Claude Code, Codex and LLMs every day: agents draft, I design, review and test. In an expert's hands, AI brings the best of both: 10 years of expertise, plus the speed and productivity of agents.
Jean-Baptiste TheryFounder of jCodeAfter 10 years in the trade, I believe that today, refusing AI would be unprofessional, like a scribe turning away from Gutenberg's press. Using it without rules would be just as unprofessional. I use it every day, responsibly, within each client's rules.
Agents take the repetitive work; I keep the decisions and spend the time saved on quality.
Agents map a codebase and what a change touches before I write a line, on a new project or on an application I take over.
Features, migrations and documentation start from an agent's draft that I correct and complete.
Agents write the tests a tight schedule usually skips, and I check what each one actually asserts.
An agent reviews every change before my own review, to catch what a tired eye misses.
The time saved goes into architecture, security and edge cases, where experience counts most.
The same method serves to design an application from scratch, take one over, evolve it, fix it and maintain it.
I use coding assistants only with enterprise privacy settings, or with the models your company already runs.
Under Anthropic's commercial terms or through your own cloud (Bedrock, Google Cloud, Azure), code and prompts do not train models; zero data retention can be enabled for eligible Enterprise organisations.
On ChatGPT Business, Enterprise or the API, OpenAI does not use inputs or outputs to train its models by default.
With Privacy Mode, Cursor holds zero data retention agreements with its model providers.
Your company's accounts, an open-source agent such as OpenCode with the LLMs you choose, or a self-hosted model: I work with what you already run.
Tool terms change: I check them for each engagement, and your contract prevails. Terms checked in October 2026.
I start from the need, not from the tool, whether for your product or for your company's own processes. Three answers, from the simplest to the most capable.
When the rules are clear: a webhook, a scheduled job or an n8n workflow, or custom code when the volume or the logic calls for it. Cheaper, faster and easier to audit than a model.
When rules fall short, on free text for instance. Prompting first, then RAG on your documents; fine-tuning only if both fall short, always evaluated on real examples.
When a task takes several steps and some judgement: tools with narrow permissions, guardrails and a human approval before anything irreversible.
The right answer often mixes them: an automation that calls a model at one step, or an agent that leaves everything predictable to plain code.
Six steps, the same on every project, so speed never costs quality.
The repository's AGENTS.md, the ticket and a Ragmir search give the agent the rules and the relevant code.
The agent proposes a plan or a spec that I approve, split into tasks small enough to review in one sitting.
The agent drafts the change and its tests within the limits I set; where it fits, a failing test comes first.
Type checks, tests, linting and, for an interface, a check in a real browser.
I read every line, ask an agent for a second review and fix what is wrong.
Continuous integration runs, and nothing is merged or deployed without my approval.
Client work stays confidential. My own products show the method in the open.
A personalised workout generator with an adaptive AI coach, built as its sole technical owner from June 2025 to September 2026, with coding agents under the rules on this page.
Open-source agentic RAG I created in 2026: coding agents search code and documents and cite their sources. It runs in my own setup every day.
The same harness on every project, maintained like a product.
Claude Code first, Codex on the same instructions, GitHub Copilot in VS Code, and one AGENTS.md file per repository.
My own versioned skills in the open Agent Skills format, shared by Claude Code and Codex: browser proof, code review, SEO audits, upgrades and delivery.
Ragmir for cited context, a code graph for impact analysis, current library documentation and a private gateway where every write waits for my approval.
Subagents in parallel and scripted workflows that always end with a verification pass.
Hooks that keep the agent's context focused, and a memory that carries decisions from one session to the next.
Tasks with dependencies in Linear, Notion or GitHub Projects, human checkpoints, and no task closed without proof.
Agentic engineering, not vibe coding. These are the skills I practise, and what to expect from anyone who uses LLMs on your code.
The right files, constraints and examples, and nothing more.
Plans, subagents and workflows: a task split across agents, and every result checked.
Tools with narrow inputs, clear errors and the least privilege they need.
Chunking, embeddings and BM25 search, as built into Ragmir.
Reference cases that measure a prompt, a model or a search before and after each change.
Tests, type checks, CI and browser evidence: what an agent claims is checked, not trusted.
Against prompt injection, content read is data, never instructions, and anything sent outside waits for human approval. Secrets stay out of prompts.
The right model and reasoning effort for each task, an eye on token cost, and an agent's solution rejected when it is wrong.
What you can hold me to on every engagement.
Tell me what you want to build or improve, your budget and your target date.
A developer who has coding agents such as Claude Code or Codex write and run part of the code, while keeping the design, the review and the responsibility for what ships. It is engineering with agents, not vibe coding.
No. Your company's rules come first: on your code, I use only the assistants, accounts and data it allows.
No. I work on client code through business or API accounts that do not use it for training. A passage sent to an external LLM still leaves your infrastructure and falls under that provider's terms.
You do, as for the rest of the code. The contract provides for the assignment of rights to the code I deliver, including code drafted by an agent.
Not without me. Agents do not commit, push, merge, deploy or delete anything without my explicit approval, and they never sign in to your services: I type every password and one-time code myself.
Usually, never at the expense of review. Studies disagree on the size of the gain, and AI mostly amplifies existing practice, so tests, CI and review come first. I quote the work itself, as a fixed price or at a day rate, rather than promise a percentage.
Rarely as a first step. A well-built prompt, then a search across your documents (RAG), covers most needs; fine-tuning pays off when evaluations show that both fall short, and it needs clean example data.
Describe your project on Malt
Ask a question on LinkedIn