6 months of Claude Code,
day in, day out.
Seen by a non-dev.

— The numbers

Six months later, here's what it added up to

Before I hand you the lessons, let me put the numbers on the table. Everything is measurable, everything sits in my ~/Projets/ folder or on GitHub. No extrapolation, no projection. Just what came out of the keyboards and is running today.

The raw numbers at six months

• 11 projects shipped · jerwis.fr (this site), a local admin back office with 11 modules, two automated newsletters (AI Playbook + Business Radar), an idea-brainstorm agent (476 items / run, scored across 5 axes), a social-repurposing agent, an HTML sales tool for my sister in sales, two narrative podcast series in the Wondery style, an ATVAI Sanity agent for Eurofiscalis, an article transcreation pipeline in 7 languages.
• ~1,850 Git commits across all my personal repos. I counted by running git log --since="6 months ago" --oneline | wc -l on each repo, then added them up. Before Claude Code, I'd never committed in my life. Today it's become a reflex every session.
• 0 lines of code written by hand. 100% through Claude Code, in plain language. I review, I validate, I commit. But I haven't touched the keyboard for a single line of JS, CSS or Python. Which doesn't mean "zero effort" — the human part just shifted toward the prompt, the plan, the critical review.
• ~12 hours saved per week on editorial production. Before: 3 h to write an article + 1 h to publish it + 1 h to repurpose into social posts × 2 publications/week = 10 h. Today: 45 min of assisted writing + 5 min to publish + 5 min of social repurposing × 2 = 1 h 50. Counted hour by hour over 4 weeks of March 2026.
• ~€340 / month of stack · Claude Code Max $200 plan + Vercel Hobby (free) + Resend (free up to 3,000 emails) + ElevenLabs Creator $22 + R2 Cloudflare (~€0) + a domain name at €1.20/month. Before, I had none of these tools — so it's 100% new investment, but also €0 in agency or freelance fees.
• ~€3,000 / month of freelance dev avoided · at a €600/day rate for the equivalent of 5 days/month (a very conservative estimate for a mid-level dev on HTML + Vercel + Node scripts). Net delta: ~€2,660/month recovered, or ~€16,000 over 6 months.
• 14 articles published on jerwis.fr, including 5 in April 2026 over the first 4 weeks the site was live.
• 10 downloadable freebies (an anonymized CLAUDE.md, a pack of 26 skills, an email-course guide, etc.) — all generated and structured with Claude Code.
• Zero frameworks learned over these 6 months. No React, no Next.js, no Vue. Just vanilla HTML / CSS / JS when it made sense, with Claude explaining what I was touching as I went.

My daily stack

Here are the tools that run every day at my place, at a glance.

Tool	Job	Cost / month	Why I keep it
Claude Code Max	Main brain · code, writing, research, agents	$200	The only one I can't replace · it drives everything else.
superpowers plugin	Forces the brainstorm → plan → execute reflex	Free	Keeps non-trivial projects from going off the rails.
Claude Code loops	Auto-piloted recurring tasks (watch, repurposing)	Included in Max	The thing that runs while I live my life.
dev-browser	Browser driven by Claude · scrape, verify, screenshot	Free	Essential the moment you touch the web.
Claude Code sub-agents	Research and tasks in parallel	Included in Max	3-4 sub-agents = 3-4× faster on research.
VS Code	Reviewing the diffs before commit	Free	I don't code in it · I only review.
Vercel Hobby	Hosting sites and serverless	€0	Auto-deploy on push · zero friction.
Resend	Transactional emails and newsletters	€0 (up to 3,000 emails)	Clean API, Claude Code integration in 2 min.
ElevenLabs Creator	Podcast voices / audio content	$22	Essential for the Wondery podcast pilots.
Cloudflare R2	Hosting the podcast MP3s	~€0	10 GB free + 0 egress.
Domain name	jerwis.fr	€1.20	Obvious.

The concrete delta · before / after

To make the comparison less abstract, here's the same day-to-day with numbers in both setups.

Task	Before Claude Code	Today
Publish a full article	3 h writing + 1 h publishing	45 min + 5 min
LinkedIn / X repurposing	1 h per publication	5 min (dedicated skill)
SEO audit of a page	25 min by hand with Claude.ai	4 min (skill on)
Ship an 18-min podcast pilot	Unthinkable solo	~22 h of active work
Total monthly cost	€0 (nothing shipped)	~€340 tools, €0 agency
Live projects in parallel	0	11

— Two profiles

Why a non-dev has the edge

Before the lessons, a quick detour. There are two paths to Claude Code, and depending on your path, you're not using the same tool at all.

The senior dev

They come from their usual code editor · VS Code, Cursor, Zed (the software where devs write their code). They already have a workflow that works (their way of stringing together the steps of the job). They have fifteen years of ingrained habits. They're looking to automate what they already do, get faster, industrialize. They naturally talk about tests, linting, refactoring. They judge the quality of the code Claude produces.

Claude Code shows up in their life like a fellow engineer. They use it as pair programming (two people on the same keyboard, one codes, the other reviews). They correct its suggestions, they disable what doesn't work for them, they try to optimize every second.

The non-dev entrepreneur

Me, I come from somewhere else. I want to publish a site, send a newsletter, have a dashboard to track my article ideas. I don't start from an editor. I start from a need.

I have no opinion on code quality. I validate the result · is the site live, does the form work, does the email arrive. The rest, I trust Claude with.

Claude Code shows up in my life like an intern who already knows everything, never asks for a lunch break, and never complains.

Three differences that change everything

1. The senior dev knows what's "normal." I don't. So I ask about everything. I don't take a mental shortcut like "ah, it's probably a permissions error." I copy-paste the error exactly as it is to Claude and let the tool explain. The result: I never have a blind spot.
2. The senior dev has a technical ego. They judge the code. They want it done "right." Me, I judge the result. If the site works and the code is ugly, I don't care.
3. The senior dev wants to master it. Me, I want to ship. Two different goals, two ways of talking to Claude.

— 7 concrete lessons

The 7 lessons that stuck with me after 6 months

No theory, no bullet points copy-pasted from some senior dev's blog. Each lesson has a title, the trap I fell into, what works now, and a measured number. You can use every one of them tomorrow morning.

— End-to-end case

36 hours to ship a narrative podcast trilogy

To show you one up close · the wildest project of the 6 months. A narrative podcast series in the Wondery style about the AI war, OpenAI vs Anthropic. Three episodes, writing + voice + editing + going live. Claude Code end to end. Never done before.

Overview · who does what

Step	What I do	What Claude does	Real time
Initial brief	2 paragraphs of framing	Reframes + asks 3 questions	15 min
Source research	Launches 4 sub-agents	Scrapes, sorts, synthesizes	4 h (machine)
Script writing	Cuts + reviews + reframes	Generates + applies the skill	6 h (active)
Voice + audio	Picks voices + ambiences	Drives ElevenLabs + ffmpeg	4 h (machine)
Going live	Validates + commits + pushes	Generates cover + RSS + page	4 h

The brief I gave

"I want three episodes of 18-22 min each. French Wondery style · omniscient narrator, scenes with dialogue, sound ambiences. Subject · the AI war, OpenAI vs Anthropic, seen from 2026. No bullshit, solid sources."

The hour-by-hour breakdown

• H+0 to H+4 · 4 Claude Code sub-agents launched in parallel to gather the sources (OpenAI business timeline, Anthropic's founding, technical analysis of the models, 2024-2026 highlights). Output · 4 synthesized reports, ~80 pages total.
• H+4 to H+10 · generation of the "OpenAI vs Anthropic" pilot script with a "narrative-podcast-fr" skill written for the occasion. Three rounds of review where I made the calls ("cut scene 3, keep this one", "the narrator says 'tu' not 'vous'"). Output · 14,200 characters of clean script.
• H+10 to H+18 · sleep break + a bit of family time.
• H+18 to H+22 · a Python pipeline (written by Claude Code) that slices the script into blocks, sends each block to ElevenLabs v3 with pre-made voices, downloads the MP3s, cleans them, mixes them with ambiences and sound beds. Output · raw 19-min pilot.
• H+22 to H+28 · post-production (still via Claude driving ffmpeg) · cross-fades, EBU R128 -16 LUFS for distribution, background music integration, mastering. The numerical settings saved to memory for reuse on episodes 2 and 3.
• H+28 to H+32 · episodes 2 and 3 generated in series, reusing the same scripts/voices/settings. Automatic cover generation (duotone glitch style, JetBrains Mono).
• H+32 to H+36 · upload to Cloudflare R2 (10 GB free, €0 egress), generation of the RSS feed compliant with Apple Podcasts + Spotify, submission, going live on jerwis.fr/podcast.

The project's numbers

• 3 episodes of 18-22 min each, 60 min of narrative content total
• €0 of agency, ~€14 of ElevenLabs API for the whole thing (voices cost by the character)
• ~36 calendar hours, of which 22 h were active work on my part. The rest, it ran while I slept or lived my life.
• Estimated agency cost · €8,000 to €15,000 for an equivalent deliverable

Everything is documented in the "AI Wars Making-of" at the root of the podcast repo (10 acts, anecdotes, meta-lessons). If you want the full narrative account, it's readable.

— 3 genuinely expensive mistakes

The 3 blunders I made in 6 months

No rosy picture. Here are the three mistakes that really cost me something · a weekend of prod, an emergency key rotation, and one memorable jolt of stress. You'll be able to avoid them.

Mistake 1 · Pushing to prod without reviewing the diff

What happened · Saturday, April 12, I was working on the home page in dark mode. I told Claude "optimize the colors." It replaced background: var(--ink) with background: #0A0A0A across 6 sections. Except --ink in dark mode = cream, so some blocks meant to be ALWAYS black became cream on cream = unreadable. I pushed without reviewing and went to lunch.

What it cost me · the jerwis.fr home page was broken for 4 hours in dark mode. I got two DMs "there's a weird bug on your site." I spent my Sunday afternoon redoing the whole thing.

My rule since · an absolute obligation to test locally in light + dark before any git push. And the "semantic variables vs fixed colors" rule is carved into the project's CLAUDE.md.

Mistake 2 · Trusting Claude with a secret

What happened · on the "Importateur de Buzz" project (NANAKIA, for someone close to me), I had a Cloudinary API key lying around in a config file. I asked Claude to "clean up the project." It moved the key into .env.local, great. Except before the move, the config file with the key in plain text had been committed twice. The key lived 5 days in the public Git history.

What it cost me · an emergency key rotation, and a security audit of the whole repo. Plus a session of about 3 h to understand how to remove a secret from Git history cleanly (answer · git filter-repo).

My rule since · every secret goes into .env.local BEFORE the first commit. A pre-commit check looks for API-key patterns (Resend, Stripe, OpenAI, etc.) in any added file. And I review every diff before push, explicitly looking for sk_, re_, pk_.

Mistake 3 · Asking Claude to handle a prod migration without a backup

What happened · on the Eurofiscalis CRM (another project), I wanted to reorganize the bookings table. I told Claude "can you migrate the table so it has the new structure?" It did. Without warning me I needed a backup. The migration failed on 3 rows (badly formatted data). No way to roll back.

What it cost me · an entire weekend rebuilding the 3 rows by hand from Brevo emails. And a monumental jolt of stress.

My rule since · any destructive action (migration, delete, mass update) must be preceded by an explicit backup + a rollback command documented in the plan. This rule is in my global CLAUDE.md, and Claude reminds me of it when I skip it.

— My routine

What a typical day looks like

People often ask me how it actually shakes out when you work with Claude Code every day. Here's my current routine, the way I settled into it by late April 2026. Three sessions a day, never more than 90 min each, and the rest of the machine running in the background while I live.

Morning · 6:30 — 9:00

• 6:30 — 7:00 · coffee + reading 2 newsletters (Hormozi + one depending on my mood), skimming HackerNews and Reddit r/ClaudeAI. No opening of Claude Code. Just raw material.
• 7:00 — 8:00 · I open my local back office (localhost:3001), I click "Find me some ideas." The brainstorm scans 470+ items, scores them, sorts them into 6 clusters, and hands me 15 ideas. I keep 2-3.
• 8:00 — 9:00 · I launch a Claude Code session on the idea of the day. I give it the brief, I tell it "propose a plan." We validate, we code, I commit every step.

Afternoon · 14:00 — 17:00

• 14:00 — 15:00 · a deep-work Claude Code session. Often a heavier project · Eurofiscalis, podcast, reworking an admin module. This is where I split into sub-agents if I have to do several things in parallel.
• 15:00 — 16:00 · away from the screen, I walk, I let Claude run its long tasks (7-language transcreation, audio generation, multi-source scraping). I come back when the terminal is done.
• 16:00 — 17:00 · review + commit + push. I test locally in light + dark, I review the diff, I publish. If a newsletter needs to go out, this is where I click "send" in my admin.

Evening · 21:00 — 22:30

• 21:00 — 22:00 · a "tidy up" session. I update the CLAUDE.md files of the projects I touched, I fill in the CHANGELOG.md files, I tag the backlog ideas I want to handle next week.
• 22:00 — 22:30 · planning the next day. I list 3 priorities max, I leave Claude Code closed for the night (for real · never any overnight auto-deploy).

There's nothing optimized about this routine · it settled in on its own over the 6 months. It lets me produce 2-3 serious deliverables a week without burning out.

— What's still hard

What Claude Code still gets wrong

I'd come across as a snake-oil salesman if I stopped at the nice stories. Here's what still doesn't work, after 6 months.

Hallucinations on library versions

Claude regularly proposes syntaxes that disappeared a year ago, or npm packages that don't exist. On recent Next.js (App Router, Server Actions), you have to systematically cross-check against the up-to-date docs. The context7 plugin helps enormously · it pulls the official docs at the moment of answering, rather than relying on its training memory. Without it, I reproduce the bug.

Sessions that run too long

Past ~90 min, Claude starts losing the thread. It forgets a decision made at the start, or it "invents" a function it thinks it wrote. The fix: restart a session with a 5-line recap prompt ("here's where we are, here's what we need to do, here are the constraints"), and close the previous one.

Time and cost estimates

Claude is very bad at estimating how long a task will take. It says "5 minutes" and it takes 45. It says "this'll cost €2 in API" and you end up at €18. I no longer ask it to estimate · I measure it myself afterward, and note it in the CHANGELOG.

Very recent topics (post-November 2025)

Claude models have a knowledge cutoff. On anything that happened after their training date, they're off or make things up. For real, current research, I go through a sub-agent that has web access (dev-browser, or a web MCP). Never Claude on its own.

— Quick FAQ

The questions I get a lot

"Do you really need the $200 Max plan?"

For how I use it (daily sessions, sub-agents, projects in parallel), yes. I'd started on the $20 Pro plan and I'd hit the token ceiling in 2 days. Above 1 serious session a day, the Max plan becomes the right ratio. Below that, Pro is plenty.

"Cursor, Continue, Zed AI, Claude Code · which one?"

For a non-dev, Claude Code straight in the terminal is simpler. No IDE to learn, no plug-in to configure. For a senior dev who already lives in VS Code or Zed, the opposite may be true. The real difference isn't the tool, it's the discipline (CLAUDE.md, plan, CHANGELOG) — and that one works everywhere.

"Aren't you going to end up learning to code anyway?"

Probably a little. I understand better than before what a variable is, an import, an async function. But I won't become a dev. That's not my goal. My goal is to ship tools that serve me, without depending on a freelancer for every little tweak.

"And what about security in all this?"

It's my biggest fear. I lean on Claude's security-review sub-agent after every sensitive feature (API, payment, auth), I systematically check my .env files, and I rotate keys at the slightest alert. On truly critical projects (client SaaS, third-party data), I'd still have a human dev review it. But for my personal tools, that level is enough for me.

"How long before you're self-sufficient?"

For a non-dev who gets serious about it, 3-4 weeks to ship your first useful project (a simple site, an automated script). 3 months to build the reflex of breaking work into plans, writing a CLAUDE.md, asking for a CHANGELOG. 6 months to truly feel that you're driving the tool and not the other way around.

"How many hours a week to use Claude Code seriously?"

Count on 8 to 12 hours a week for the first 3 months if you start from zero · one to two sessions of 90 minutes a day, plus reading docs and digesting your mistakes. After that the pace drops to 6-10 hours a week, but the output per hour climbs sharply (you know what to prompt, you verify faster, you cut the sessions that go off the rails).

"What's the minimal non-dev stack around Claude Code?"

Five building blocks are enough · Claude Code Max ($200/month), a free VS Code editor to review the diffs, Vercel (free) to host, Resend (free up to 3,000 emails) for transactional email, and a GitHub account. Total · $200/month to go from zero to several projects in production. Add ElevenLabs ($22) if you do voice, Cloudflare R2 (free) if you host heavy files.

"Do you need to know git before starting?"

No. Before Claude Code I'd never committed in my life. I ask Claude to explain every command the first time (commit, branch, merge, revert), and today it's become a reflex every session. Count on 2 weeks to be comfortable with the 6-7 common commands. The trap · NEVER do git push --force on main without understanding what it overwrites.

"What's the worst mistake to avoid when starting?"

Running a database migration without a backup. It happened to me once (told above), I lost a weekend rebuilding my data by hand. Golden rule · before any DELETE, UPDATE or schema migration, ask Claude to first make a SQL dump and confirm that the backup file is readable. It's 30 seconds, it spares you the cold sweats.

"Can Claude Code really replace a freelance dev?"

For internal tools, personal sites, automation scripts and MVPs · yes, and I save around €3,000 a month in equivalent freelance dev day-rate (estimated at €600/day × 5 days/month). For billed client SaaS or critical systems (payment, medical data, contractual SLA), no · legal liability and long-term technical debt still call for a human who signs off.

"Can you show a concrete end-to-end case?"

Yes · the full making-of of the Eurofiscalis booking (an appointment-booking widget embedded in 4 WordPress sites) is documented hour by hour in this article. You'll see the initial brief, the traps I hit, the real cost and the trade-offs I owned. It's probably the most representative of what Claude Code changes for a non-dev on a real client project.

— The essentials for a non-dev

Three rules worth all the tips

A senior dev will talk to you about optimization, test coverage, clean architecture. For an entrepreneur who wants to ship, three things really matter.

1. Does it work for you, today. Not "will it scale to 10,000 users." You're not there yet.
2. Do you understand what you're putting into production. Not "every line," but "what the script does overall." If Claude hands you something you don't dare review, you stop.
3. Can you fix it if it breaks. Tomorrow, a script stops running. You need to be able to open Claude Code, copy the error, get a diagnosis. If you depend on a complexity you never understood, you're stuck.

These three rules are worth all the senior-dev tips in the world. They don't stop you from reading the tips, of course. They just tell you · apply them at your own pace, and not because someone else does.

Going further

If you want to keep the momentum going:

1. Getting started with Claude Code — the starting point if you've never opened a terminal
2. Claude Code loops explained — the decision loop that makes the tool powerful
3. Building your first Gmail agent — a concrete project to get hands-on
4. Superpowers · the plugin that forces Claude to think — the plugin I use to structure every non-trivial session

If you want to follow my concrete field reports as the months go by, my Friday watch sends them straight to you. No ads, unsubscribe in one click.

If you have a question or feedback on this article, just reply to the newsletter. It lands in my inbox and I read everything, I can be wrong on some points and your feedback helps me set things straight.