Skip to Content
Meubles Morinière

The project14 min read

My contribution and AI

AI as leverage, engineering as the standard. How I design and build with Claude Code, what I frame, decide and verify, and the records that show it.

AI as leverage. Engineering as the standard.

I build artificial intelligence into my development process to explore, design and build more effectively. From the first idea to the implementation, I use Claude Code to go further, while keeping control of the technical choices, the consistency of the product and the quality of what I deliver.

WhoWhat is theirs
The store’s managerthe business rules that commit the store, which he approves in the ERP under “Décisions à valider” (decisions to approve)
Methe needs and the business frame, the trade-offs, the rules and guardrails that frame the AI, the consistency of the product, the tests, checking the result on screen and shipping to production
Claude Codeexploration, proposals and implementation (the code, the tests, the migrations, this documentation), within the frame I set and under the repository’s guardrails

What follows rests on records gathered on November 13, 2026: the git history, my messages in Claude Code sessions, the old decision log and the instructions that frame the sessions. The records and the survey are detailed at the end of the page. The same split, one architecture decision at a time, is in What is mine in these decisions (in French).

Design

What comes from me:

  • The business framing, which I give the sessions again and again and against which I have them judge their proposals: “audit all the business decisions, check that they are consistent with meublesmorinière SME furniture consumers businesses family business a single warehouse VAT 10% and 20%. 2026-2027” (October 22).
  • The needs I notice in use: “when I try to delete a piece of furniture and there are still units in stock I don’t want the user to be blocked” and “erp product list the drag handle is confusing you can’t tell what it’s for, what it reorders and why, to what end” (October 22).
  • The link with the store. Questions for the manager go through me, and feedback from the people who use the ERP comes back to me: “check that nothing was forgotten don’t forget that the manager knows nothing about computers too” (October 26).
  • The first data schema, according to the log: “schema provided by Adrien, applied word for word” (August 24, reported record).

The AI helps me explore: it proposes screens and architectures, runs audits, evolves the data model and suggests most of the names and shapes. I judge these proposals against the business frame, and I keep the ones that serve the store.

Trade-offs

Here is what I decided:

DateDecisionRecord
August 23business customers can buy online, instead of the “strict B2C” considered the same daylog
August 24refunds are left out of V1, against the two external reviews from the same daylog
August 31the production migration is run by hand, before the mergelog
October 12the AI runs no tests, because parallel sessions were saturating the machineinstructions
October 25“Migrer la prod” (migrate production) runs without a copy of the database, and rollback goes through Neon’s point-in-time restoreapps/erp/scripts/pre-deploy-migrate.mts
October 26Fisher no longer has any scope limit, and runs without zero data retention, for lack of a Pro planinstructions, and the comment below
November 9delete before adding, and copy nothing the code already knowsinstructions

A trade-off that has a cost is written down where it applies. The one from October 26 accepts that exchanges with Fisher are kept for thirty days by Anthropic, and the storefront’s Privacy page says so.

isAiAvailableapps/erp/shared/server/ai/availability.ts · lines 5–18
/**
 * Fisher se montre dès que la passerelle peut s'ouvrir. Décision d'Adrien (26/10/2026) : sans plan
 * Pro, la non-conservation (ZDR) n'est pas offerte par la passerelle ; Fisher passe quand même, sans
 * entraînement chez Anthropic, qui efface les échanges sous 30 jours. `AI_ZERO_DATA_RETENTION` la
 * redemande le jour où l'offre la permet (`aiModelOptions`).
 */
export function isAiAvailable(): boolean {
	if (env.AI_DISABLED !== undefined) return false;
	return (
		env.AI_GATEWAY_API_KEY !== undefined ||
		env.VERCEL_OIDC_TOKEN !== undefined ||
		env.VERCEL_ENV !== undefined
	);
}

For everyday technical calls, I give the criterion rather than the solution: “and go ahead fix everything make the best choices without over-engineering to settle the decisions and suited to meublesmoriniere as simply as possible” (November 13). The AI compares and applies within that frame. Whatever commits the product is written down, dated, where it applies (the instructions, the ADRs), so that the reason for a choice can be reviewed like the code.

The business rules that commit the store belong to the manager. They are put to him in the ERP, under “Décisions à valider” (decisions to approve), and each answer keeps the verdict, who gave it and the rule in force at that moment (DecisionAnswer, packages/db/prisma/schema.prisma). On November 9, the list had 302 decisions (reports/ERP Morinière registre des décisions.md), and I asked to keep only the real ones: “Analyse each decision asked of the manager. Remove the unnecessary choices, group those that overlap and keep only the genuine business decisions that require his input.” It went down to 44 on November 11 (6327df8f1c), then to 8 on the 13th (7703182e16). His answers are in the production database, not here.

Development

Implementation goes through Claude Code: the code, the tests, the migrations, the commit messages and the documentation, this page included.

What comes from me is the frame that makes this speed safe:

  • The instructions (AGENTS.md), on how ERP pages load, toasts, Fisher, maintainability and the production migration.
  • The guardrails that protect the machine, Claude Code hooks born from measured overloads, each carrying the decision that gave rise to it: no tests run by the AI, at most three dev servers (seven took up 27 GB out of 24), two tsc at a time, every command at background priority.
  • Two commands that frame an audit, /audit-backend and /audit-frontend: a scope, findings on the real database or the real rendering, then the fix. They come from a specific request: “I don’t want to describe the goal I want to define a feature, a domain, a scope with this command and fix/apply/improve” (October 22).
  • Orchestrating the sessions, each in its own worktree, and integrating their work into main.

What proves it: the instructions, the hooks, their wiring (.claude/settings.json) and the two commands have been in the repository since November 13. Only the ERP’s saved session, my permissions and the worktrees remain on my machine. From October 14 to November 13, git ignored them, so that month has no history. Here is a guardrail as a session meets it:

MESSAGE.claude/hooks/no-tests.mjs · lines 5–11
const MESSAGE = `🚫 Tests interdits à l'IA sur ce poste (décision d'Adrien, 12/10/2026).
Adrien lance les tests lui-même, à la main, quand c'est nécessaire : des runs
IA en parallèle ont surchargé son Mac (charge 84 sur 10 cœurs, swap saturé).
Ne lance NI vitest, NI playwright, NI \`pnpm test*\` / \`pnpm e2e*\` / \`pnpm validate\`,
même sur un seul fichier. Écris le code (et les tests demandés) sans les
exécuter, puis donne dans ta réponse la commande exacte qu'Adrien jouera, ciblée
sur les fichiers touchés.`;

Validation

What comes from me:

  • The tests. Since October 12, I run them myself, by hand, on the files touched. The AI writes the tests and gives the command.
  • The real rendering, in use: “footer logo is huge just add the logo and not the name” (November 2), “it’s too long” (November 12), “site menu sheet I don’t want a fixed close X it’s ugly” (November 13).
  • What the AI asserts, which I question (next section).

On top of that comes the tooling I had set up: the pre-push gate, which checks the head commit of every push to main in a detached worktree (Quality), and the tests that read the code.

And some actions cannot be proven. The gate’s attestation log records four Postgres suites declared as run in November, and fifteen production migration receipts are on my machine. But all of them carry my system user, which my sessions also run under, so they cannot tell my action from an agent’s.

Understanding and questioning

Here is what the records show, episode by episode, with my messages translated from the French I typed them in.

A wrong idea, corrected by the AI, then checked

On November 10, I write: “I don’t want the erp to have point-of-sale software because the law requires I think an external tool?”. The AI replies, citing the tax authority’s text, that the law does not require an external tool but a secure till and a proof of compliance, and refers one point to the accountant. I then ask it: “check the erp against the four conditions”. The next day, a commit seals sales from their first payment and makes closings unalterable (6327df8f1), and the link is inferred from the order of events. On November 13, I am still writing: “what to do I don’t understand point-of-sale software?”. The choice between a certificate, a proof of compliance and a separate certified till remains to be made with the accountant (ADR 0008). That choice is mine, not the tool’s.

A question that changed the product

On October 26, I ask: “why is ask fisher on every page if it’s only for searching the manual?”, then “why not centralize (drop a file etc)”. The decision dated the same day makes Fisher a single chat, on every page, that reads ERP data under the person’s rights and receives the documents they drop (Fisher).

Demanding the source

On November 12: “for the loading are you sure? inspect the next js documentation”. The AI rereads the Next.js documentation installed in the repository, and corrects one point of its previous answer.

A question that finds no error

On November 13: “featured products why isn’t product card reused?”. The AI justifies the difference: the “Aussi à la une” (also featured) row has neither the layout nor the information of ProductCard. Questioning did not mean being right.

Holding back what the tool produces

On November 3, I write “remove the over-engineering frontend public site”, and on November 9 “I want to avoid a maintenance burden that is too heavy for the ERP”. That second message comes just before the inventory and the slimming plan dated the same day (git show b3f1e845d2:docs/allegement-erp.md), and then the rule “delete before adding”.

Mastery is checked in an interview

A repository shows what was built, and this volume is real:

lines of TypeScript excluding tests: routes, modules and shared code of both applications, package sources
408,398
apps/{erp,site}/{app,modules,shared}, packages/*/src
lines of tests: unit, on Postgres, source-reading, end-to-end
381,832
*.test.ts(x) and *.spec.ts under apps/ and packages/
Measured in the repository at each build of this documentation.

Mastery, on the other hand, is checked in an interview, without AI. What I have been building in a team since 2022, outside this repository, is described in the case study (Outside this project). Here is what I suggest asking me:

  • Follow an order from the cart to the invoice, saying at each step which application writes what, and under which rights (The path of an order).
  • Explain why the storefront cannot mark an order as paid, and what stops it in the database (Security).
  • Read a CHECK constraint and its refusal test, and say what it protects (Quality).
  • Change a rule in packages/domain and predict what breaks.
  • Defend an ADR, and say what would make me change my mind (Decisions).

What I take from it

A decision that matters has to be written down. When I give a criterion rather than a solution (“best choice”), the reasoning behind the choice stays in the session’s report. I drew one rule from it: a decision that matters is written down, dated, where it applies, and the ADRs separate the confirmed reason from the assumption.

AI speeds everything up, including the growth of the code: thirty-two audit reports in eight days at the end of August, and 605,000 lines of TypeScript in the ERP as of November 9, including 302,000 lines of tests. Volume does not prove quality, and every addition is paid for at every change. Hence the November 9 rule and what has been removed since August: the decision log, almost all of the decisions submitted to the manager.

Decisions made quickly got reversed. Stock was taken out of the model and came back the same day, after a third audit (August 24). Refunds, cut from V1 on August 24, are today the Refund model. On August 29, I refused a history of order steps (“I don’t want an order status event”). On November 5, when nothing recorded the date an order went into fabrication any more, the OrderStep log arrived (2f4e60058c). The quarantine on new dependency versions, removed by my decision on August 29, is now one day (minimumReleaseAge, pnpm-workspace.yaml). Cutting fast costs almost nothing in the code, but coming back costs a migration. What I take from it: ask about the need before cutting, as before adding.

Orchestrating AI is a craft in itself. When my sessions saturated my machine (a load of 84 on 10 cores, swap full), I turned it into rules and hooks. They have a price: the AI does not run the tests, so a change can land without it having run them. The push gate and my manual runs make up for it, but the browser tests that walk through whole user flows remain manual.

The most personal part of my work stayed outside the repository for a month. From October 14 to November 13, git ignored my instructions and my hooks, even though several pages of this documentation cited AGENTS.md as a source. They have been versioned since then, because what frames the AI has to be reviewed like the code it writes.

Three kinds of records

RecordWhat it isWhat it is worth
Directwhat I typed myself, kept as is: my messages in Claude Code sessionsshows what I asked for, refused or questioned
Reportedwritten by the AI at the time: “Adrien’s decision” in the log, an instruction, a comment in the codedates a decision and attributes it to me, but the wording is the AI’s, and the exchange that produced it is no longer kept before October 22
Inferreddrawn after the fact from the code or the order of commits, this page includedan assumption, presented as one

The survey

This survey dated November 13, 2026 has not been checked again since.

SourceWhat it shows
main, 3,076 commits from August 23 to November 13, 20261,188 have Claude as their author: cloud sessions, from October 30 to November 9. Of my 1,882 commits, 1,456 carry Co-Authored-By: Claude ….
My 426 commits without that trailer184 merges, 39 wip: snapshots from parallel sessions and 203 others (70 in August, 131 in September, 2 in November).
Synclune, a sister projecta copy of this repository that became a jewellery store on September 23, and came back into main in a single commit on October 14 (72332fcd97, 3,770 files). Its 24 commits, outside main, all carry Claude’s trailer.
My Claude Code sessions, from October 22 to November 13462 sessions on my machine, 818 typed messages. Older sessions are no longer kept, and cloud sessions are not included.
The old decision log, from August 23 to September 1390 dated entries: 40 attribute a decision, a request, a ruling or a refusal to me, and 196 rely on an audit. Written by the sessions, removed on September 1.
The session instructions (AGENTS.md, .claude/)seven dated decisions and requests, still in force, versioned since November 13 along with the hooks that apply them. git had ignored them since October 14, so that month’s decisions have no history, and their date is the one written in them.

What can be replayed from the repository:

git shortlog -sn HEAD
git log --author=Adrien490 --format='%(trailers:key=Co-Authored-By,valueonly,separator=%x2C)' | grep -c Claude
git show '2fd35cdb09^:docs/decisions.md' | grep -E '^- 2026-' | grep -c Adrien
git log --format='%h %ad %s' --date=short -- AGENTS.md .claude/

The sessions themselves are on my machine (~/.claude/projects) and cannot be replayed from the repository.

Method

  • This page was written with Claude Code, from this survey of November 13, 2026. The French version quotes my messages exactly as I typed them, typos included, and this English version translates them.
  • The messages counted are the ones I typed in the sessions on my machine, excluding subagent messages, command templates and tool output. A go-ahead is a message of a few words that approves, relaunches or has a commit made.
  • The episodes in Understanding and questioning were chosen by reading, not counted. They show what is possible, and say nothing about a frequency.
  • This survey does not judge the quality of the software. The other pages in the Skills and evidence section do, each with its sources.

On this page