Claude custom instruction system (CLAUDE.md + extensions)

Context

I needed to standardize Claude's behavior across all my projects and both accounts (Gmail for business strategy, hlahutte@cobra.fr for Odoo/Shopify dev). The immediate trigger: an Odoo production incident the day before (merging the entire preprod3 branch, which contained mixed commits from me and the Irokoo agency, breaking several things) — something to hardcode as a rule to prevent a repeat.

What was built

CLAUDE.md (universal base, ~80 lines): Karpathy's 4 principles (think before responding, simplicity first, surgical edits, goal-driven execution) + general behavior rules (casual tone, no preambles, intellectual honesty, mandatory web search, context management, living document).
CLAUDE-strat.md (business extension, Gmail account): prose format by default, strategy/execution separation, deliverable types, competitive research in the audio market, business honesty (risks before opportunities), a list of what I don't want (buzzword bingo, off-the-shelf frameworks like SWOT/AARRR).
CLAUDE-dev.md (dev extension, cobra.fr account): hard rule at the top on targeted Odoo deployment (never merge all of preprod3, always cherry-pick, mandatory question before any deployment), Odoo conventions (git, Shopify sync, barcode matching) and Shopify, code discipline, a live "Current topics" section.
Clarified deployment workflow: CLAUDE.md in Project Instructions everywhere, extensions uploaded per account.

What was tricky / unexpected

Successive pivots during the session: v1 started with token optimization for autonomous agents (legacy Jarvis/Paperclip), then got reworked once I explained I no longer use the Mac Mini, then again once I clarified the strategy/dev duality. Output quality depends on the clarity of the context provided.
Over-engineering dropped in real time: at several points (weekly briefing agent, personal preprod-hugo branch), Claude steered me toward the simple solution — a direct application of Karpathy principle #2.

Stack

Markdown. Inspiration: Andrej Karpathy's 4 principles for working with LLMs and Boris Cherny's (Anthropic) *Compound Engineering* concept — every mistake becomes a rule, the file improves with use. No code executed, just structured documentation.

What this illustrates

A concrete case of "context engineering" for a non-technical CEO. The value doesn't come from one-off usage but from building a personalized working environment that injects business context, behavior rules and lessons learned into every session. The file becomes an asset that compounds over time — for a CEO who wants to expand their scope (dev, technical SEO, strategy) without replacing their teams.