Anthropic’s guide to Claude Code says instructions written in a prompt, a CLAUDE.md file or a skill are requests the model can decide not to follow, while hooks, deny rules and managed settings are enforced every time (Claude Code best practices, read November 8).
Addy Osmani, whose X bio reads “Claude Code @anthropicai”, wrote in an X article on November 8: “prompts, CLAUDE.md, and skills are requests that the model can consider and choose to act on; hooks, deny rules, and managed settings are requirements that are always enforced”.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a harness around models shapes what they deliver. It teaches the general idea only and does not discuss Claude Code’s guide.
The key insight: Anthropic’s Claude Code best-practices guide, read on November 8, says the prompt, CLAUDE.md and skills are requests the model may decline, while hooks, deny rules and managed settings are applied by Claude Code every time.
The guide separates two kinds of instruction. In the first, Claude decides whether a line applies. In the second, Claude Code, the program around the model, carries out the rule itself. Of the second kind, the guide says: “Claude Code applies this every time, no matter what the model decided to do”.
Its advice for anything that must never be skipped is direct: “If something has to happen every time, make it a hook, a deny rule, or a managed setting.” The same page says CLAUDE.md is for guidelines, and that Claude is allowed to decide a guideline does not apply.
By our count, the guide names three places for requests, which are the prompt, CLAUDE.md and skills, and in its definition three enforced controls, which are hooks, deny rules and managed settings. Its comparison table also lists the sandbox in strict mode on the enforced side. The guide adds that hooks which ask another model to decide are not deterministic either, and that real enforcement needs a command or http hook.
How Anthropic’s Claude Code guide sorts instructions: the prompt, CLAUDE.md and skills are requests the model weighs; hooks, deny rules and managed settings are enforced by Claude Code. Data: Anthropic, Claude Code best practices, read 8 Nov 2026. Grouping and counts are ours.
What Was Posted on X
Holy: An Anthropic engineer building Claude Code says writing “NEVER” in CLAUDE.md doesn’t make it enforceable: prompts, skills and project instructions are requests; hard boundaries need hooks, deny rules or managed settings.
Old habits now backfire too. ALL CAPS, “think harder” and “don’t be lazy” can push current models into overwork. If two corrections fail, clear the session and rewrite the prompt instead of making it re-read a third failed attempt.
Scaling agents creates an “orchestration tax”: past the first few, each extra agent can demand more attention than the last.
And green tests plus plausible code aren’t enough. Approving code you don’t understand isn’t code review.
An X post summarised Addy Osmani’s article, saying writing “NEVER” in CLAUDE.md does not make it enforceable and that hard boundaries need hooks, deny rules or managed settings. This piece checks those points against Anthropic’s own Claude Code guide.
Seven Habits the Guide Says to Drop
The guide’s section titled What you can stop doing lists seven older pieces of advice, by our count. It says: “Every one of these was good advice when it was written”, and that most were workarounds the models have since outgrown.
On writing rules in capital letters, the guide says “piling on emphasis mostly makes Claude over-apply the rule”. On telling the model to keep going or not to be lazy, it says lines like that “are more likely to cause over-work than fix a stall”. It says effort level, not phrases such as think harder, is the control for how hard Claude thinks.
Osmani’s article puts the same point this way: “With current models, these tend to lead to over-enforced rules and overwork”.
Two Corrections, Then Start Over
Both texts give the same rule for a session that goes wrong. Osmani writes: “If you’ve already made two failed attempts to correct the agent, clear the session and write a better prompt”. The guide says a third correction mostly adds to a conversation that already contains two failed attempts for Claude to keep re-reading.
Osmani’s article compares engineering managers’ practices “with what we’ve learned from using and creating Claude Code”. He gives two names to problems he describes. On running many agents at once: “It’s what I call the orchestration tax”. On approving large diffs because tests pass: “This is not code review, this is code ratification”.
The Business Lever
For teams rolling out coding agents, the guide’s own wording sets where a rule has to live. It says that if something has to happen every time, it belongs in a hook, a deny rule or a managed setting, and that CLAUDE.md is for guidelines Claude may decide do not apply.
As we read the guide, its split moves a policy question into configuration: a rule that must hold every time is written as a hook, a deny rule or a managed setting rather than as a sentence in CLAUDE.md.
As we read Osmani’s article, the same split runs through his advice to let the agent decide by default and to enforce only the few things that must always or never happen.
As we read both texts, they describe how the controls are meant to work; neither measures how often instructions in CLAUDE.md are ignored in practice.
Addy Osmani, X article, 8 Nov 2026
“With agents, over-process is the new micromanagement.”
Three Implications
FOR TEAMS SETTING RULES The guide says anything that has to happen every time belongs in a hook, a deny rule or a managed setting, and that CLAUDE.md is for guidelines.
FOR LONG-TIME USERS The guide lists seven pieces of older advice it says users can stop following, including ALL CAPS instructions and telling Claude not to be lazy.
FOR SECURITY REVIEWS The guide says a WebFetch deny rule for a domain does nothing about curl to that same domain unless the sandbox is on.
The essay describes “an operating system that wraps your AI”, “a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”
As we read it, the guide’s hooks, deny rules and managed settings are the enforced part of that harness: they are applied by the software around the model rather than left to the model. The essay does not discuss Claude Code’s guide.
What Is Not Established
The guide lists limits on enforcement too. It says a WebFetch deny rule for a domain does not stop the same domain being reached with curl unless the sandbox is on, and it says “Auto mode is a classifier model judging each action rather than a hard rule”. This piece does not test any of these controls, and it does not show how often instructions in CLAUDE.md are ignored in practice.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Anthropic’s Claude Code guide says prompts, CLAUDE.md and skills are requests the model may decline, and that hooks, deny rules and managed settings are enforced every time. By our count it names three of each in that definition, and seven pieces of older advice it says users can stop following.
A note on sourcing. Quotations are from Anthropic’s Claude Code best-practices documentation, read on November 8, 2026, and from Addy Osmani’s X article of the same day. Documentation changes over time. The grouping and counts are ours. We did not test the controls described.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
