Your AI Instructions Are Too Long. Anthropic Just Proved It.

The company that builds one of the best coding AIs in the world just spent three days publicly telling its own engineers to delete most of their instructions.

It is 6 p.m. On a Thursday. Priya runs a six-person marketing studio out of a converted garage in Austin. Before she lets Claude touch a client email, she pastes in a 900-word house-style document: tone rules, banned words, formatting quirks, three past mistakes she never wants repeated, and a paragraph reminding it not to use the word "delve." She has rewritten that document four times this year. It keeps getting longer. It keeps getting less obeyed.

She is not doing anything wrong. She is doing what almost every guide on the internet told her to do: give the AI more rules, more examples, more guardrails. Anthropic just told its own team that this advice is out of date for the newest models, and published the receipts.

You are not managing an assistant. You are managing a rulebook nobody rereads

On July 24, 2026, Claude Code engineer Thariq Shihipar posted a thread that has since pulled in more than 4 million views and over 31,000 bookmarks, an eye-watering number for a technical post about system prompts. His claim: Anthropic deleted more than 80% of Claude Code's system prompt for its newest models, Claude Opus 5 and Claude Fable 5, and measured no drop in coding performance.

The same day, Anthropic published the full reasoning on its blog, in a post titled "The new rules of context engineering for Claude 5 generation models." The admission at the center of it is blunt: "Overall, we found that we were overconstraining Claude Code, both through our system prompt and in our CLAUDE.md files and skills." Priya's 900-word document is exactly the pattern Anthropic is describing, just built by a small-business owner instead of an AI lab.

What is Claude context engineering?

Context engineering is Anthropic's name for everything the model sees besides your actual message: the system prompt, any "skills" it has loaded, a CLAUDE.md or custom-instructions file, memory from past sessions, and whatever documents you attach. Anthropic's blog puts it plainly: "When you send a message to Claude, the prompt is only a small part of the context it gets. Much of your context is assembled from your system prompt, Skills, CLAUDE.md files, memory, and other sources." For a business owner using Claude or a Claude-powered agent day to day, your context is whatever standing instructions, brand guides, and saved preferences you have loaded before you ever type a word.

Anthropic's discovery is that stuffing that context with exhaustive rules stopped helping once the model got good enough to use judgment. Worse, it started actively hurting. The blog gives its own internal example: engineers reading transcripts of Claude Code kept finding contradictions like "leave documentation as appropriate" sitting next to "DO NOT add comments," both inherited from different layers of instructions written at different times. The model had to spend its reasoning untangling the fight instead of doing the work.

The old rule and the new one

Anthropic's post walks through several of these reversals directly. Here is the shift, in their own framing.

That last row is the one most small-business AI setups get wrong without realizing it. Every time you paste your full house-style guide into a new chat instead of saving it once as a reusable skill, you are repeating instructions the model already had. Anthropic found that repetition does not add safety with these models, it adds noise.

Isn't a shorter rulebook a worse rulebook?

It sounds backwards. More rules should mean more control. Anthropic's own data says the opposite happened for coding tasks, and the logic holds for a general business assistant too. A rule exists to prevent one specific bad outcome. Every rule you add is also a rule that can quietly contradict a different rule, or your customer's actual request, three months later when you have forgotten it is even in the document. The model then has to guess which instruction wins. Sometimes it guesses wrong, and you blame the AI instead of the pile of instructions it was handed.

The fix Anthropic settled on is not "give it no instructions." It is a split: a short, always-loaded core (who you are, what you sell, the two or three things that must never happen) plus a set of small, specific skills the AI pulls in only when the matching task shows up. That second half is the part any small business can copy today, no engineering team required.

How to build a business skill without writing a line of code

Claude's Skill Creator does this through a plain-language interview, not a code editor. You describe a task once, answer a handful of questions, and Claude assembles a reusable skill behind the scenes. Here is the version Priya actually ran, for the client weekly-status update she used to rewrite from memory every Friday.

1. Name the recurring task. She told Claude: "I want a skill for my client weekly status update."
2. Define the shape of the output. Claude asked what sections it should always include. She said: what shipped, what is blocked, what is coming next week, and a one-line summary at the top for clients who only read the first sentence.
3. Set the tone once. Direct, no corporate padding, short sentences. She said it in one line and never has to repeat it.
4. Tell it what to never do. Never pad a section to make it look full. If nothing is blocked, say "nothing blocked" and stop.
5. Let Claude assemble the skill. No file editing, no YAML, no markdown syntax. The interview produces a working skill she can trigger just by asking for her weekly update.

The same pattern works for a customer-support reply skill (tone, the three things support agents are never allowed to promise, how to escalate), a quote-follow-up skill (how many days to wait, what discount ceiling is allowed, when to loop in a human), or an inventory-reorder report. Each one stays a separate, small file. None of them needs to repeat what the others already say, which is precisely the progressive-disclosure structure Anthropic just rebuilt Claude Code around. You can see the same before-and-after shape in the diagram below: one long always-on rulebook that the agent has to fight through, versus a short core plus a handful of skills it opens only when needed.

What does claude doctor actually do?

Anthropic backed the advice with a working tool. The new claude doctor command (run as /doctor inside Claude Code) audits a developer's system prompt, skills, and CLAUDE.md files and flags exactly the problems described above: duplicate instructions, contradictions between layers, and content the model could already figure out on its own. Thariq's advice to developers was direct: "Across your system prompt, skills, and CLAUDE.md files, you may need to simplify just like we did."

You do not need to install Claude Code to borrow the habit. Run your own five-minute version of a doctor pass on whatever standing instructions you have loaded into your AI tools right now: list every absolute rule you can find (every "always," "never," and "do not"), cross out any two that quietly contradict each other, delete anything the AI could already infer from the conversation, and move anything task-specific into its own skill instead of the main document. That single pass is the small-business equivalent of what a multi-billion-dollar AI lab just did to its own product.

Why this is happening now

The context-engineering rewrite landed the same day Anthropic launched Claude Opus 5, priced at $5 per million input tokens and $25 per million output tokens, about half of Claude Fable 5's $10 / $50 pricing, according to Anthropic's model card and reporting by The Register. Coverage from that launch also flagged a side effect worth planning around: Opus 5's default responses run longer than earlier Opus models, so if you want shorter answers from your own AI setup, Anthropic's guidance is to ask for it explicitly rather than assume brevity by default.

Cheaper, more capable models change the math on what is worth automating. A shorter, well-structured context also means more of that cheaper token budget goes toward doing the actual work instead of the model re-reading a rulebook on every single message. If you have been holding off on giving Claude or a Claude-based agent a bigger role in your business because it felt unpredictable, this is the moment that predictability got a genuine engineering fix, not a bigger prompt.

What you have after this

Give this a week: trim your AI's standing instructions to the handful of things that must never move, and turn your most-repeated task into one Skill Creator interview.

Give it a month: three or four skills covering your weekly reports, your support replies, and your follow-up emails, each one small enough that you can read the whole thing in under a minute.

Give it six months: an AI setup nobody has to babysit, because it was never built as one giant rulebook to begin with. If you want a structured first pass at this, that is exactly what we build in the 14-Day AI Readiness Sprint, alongside the specification habits we cover in Stop Your AI From Guessing: Get Grilled Before You Build.

Same business. Same AI subscription. A completely different relationship with it.