Patch or Spec: The Question to Ask Before You Delete Your AI Instructions

Anthropic removed more than 80% of Claude Code's system prompt for its newest models and lost nothing measurable on its coding evaluations. The advice traveling with that number, delete your instructions, is half right and dangerous at scale. Instructions that patch a model's weakness should go. Instructions that encode your company's judgment are the only asset you have been quietly accumulating, and deleting those hands your differentiation to every competitor running the same model.

On July 24, 2026, Anthropic published that it had removed over 80% of Claude Code's system prompt for its Claude 5 generation models with no measurable loss on coding evaluations. The number moved fast. Four fifths of the guidance came out and the work held.

Days later, Boris Cherny, the creator of Claude Code, told a Y Combinator Startup School audience that everyone should be doing this. Delete your configuration files. Delete your skills. Delete your hooks. Every six months, see what the model does without them. In his account of Anthropic's internal testing, stripping the prompts entirely shows that "the model is actually a little bit more intelligent without these prompts."

Applied evenly across a company, that advice will destroy something you cannot rebuild in a quarter. There is a sharper test.

Why did Anthropic delete 80% of its own instructions?

The reason matters more than the number. Anthropic found it had been overconstraining its own product through the system prompt, the configuration files, and the skills layer. Reading transcripts of internal usage, engineers watched rules collide inside a single request: one line asking for documentation where appropriate, another forbidding comments outright. The model spent reasoning reconciling contradictions before it could touch the work.

Those contradictions were not carelessness. They were guardrails written for weaker models that could not be trusted to read a situation. The old rule was rigid, forbidding multi-line comment blocks. The replacement asks the model to match the comment density, naming, and idiom of the surrounding code. One freezes a single answer. The other hands over a judgment.

Anthropic's prompting guide for Claude Opus 5 is blunter. It instructs users to remove explicit verification instructions, because the model already checks its own work, and the extra instruction stacks on top of behavior that exists, producing over-verification that costs tokens and buys nothing.

What separates a patch from a spec?

Every instruction you have ever written to an AI system is one of two things.

A patch compensates for something the model got wrong. Double-check your answer. Do not invent statistics. Verify the source before you cite it. These are bug reports written in English, and they expire the day the bug is fixed. Anthropic just told you which of yours expired.

A spec transmits information that exists nowhere except inside your company. Our deck template uses this color and puts the source in the lower left. Our legal team has never once approved that word. Our CFO reads variance before totals. No amount of model intelligence produces any of that, because none of it is derivable. A frontier model could be ten times smarter and still have no path to your CFO's preferences.

The test fits in one sentence. Could a smarter model have worked this out on its own? If yes, you wrote a patch, and it is now dead weight. If no, you wrote a spec, and it is the only part of your AI setup a competitor cannot copy by buying the same subscription.

Which instructions did Anthropic actually keep?

The company that deleted 80% was precise about the remaining 20%. Its guidance says skills work best when they encode "opinions, knowledge, or best practices that are particular to you, your team, or product." The same document tells you to keep repository-specific gotchas and cut anything obvious from looking at the files. The Opus 5 guide tells you to prompt explicitly for the styles and templates the model must follow, and to prompt explicitly for response length, because the model will not guess either.

Read together, the instruction is narrower than the headline. Delete what the model can derive. Keep what only you know.

This is what I mean by institutional taste. For most of business history, judgment lived in individual heads and scaled only as fast as you could hire and train. The harness layer changes that, because taste written down becomes taste applied at volume. An organization can now run its standards across ten thousand documents instead of the twenty a senior person had time to review. Deleting the spec layer to chase a benchmark deletes exactly that.

How should a company run this audit?

Not as a purge. Gartner warned in May that applying uniform governance across AI agents produces failure in both directions: over-restricting simple agents slows delivery and pushes teams into shadow development, while under-restricting the autonomous ones outruns human oversight. The same asymmetry applies to instructions. A blanket delete is uniform governance wearing a different hat.

The pressure to move fast is real. CrewAI's February survey of 500 executives at large enterprises found 65% already running agents and 100% planning to expand this year. The strain is equally real. WRITER's 2026 survey found 79% of organizations hitting adoption challenges, a double-digit jump from 2025, with 54% of C-suite respondents saying AI adoption is tearing their company apart.

So run it on one workflow. Take a single documented process, strip only the lines that patch a model weakness, keep every line that carries your judgment, and rerun it against work you have already delivered so you have something to compare. Researchers call this ablation: remove one thing, measure, repeat.

Pay attention to who owns each layer. Patches usually come from whoever is technical enough to have written them. Specs almost always come from the middle of the org chart, from the manager who knows which client hates bullet points and why the third slide always fails. That layer is the one being cut hardest right now. Firing it before you have written down what it knows is how a company loses its taste and only finds out two quarters later.

Here is the test for Monday. Open the longest instruction file your team maintains and mark every line P or S. If most are P, you have been managing a model that no longer exists. If you cannot find a single S, the deletion debate is not your problem. Nobody has written down your judgment yet.

Frequently asked questions

What is the difference between prompting and context engineering?

A prompt is what you type in a single request. Context engineering covers everything else the model sees: the system prompt, saved instruction files, skills, memory, tool descriptions, and referenced documents. Because context is reused across many requests, it cannot be tailored to any one of them, which is why generic rules written into it cause more damage than a badly worded prompt.

Does deleting instructions actually save money?

Partly. Every instruction is read on every request, so removing dead ones cuts tokens. Anthropic's guidance notes that redundant verification instructions burn tokens with no quality gain. The larger saving is indirect: contradictory rules force the model to spend reasoning resolving them before starting work, so the cost appears as slower, longer runs rather than a line item.

How often should we audit our AI instructions?

Tie it to model releases rather than the calendar. Cherny's suggested cadence is roughly every six months, which tracks how often frontier capability has been moving. The practical trigger is a vendor publishing migration guidance that names instructions to remove, which is what Anthropic did on the day Opus 5 shipped.

Is there a real risk in removing guardrails?

Yes, and it varies by autonomy level. Gartner's guidance separates agents that suggest from agents that execute, and reserves rigorous controls including monitoring, rollback, and circuit breakers for the autonomous end. Deleting a stylistic patch on a drafting assistant carries little downside. Deleting an approval gate on an agent with permission to act is a different category of decision.

What happens to these instructions when the next model ships?

The patches expire again, which is the pattern worth internalizing. Anything written to correct model behavior has a shelf life set by someone else's release schedule. Anything written to transmit your standards survives every generation, because it was never about the model's capability. Sort your instructions once and the next migration becomes an afternoon rather than a rebuild.

Sharon Gai is an AI transformation strategist, keynote speaker, and author of How to Do More with Less. She advises Fortune 500 companies on AI adoption and organizational redesign.