Is it me, or did Claude get worse?

Many experienced Claude users are asking the same question. The evidence suggests a broader shift: Claude has evolved, Claude Code has changed around it and prompting habits that worked a few months ago now need re-testing.

"Either Opus 5 is worse than 4.8 or it expects something different from me."

That comment from an ordinary Claude user gets remarkably close to the heart of what has been happening.

Over the last few weeks, Reddit, Hacker News, GitHub and developer communities have filled with variations of the same questions.

• Why is Claude suddenly so argumentative?
• Why does it keep changing things I did not ask it to change?
• Why is it ignoring my instructions?
• Should I go back to Opus 4.8?
• Should I delete my CLAUDE.md?
• Should I just move to Codex?

Claude has been one of our primary working AI tools for months. We have built real client workflows around frontier models, using Claude Code and working across several providers rather than treating any single one as the answer to everything.

Recently, we noticed our own behaviour changing.

Tasks that would have automatically gone to Claude started being handled elsewhere. The friction increasingly came from the interaction itself: unsolicited objections, arguments about client decisions and a model that sometimes seemed more interested in reconsidering the brief than completing it.

We have found ourselves switching to ChatGPT for more of our work than before.

That made us ask the same question everyone else was asking.

Did Claude get worse?

We therefore set out to gather evidence and better understand the changes.

The change is real. "Worse" is harder to prove

There is no compelling evidence that Claude 5 is broadly less capable than Claude 4.x. There is plenty of evidence that it behaves differently.

Recent user reports repeatedly describe Opus 5 as more argumentative, opinionated, autonomous and willing to reinterpret instructions. We went looking for this because it closely matched what we were experiencing.

Some experienced users have switched back to Opus 4.8. Others are splitting work between Claude, Codex, ChatGPT and Gemini.

A benchmark can improve while the experience of working with the model gets worse.

Complaints on Reddit do not establish an underlying capability regression. They do tell us that a meaningful group of experienced users are encountering a different working relationship with Claude.

Anthropic's own documentation gives us a clue as to why.

Anthropic removed more than 80% of Claude Code's system prompt

On 24 July, Anthropic's Thariq Shihipar explained that the company had removed more than 80% of Claude Code's system prompt for models including Opus 5 and Fable 5.

Anthropic reported "no measurable loss" on its internal coding evaluations.

This choice of wording is significant, as it shapes how we interpret the results.

Anthropic did not claim that performance improved by removing 80%. It also did not publish enough detail about the internal evaluation suite for anyone outside the company to reproduce the result.

The more important point was what had accumulated inside the prompt.

Instructions designed to compensate for weaknesses in earlier models were increasingly unnecessary. In some cases, the additional rules were over-constraining newer models.

This shift fundamentally changes the rules of prompting.

For several years, one standard response to an AI system behaving badly was to add another instruction.

• Claude forgot something? Add it to CLAUDE.md.
• It failed to verify its work? Tell it to double-check.
• It did not use a tool? Make the instruction stronger.
• It ignored a workflow? Add examples.

Over time, prompts evolved layer by layer into complex operating manuals.

Newer Claude models need a different approach.

Your old prompts may be part of the problem

Anthropic's current guidance is more specific than "write shorter prompts".

The useful principle is minimum sufficient context.

Some historical workarounds can go. Scope, legal requirements, safety constraints and quality standards stay.

Conditional procedures increasingly belong in Skills or referenced files instead of being loaded into every interaction. Hard requirements are stronger when they also exist in permissions, schemas, tests and code.

Anthropic specifically warns that some inherited verification instructions can cause Opus 5 to over-verify, adding tokens and latency without improving the result.

Aggressive instructions designed to make earlier Claude models use tools can now cause newer models to over-trigger them.

This represents a fundamental change in how prompts should be constructed and maintained.

A CLAUDE.md file that worked brilliantly with Opus 4.8 may contain exactly the instructions Opus 5 no longer needs.

The previous implementation may have been completely sensible at the time. It can still become outdated.

Claude Code complicates the diagnosis

There is another reason "Claude got worse" is hard to diagnose.

Claude Code is more than a model.

The experience also depends on reasoning effort, system prompts, context management, compaction, memory, permissions, model routing and product defaults.

Anthropic's previous updates have already demonstrated how significantly those layers affect user experience.

In April, after months of complaints about Claude Code quality, Anthropic published a detailed postmortem. It found several product changes that had genuinely made the experience worse.

One of these changes involved lowering the default reasoning effort, which altered the balance between speed and output quality.

Another change repeatedly discarded older reasoning history, which led to forgetfulness, repetition and unusual tool choices.

A further adjustment (a system-prompt instruction intended to make Claude more concise) resulted in a 3% decline in one coding evaluation metric.

These were ultimately product-level problems rather than evidence of any sudden deterioration in the underlying model weights.

This distinction is essential to keep in mind whenever someone claims "Claude has got worse."

They may be entirely correct about the experience and mistaken about the cause.

How to tell what has actually gone wrong

Before rewriting your entire Claude setup or cancelling your subscription, isolate the failure.

What you are seeing	What to test first
Claude behaves differently immediately after changing model	Run the same task against the previous pinned model
A mature workflow performs worse on Opus 5	Run a clean-context version with only the essential instructions
Claude argues with or repeatedly reinterprets instructions	Audit CLAUDE.md, Skills and inherited context for conflicting or overly prescriptive rules
Claude forgets decisions during a long task	Check context and compaction behaviour
Claude starts doing work outside the requested scope	Review permissions, autonomy settings and explicit scope boundaries
Claude suddenly feels worse across unrelated tasks	Check recent Claude Code changes, service incidents and model routing
The new model performs well in a clean test but poorly in your existing environment	Treat the workflow as a migration problem

The most important principle is to change only one variable at a time.

If you change the model, system prompt, CLAUDE.md and tools together, you learn almost nothing.

Should you delete your CLAUDE.md?

No.

Review it carefully.

Anthropic's 80% figure came from Claude Code's own system prompt. It is not an instruction for every user to delete 80% of their CLAUDE.md.

A better process is:

Remove historical workarounds. If an instruction exists because an older Claude repeatedly failed in a particular way, test whether the current model still needs it.

Keep real constraints. Client requirements, scope boundaries, legal rules, safety instructions and acceptance criteria still matter.

Move conditional information out of permanent context. A procedure needed once every twenty tasks does not necessarily need attention during the other nineteen.

Put hard controls somewhere harder than prose. Tests, permissions and schemas are more reliable than hoping a model remembers the sentence halfway down a large instruction file.

Test every deletion. Prompt simplification should be an evaluation exercise, rather than a spring clean.

Of all these steps, testing every deletion is the most critical.

We have seen a rush towards "delete your CLAUDE.md" advice since Anthropic published the 80% figure. The evidence supports selective subtraction.

Blanket deletion simply replaces one untested assumption with another.

Anthropic did publish guidance. Finding it is another matter

KINTAL is part of the Claude Partner Network.

This provides a slightly uncomfortable example of the broader change-management problem facing many teams.

We were invited to Anthropic's Partner Series webinar on 22 July, two days before the Opus 5 announcement and the public explanation of the system-prompt reduction.

We did not attend.

The published agenda described model selection, evaluations and orchestration patterns. It did not signal that established Claude implementations might need prompt and context revalidation.

If it had been described that way, we would have made time for it.

Anthropic has since published substantial model-specific prompting guidance. Its documentation on CLAUDE.md, Skills, context and verification is useful.

A developer actively looking for "how should I change my CLAUDE.md for Opus 5?" can find the answer.

An organisation with a functional Claude setup built several months ago may not yet have any reason to ask that question.

The first indication of the need for migration may simply be that something in the workflow starts behaving unexpectedly or inconsistently.

It is important to note that publishing migration guidance is not the same as actively managing a migration process.

For partners and organisations running AI in real workflows, that distinction matters.

Frontier-model upgrades are becoming migrations

Software teams already understand this pattern elsewhere.

When a database, framework or API introduces a meaningful breaking change, mature teams assess compatibility before upgrading production systems.

Increasingly, frontier AI models deserve the same careful treatment as other production-critical software components.

A model name changing in a dropdown can look like a routine upgrade. Underneath it, reasoning behaviour, autonomy, instruction sensitivity and the surrounding harness may all have changed.

As a result, asking the old question, "Is the new model better?", is becoming less useful in practical terms.

Instead, the most practical question to ask is:

Is the workflow we built around the previous model still valid?

Sometimes the answer will be yes.

Sometimes a few instructions need removing.

Sometimes the newer model is genuinely worse for the specific behaviour you care about and staying on the older version makes sense.

Sometimes another provider is simply a better fit.

Multi-model working is already normal at KINTAL for precisely this reason.

We care more about whether the work is controllable and useful than which logo sits above the chat window.

Should you switch back to Opus 4.8 or move to Codex?

Treat that as an engineering decision rather than a loyalty test.

If Opus 4.8 currently gives you more reliable results on a critical workflow, pinning it while you test Opus 5 is rational.

If Codex produces better results for part of your development workflow, use Codex.

The mistake is switching before you understand what failed.

A badly migrated Claude workflow can make Opus 5 look worse than it is. Equally, endlessly rewriting your prompts to accommodate a model that performs worse for your specific task wastes time.

Run a direct comparison between models using your actual workflows and requirements.

Document the evidence you gather during these tests.

Then, make your decision based on these findings.

What this tells us beyond Claude

Anthropic provides a particularly clear example because the company has publicly explained some of what changed.

The same implementation risk exists across frontier AI.

A workflow tested against GPT, Gemini or Claude today cannot automatically be considered tested against the next generation carrying the same brand name.

The smarter these systems become, the more judgement they are being given.

That changes the controls they need.

AI governance therefore has another practical job: knowing when the assumptions built into an existing workflow have expired.

For organisations already using AI seriously, model migration needs to become part of normal operations.

Otherwise every major release carries the same bewildering first question:

"Is it me, or did the AI get worse?"

---

Frequently asked questions about Claude 5 and Claude Code

Did Claude get worse?

There is no strong evidence that Claude 5 is broadly less capable than Claude 4.x. There is credible evidence of changed behaviour, including greater autonomy and reports of more argumentative interactions. Claude Code product changes can also affect quality independently of the underlying model.

Is Opus 5 worse than Opus 4.8?

It depends on the task and the surrounding workflow. Some experienced users prefer Opus 4.8 because they find it easier to control, while Opus 5 performs strongly on more complex and longer-running work. Existing prompts designed around 4.8 can also behave differently on Opus 5.

Why is Claude more argumentative now?

Users have repeatedly reported Opus 5 challenging instructions, reframing requests or giving unsolicited objections. The evidence establishes a recurring user-experience pattern. It does not yet establish a single technical cause. Changed model behaviour, stronger autonomy and inherited context can all contribute.

Why does Claude ignore my instructions?

Start by checking whether your instructions conflict with other context, CLAUDE.md files, Skills or system-level behaviour. Newer Claude models follow some instructions more strongly, which means old workarounds and duplicated rules can create unexpected conflicts.

Should I delete my CLAUDE.md for Claude 5?

No. Audit it. Remove historical workarounds that the newer model no longer needs, retain genuine project and safety constraints and test performance before and after each substantial change.

How should I change my CLAUDE.md for Opus 5?

Keep information that Claude cannot safely infer, especially stable project rules, important constraints and relevant architecture. Move conditional procedures into Skills or referenced files and remove repeated or obsolete behavioural instructions. Test the revised file against real tasks before rolling it out across a team.

Do old Claude prompts still work with Claude 5?

Many will. Anthropic also documents several areas where inherited prompting can become inefficient or counterproductive, including repeated verification instructions and aggressive tool-trigger wording. Existing production prompts should be re-tested rather than assumed compatible.

How do I make Claude follow instructions?

Use clear outcomes and scope, remove conflicting or duplicated context and give hard requirements external enforcement where possible through tests, schemas and permissions. Adding more forceful prose is not always the answer with newer models.

Why does Claude Code do things I did not ask it to do?

Claude Code's behaviour is affected by both the model and its permissions or autonomy settings. Opus 5 and other newer Claude models are designed to be more agentic. Review scope instructions, permission modes and whether inherited prompts encourage excessive initiative.

Should I switch from Claude Code to Codex?

Compare them on the work that matters to you. Some users are switching or splitting workflows between the two, but current evidence does not establish a broad permanent move away from Claude. Cost, usage limits, control and task performance all influence the decision.

Which Claude model should I use?

Choose using your actual workload rather than a general model ranking. Test representative tasks for quality, controllability, latency and cost. A newer model is not automatically the best production choice for every established workflow.

How can I tell whether Claude, Claude Code or my prompt is causing the problem?

Change one variable at a time. Compare models using the same task and context, then test the newer model in a clean context. Check Claude Code settings and recent product changes separately. If the clean version works and your mature setup fails, your implementation is the first place to investigate.

---

Sources

This article draws on Anthropic's Claude 5 model-specific prompting guidance, Claude Code documentation, Anthropic's April 2026 Claude Code quality postmortem, its July 2026 explanation of Claude Code system-prompt reduction, Claude Partner Network material, GitHub reports and independent practitioner research reviewed by KINTAL in August 2026.

AI was used to compare our own experience with public discussion, organise the research and support the drafting process. The final conclusions and editorial judgement are KINTAL's.