Claude Opus 5 launched July 24, and it's been the most discussed model release on X all week. Not for its benchmark scores — for its prompting.

The week's two top AI stories on X's trending tab were literally "Claude Users Adopt Simplified Prompts for Clearer Responses" and "Practical Tweaks Make Claude Output Clearer and More Human." And the most viral Claude post of the week came from Matt Shumer, who put the thesis bluntly:

"Claude Opus 5 is way better than people give it credit for. If you're using it like previous Claude models, it's going to suck. Two things make a huge difference: Delete ALL your skills/MCPs/Claude.md/etc. Start fresh. Stop telling it how to do the thing. Just say what you want."

~756K views. That's a signal.

Why Opus 5 Is Different

Opus 5 is Anthropic's most agentic model yet — built for long-running agents that work across files, tools, and time. It has stronger instruction-following, better tool use, and notably, it's better at figuring out how to do things itself.

That last point is the whole story. The prompting style that made earlier models perform — heavy system prompts, skills, MCP servers, and step-by-step scripts — was a workaround for models that couldn't infer intent. Opus 5 can. When you keep loading it with the old scaffolding, you're not helping it. You're:

Boxing it into a path that doesn't account for edge cases it would otherwise handle
Adding noise to a model that already knows how to plan
Constantly invalidating its context with instructions it doesn't need

This is the same shift OpenAI signaled with their simplified prompting guide in July: lead with the result, not the recipe. Anthropic's users are arriving at it independently, one viral post at a time.

What "Simplified" Actually Means

Simplified doesn't mean one-liners. It means rebalancing your prompt toward goals and constraints, away from process.

Before (Opus 3-4 style)

You are a senior staff engineer. Follow these steps exactly:
1. Read src/auth/*.ts to understand the current flow
2. Create a diagram of the auth flow
3. Propose 3 refactoring options with tradeoffs
4. For each option, list files affected, risk level, and effort
5. Recommend one and justify it
6. Do not write any code yet
Output format: use tables where possible...

After (Opus 5 style)

Refactor src/auth/ so that token validation is a single
reusable middleware. Keep the public API identical, keep
all tests green, and don't change any behavior. If you see
a cleaner structure than what I described, do that instead.

Both prompts take about the same effort to write. The second one works better on Opus 5 — because it states the outcome, the constraints that matter, and explicitly hands the how to the model.

The Skills and MCP Question

The aggressive part of Shumer's advice — delete all your skills/MCPs — deserves nuance, because "skills" is doing a lot of work in this sentence.

On Opus 5:

Skills that encode generic process ("how to write a commit message," "how to review a PR") get in the way. The model knows these. They're the recipe problem in file form.
MCP servers that just add tools you rarely use waste context and complicate tool selection. Opus 5 picks tools better when there are fewer, more relevant ones.
Context files full of step-by-step procedure (CLAUDE.md with "always do X then Y") have the same effect as over-specified prompts: they constrain instead of inform.

What still matters:

Domain knowledge — facts about your codebase, your conventions, your infrastructure. Opus 5 doesn't know those; that's not "recipe," that's context.
Constraints and guardrails — the one or two boundaries that genuinely must hold ("never touch the payment files," "only use sources in the repo").
A clear outcome — what done looks like, and how you'll verify it.

The pattern that's emerging from the X discourse: context files get smaller and more factual; prompts get more goal-oriented; scaffolding gets deleted and re-added only when the output demonstrably improves.

The "Stop Telling It How" Anti-Pattern

The other half of the advice is the hardest for experienced users to adopt, because it reverses years of conditioning. On Opus 5, these all count as telling it how:

"First, break this down into subtasks..."
"Read these 3 files before responding..."
"Think step by step..."
"Act as a [persona]..."
"Here's the exact algorithm you should follow..."

Opus 5 plans, decomposes, and reasons internally. Pre-imposing your plan on it is like handing a senior engineer a junior's checklist — it'll follow it, but the output is worse than if you'd stated the goal and let it do its job.

The exception is when the process itself matters — regulatory steps, a specific build pipeline, a sequence that has to happen for compliance. State process then. Otherwise, state the outcome.

A Practical Opus 5 Prompting Checklist

State the outcome first — what done looks like, in one or two sentences.
Give it your facts, not your procedure — the domain knowledge it can't know.
Add constraints, not steps — the 1-2 boundaries that must hold.
Hand over the how — tell it you trust its judgment on approach.
Audit your scaffolding — for every skill/MCP/CLAUDE.md rule, ask: does this encode knowledge I have that the model doesn't, or is it recipe? Delete the recipe.
Iterate from minimal — start with the simplest version that could work, measure, add structure only where it demonstrably helps.

The Bigger Picture

This week's discourse is the clearest sign yet that the prompt engineering era is shifting. The models that dominated 2023-2025 needed prompt engineering in the literal sense — engineering to compensate for their limits. The 2026 frontier models need intent communication — clearly saying what you want and getting out of their way.

That doesn't make prompting obsolete. It makes the craft different: fewer words, sharper constraints, more domain context, less process. The system prompt engineering pages on this site are built on exactly that rebalance — and the Claude Code prompting guide has more examples of goal-first prompting for agent workflows.

If you're still prompting Opus 5 the way you prompted Opus 4 — same skills, same MCPs, same step-by-step scripts — this week's trending tab is your sign to try the opposite. Delete the scaffolding. State the result. See what the model does when you actually let it work.

Shumer's post: @mattshumer_ on X. Anthropic's Opus 5 announcement: anthropic.com/news/claude-opus-5.

Claude Opus 5 Wants Simpler Prompts — Delete Your Skills and Just Say What You Want