How Claude Killed Prompt Engineering (And What To Do Instead)

You spent the last two years learning to write longer, more detailed prompts for AI — more rules, more context, more “please do exactly this.” Turns out that instinct might now be working against you. In one week, Anthropic shipped three things that all point the opposite direction. Okay so here’s what actually happened, and what to change in how you use Claude starting today.

What Actually Happened (3 Moves, 1 Week)

For two years, the advice was consistent: write longer prompts, add more context, spell out every rule. That was the skill everyone told you to learn.

Then Anthropic shipped three things in the same week that quietly say the opposite is now true: Record a Skill, Claude Opus 5, and new guidance on context engineering. None of them are small tweaks. Together, they point toward giving Claude less instruction, not more.

Think of it like managing people. A junior employee needs a step-by-step checklist for everything. A senior one just needs the goal and the context — and honestly, handing them a checklist usually slows them down. The Claude 5 generation of models is starting to behave more like the senior employee.

Move 1 — Record a Skill: Claude Learns by Watching You

Record a Skill lets Claude learn a task by watching you do it, instead of you writing instructions for it. You record your screen, narrate what you’re doing and why, and Claude turns that recording into a reusable skill it can run on its own later.

Here’s how it works:

1. Open the Claude desktop app. This feature lives in the Cowork tab, not the standard chat interface.
2. Click the + button and select Record a Skill. This starts your screen recording session.
3. Do the task normally and explain as you go. Claude uses your narration to understand the intent behind each action, not just the clicks.
4. Stop recording. Claude processes what it saw, names the skill, writes the steps, and stores it for later.
5. Run the skill later. Next time, just trigger it. No re-explaining, no new prompt.

💡 Tip: start with a task you already do at least twice a week — formatting a report, processing a client email, updating a spreadsheet. The more repetitive the task, the more time the skill saves you.

Move 2 — Claude Opus 5: Less Instruction, Better Results

Claude Opus 5 came in within about half a percentage point of Claude’s larger Fable 5 model on coding benchmarks, at roughly half the cost — a result independently confirmed by outlets covering the launch. But the more interesting finding for how you actually use it: detailed, heavily structured prompts didn’t outperform minimal ones. In a lot of tests, they performed worse.

Here’s the difference in practice:

Old prompt — less effective with Opus 5:

You are a professional email writer. Your task is to write a follow-up email.
Follow these rules exactly:
1. Keep the email under 150 words
2. Use a professional tone but not overly formal
3. Reference the previous meeting
4. Include a clear call to action
5. Do not use phrases like "I hope this finds you well"
6. End with "Best regards"

Write a follow-up email to a client after a product demo.

New prompt — more effective with Opus 5:

Write a short follow-up email to a client after a product demo.
Goal: get them to book a next step meeting.
Tone: warm but direct.

The principle: Opus 5 was trained to understand intent. The more you constrain exactly how it thinks, the less it can apply that training. Give it the destination, not the turn-by-turn directions.

Move 3 — Context Engineering: Why Anthropic Deleted 80% of Claude Code’s Instructions

Anthropic published new guidance on context engineering for the Claude 5 generation, and the core finding is blunt: instead of building large, detailed system prompts, these newer models respond better to minimal, well-structured context. Independently confirmed reporting says Anthropic removed over 80% of Claude Code’s original system prompt — and it ran faster and performed better, not worse.

Here’s roughly how BULDRR AI would summarize the shift in plain terms (this is our own condensed take on the guidance, not a word-for-word Anthropic quote):

1. Give goals, not procedures. Tell Claude what success looks like and let it figure out how to get there.
2. Remove rules that hedge. If a rule starts with “try to” or “if possible,” cut it — hedges weaken confidence without adding control.
3. Use examples, not explanations. One example of the output you want beats ten sentences describing it.
4. Keep context relevant. Every sentence in your context window that isn’t directly useful pulls focus from what matters.
5. Let the model self-check. Ask Claude to verify its own output instead of writing separate validation rules.
6. Start minimal, add only what fails. Begin with the shortest system prompt that could work, and only add a rule when you see a specific failure.

Here’s what that looked like for Claude Code’s own system prompt, simplified:

Before context engineering:

You are Claude Code, an AI coding assistant. You help users write, debug, and review code. Always follow these guidelines:
- Write clean, readable code
- Add comments to complex logic
- Explain your changes
- Ask for clarification when requirements are unclear
- Follow the user's existing code style
- Suggest tests when relevant
- Handle errors gracefully
[... 80% more rules like this ...]

After context engineering:

You are Claude Code. Help the user build working software.

When you make changes, explain what you changed and why.
When you are unsure, ask one clarifying question before proceeding.

Same model, far less instruction, better results.

What This Means for How You Use Claude

Here’s the checklist worth actually doing, not just reading:

1. Audit your system prompts. Go through every system prompt or CLAUDE.md file you have. Cut every rule that starts with “try to,” “if possible,” “generally,” or “where relevant.”
2. Replace rules with examples. For every formatting or style rule you’ve written, swap it for one example of the output you actually want.
3. Record your first skill this week. Pick one task you do more than twice a week. Recording it takes under 10 minutes, and it runs on its own from then on.
4. Test Opus 5 on your complex tasks. If you’re currently defaulting to a bigger, pricier model for coding or analysis, compare Opus 5 — it’s landing close on performance at a lower cost.
5. Stop writing longer prompts by instinct. When Claude underperforms, the old habit was to add more rules. Try removing rules and clarifying the goal instead.

Fear removal, honestly: you don’t need to redo everything today. Pick one system prompt or one repetitive task, apply one change, and see what happens before you touch anything else.

Common Questions People Ask

Is prompt engineering completely dead?

Not yet, but it’s becoming a transitional skill rather than a permanent one. Minimal, goal-first prompting and context engineering are replacing the old habit of longer, rule-heavy prompts.

Do I need Claude Opus 5 specifically for this to matter?

No. The context-engineering and “give goals, not procedures” principles apply to prompting Claude generally — Opus 5 is just where the effect showed up clearest in benchmarks.

What’s the difference between context engineering and prompt engineering?

Prompt engineering focuses on wording a single instruction well. Context engineering is broader — it’s about what information, examples, and structure surround the model overall, and deliberately keeping that lean.

Is Record a Skill available on the free plan?

Based on current reporting, Record a Skill is available on paid Claude plans through the Cowork tab of the desktop app — check your plan’s current feature list, since availability can change.

Quick Recap

• Anthropic shipped three things in one week: Record a Skill, Claude Opus 5, and new context-engineering guidance
• Record a Skill lets Claude learn a task from a screen recording instead of a written prompt
• Opus 5 performs better with short, goal-first prompts than long rule lists
• Anthropic cut over 80% of Claude Code’s system prompt and it ran better
• The fix for you: audit your prompts, cut hedges, use examples, and record one skill this week

What Should You Do Next?

Your one action right now: open one system prompt or CLAUDE.md file you already use, delete every line that starts with “try to” or “if possible,” and run it again. That’s it — nothing to install, nothing to break, and you’ll see the difference in one test.

If you want to see this same minimal-instruction thinking applied to a real Claude Code project, check out BULDRR AI’s beginner’s guide to n8n for a hands-on example. And if you’d rather start from ready-made setups than build your own from scratch, BULDRR AI’s free workflows library has over 10,000 templates you can just download and use.