Guides
7 min read

Writing AI Prompts for Claude and ChatGPT: Which Old Tricks to Drop and What Works Now

Colorful toy bricks: a neat narrow row stacked on the left, a loose pile of larger bricks on the right
Image: AI-generated
In short

Current AI models think on their own: phrases like "think hard", commands in all caps and vague bans do little or even hurt. What works better is a first message that states the goal, material, deliverable and stopping point, plus the effort level as the dial for how much the model thinks.

If you still write prompts the way you did in 2023, you are often slowing the AI down. On November 12, 2026, Anthropic published a guide to Claude Opus 5.5 that lists the habits users should drop, and the developer docs for Opus 5.5 add the test results behind it. Most prompt guides that rank at the top of Google, by contrast, still recommend tricks from the time before reasoning models. Below: what you can cut, what still works and what a template for your team looks like.

What has changed about writing AI prompts?

Current models think before they answer. You no longer control how much they think with phrases in the prompt but with a setting. Claude calls it the effort level, OpenAI calls it reasoning effort. On Opus 5.5, thinking is always on and the default level is medium. According to Anthropic, Opus 5.5 on medium matches or beats its predecessor Opus 5 on high in internal tests on coding and knowledge work.

That leads to three shifts:

  • Less pleading, more briefing. All caps, "absolutely" and "think hard" do little or hurt. Goal, material and deliverable are what count.
  • One assignment instead of bite-sized steps. Instead of steering step by step in the chat, you describe the finished result up front.
  • The dial instead of the sentence. If an answer is too shallow, raise the effort level first. If it is too slow or too expensive, lower it. According to the Anthropic docs, a lower level reduces thinking, cost and latency more reliably than any instruction in the prompt.

Which prompt tricks can you drop?

Five common habits are outdated with current models. The table shows what works instead.

Old habit Why it no longer helps What works instead
"Think carefully", "think hard" The model thinks anyway. In Anthropic's test with a chat product, replies without such lines started sooner, with no clear drop in quality Raise the effort level if the answer is too shallow
"IMPORTANT: You MUST always ..." Already with Opus 4.5 and 4.6, Anthropic saw the models respond more strongly to the system prompt and overdo such commands, for example when using tools Phrase it normally: "Use X when ..."
"No generic design", "don't make it look like AI" According to Anthropic, vague bans mostly swap one default style for another Name exactly what should be missing: "no cream backgrounds, no pill-shaped buttons"
Steering step by step in the chat Every round costs time, and the model doesn't know the goal Put the goal, the final deliverable and the stopping point in the first message
Outline first, then draft, then file A detour. Current models deliver finished documents and spreadsheets directly Ask for the finished file right away

Our take: the biggest lever is the fourth row. If you need ten short chat rounds, you usually only figured out the goal yourself while reading the third answer. Two minutes of thinking before the first message saves those rounds.

What still works?

The basics in Anthropic's prompting best practices apply to all current models. Four of them carry the most weight:

  1. Brief it like a smart new colleague. Anthropic describes the model as a brilliant new hire who doesn't know your norms. The test: would a colleague with no background get by with your prompt? If not, the AI won't either.
  2. Explain the why. Instead of "NEVER use ellipses", write: "The text will be read aloud by text-to-speech software that can't pronounce ellipses." With the reason, the model can also handle similar cases correctly.
  3. Say what you want, not only what you don't want. "Write in flowing prose paragraphs" works better than "no bullet points".
  4. Show examples. Anthropic recommends three to five examples, as close to the real case as possible and varied enough that the model doesn't copy an accidental pattern. Mark them clearly as examples, for instance with XML tags named example, so the model doesn't read them as instructions.

With long material, order matters: according to Anthropic, long documents belong at the top of the prompt, with the question and instructions below. Attach charts and tables as an image or file instead of typing out numbers. According to Anthropic, Opus 5.5 reads values from dense charts more accurately even on the lowest effort level than Opus 5 does on the highest.

Does this apply to ChatGPT too?

At its core, yes, with one distinction. The OpenAI docs on prompt engineering separate two kinds of models: according to OpenAI, reasoning models deliver better results when you give them only the goal and leave the details to them. Classic GPT models without a reasoning step, on the other hand, need precise instructions. OpenAI currently names gpt-6-astra as the default for general text tasks.

To structure a prompt, OpenAI, like Anthropic, recommends clear sections, for example with Markdown headings or XML tags. About its latest model, OpenAI writes that it tends toward detailed, heavily formatted answers. If you want short answers, you have to specify the style explicitly.

In practice: a prompt that clearly describes goal, material, deliverable and style works in both worlds. Vendor-specific tweaks only pay off once a prompt runs thousands of times in an automation.

What does a good prompt template for your team look like?

A good template has five building blocks, based on Anthropic's advice to state the goal, the final deliverable and the stopping point right at the start:

Goal: What do you need the result for, and who will read it?
Material: Documents, data, links (paste long texts first)
Deliverable: Format, length, file (e.g. "Excel with columns A to D")
Style: Tone, house rules, specific no-gos with the reason
Stop: When is it done, and when should the AI ask instead?

An example from sales: "Goal: I'm on a call tomorrow with a prospect from the machinery sector and want to be prepared. Material: their website copy and our last email, pasted below. Deliverable: one page with five questions I should ask and three likely objections with answers. Style: matter-of-fact, informal, no superlatives, because the client dislikes marketing speak. Stop: If you're missing details on company size, ask instead of guessing."

Store your house style and brand rules once in projects or system instructions instead of copying them into every prompt. How to write such rules for AI is covered in our article on brand guidelines for AI.

What does this mean for you?

For teams in everyday office work, one afternoon is enough: go through your prompt collection, delete the thinking phrases and all-caps commands, and replace bans with concrete requests plus a reason. Leave the effort level on medium and only turn it up when needed. Anthropic advises testing a higher level before switching to a more expensive model.

For automations in Make or n8n, the review matters more, because old system prompts keep running there unnoticed. These are the points to check:

  • Old thinking settings. According to the Anthropic docs, a fixed thinking budget (budget_tokens) returns an error from Claude 4.7 onward. Opus 5.5 also rejects disabled thinking with an error. A workflow that suddenly fails after a model switch often has its cause right here.
  • Prefilled answer openings. Anthropic no longer supports prefilling the response since the 4.6 models. Instead, instruct the desired format directly.
  • Overcharged instructions. According to Anthropic, "If in doubt, always use tool X" leads to excessive tool use with current models, meaning more steps and higher cost.
  • Pasted outside content. If your workflow processes emails or web pages, they can contain instructions the model follows. Anthropic recommends marking such texts clearly as pasted content with unambiguous tags, and at the same time warns that tags can be faked. So they don't replace a safety check before a workflow sends emails or changes data.
  • Recalculate costs. According to Anthropic, Opus 5.5 costs 4 dollars per million input tokens and 20 dollars per million output tokens, Sonnet 5.5 half that. On high effort levels, though, Sonnet's cost per task gets close to Opus, because the stronger model often finishes a task in fewer steps. Where Haiku is enough is covered in our article on Claude Haiku 5.5.

Our take: if you test the prompts in your automations against ten real cases after every model switch instead of just letting them run, you avoid most surprises. Want to adapt the template to your own processes, say for quotes or lead handling, and test it against real cases after every model switch? We build that into your automation with you: Talk to us.

Frequently asked questions

Do I still need prompt engineering now that models are this good?

Yes, but differently. Tricks for coaxing the model lose value, clear briefing gains value: what the model doesn't know, such as your customers, rules and goals, it can't guess even with a lot of thinking effort.

Where do I set the effort level in Claude?

According to Anthropic, click the model name next to the send button, then Effort. On Opus 5.5, medium is the default.

Does it help to give the AI a role, like "You are a tax advisor"?

A role in the system prompt can sharpen tone and focus, and Anthropic explicitly mentions it. It doesn't replace stating the goal, the material and the result you want, though.

Do I have to rewrite all my old prompts now?

No. According to Anthropic, prompts written for Opus 5 also work well with Opus 5.5. Focus your review on prompts in automations and on prompts with thinking phrases or all-caps commands.

Sources

  1. Anthropic: Using Claude Opus 5.5 as your daily driver for knowledge work
  2. Anthropic: Prompting best practices
  3. Anthropic: Prompting Claude Opus 5.5
  4. OpenAI: Prompt engineering
This article was created with AI.
Got a project in mind?

Tell us what you're planning. You'll get an honest assessment within 24 hours.

Got it! We will get back to you asap!
Whoopsie! Something went totally wrong here... Maybe just send an email to hi@blocksandcolors.de? Sorry!
Or email me directly: hi@blocksandcolors.de