Anthropic's rules, translated. Paste, cut, test.

Anthropic's Opus 5.5 prompting guidance, rewritten for people who run a business or a channel: one prompt template, two lines to cut, one guardrail to add, and a real test you can check.

Watch the video (15 min) · Read the companion article

Download the Claude Opus 5.5 Prompting Playbook (ZIP): free, no sign-up. It includes the fill-in-the-blank prompt template, a 10-minute instructions audit, guardrail lines for when Claude works in your tools, and my real weak-vs-strong test with Opus 5.5's full answers.

Anthropic published a prompting guide for Claude Opus 5.5 and a playbook for every current Claude model. Both are written mostly for developers. This is the version for people who run a business or a channel: what to paste, what to cut, and why. The companion article walks through each tip.

A few ground rules: this follows Anthropic's docs as they read on September 23, 2026, and docs change. Lines marked My read are my interpretation for app users, not Anthropic's words. I haven't tested every tip myself. And read anything Claude writes before it reaches a customer.

What's in the ZIP

File Use it for
01-prompt-template.md The fill-in-the-blank structure, plus three filled examples: a launch email, creator captions, a client brief
02-instructions-audit.md A 10-minute check of your custom or project instructions: what to cut, what to add
03-agent-guardrails.md Lines to add when Claude works in your email, docs, sheets or CRM
04-pasting-and-long-docs.md Pasting customer emails and web pages safely; long documents first, question last
05-design-avoid-list.md Stop generic-looking pages by naming the defaults you don't want
real-test/ My weak vs strong run: both prompts and Opus 5.5's full, unedited answers

The template

Brief Claude like a brilliant new hire. Long reference material goes on top, your request goes last:

<reference>
[Optional: long material such as menus, transcripts, notes or reviews.]
</reference>

You're [one-line role].

[Why + audience: who reads this, where, and why that matters.]

Here are [3 to 5] examples of what good looks like for us:
<example>[example 1]</example>
<example>[example 2]</example>
<example>[example 3]</example>

[The ask, as an action: "Write…", "Draft…", "Change…"]
[Length, format, and where it will be used.]

What's deliberately not in it: "think carefully before you answer" and "write out your reasoning."

Two lines to try cutting

  1. "Think carefully before you answer." For chat, Anthropic suggests considering removing this line for Opus 5.5. In their testing, removing it made replies start sooner with no clear drop in quality.
  2. "Write out your reasoning step by step." Anthropic says to remove instructions like this. The model already thinks before it answers, and prompts that push it to reproduce its reasoning in the reply can be declined.

If you want faster answers, lower the effort setting first. Opus 5.5 defaults to medium, and Anthropic says lowering effort reduces thinking more reliably than prompt instructions do. My read: not every app shows an effort setting; if yours doesn't, the rest still applies.

One line to add to any agent

When Claude works inside your email, docs or CRM, add this first:

Local, reversible actions are fine. Before anything destructive, hard to
reverse, or visible to other people (deleting files, sending messages,
posting, changing shared records), stop and ask me first.

What happened when I tested it

I gave Opus 5.5 (medium effort, fresh session, no tools) two prompts. "Write me a sales email" got a template full of placeholders and five questions back. The same request with who it's for, the next step, the length, the voice and the format got an email ready to send, plus a note on the two details it had guessed. One run each: a demonstration, not a study. Both full answers are in real-test/.

Start with one thing

  1. Download and extract the ZIP, then open START-HERE.md.
  2. Run the audit in 02-instructions-audit.md on the instructions you use most.
  3. Rewrite one weekly task with the template, run old and new on the same task, and compare how fast it answered and how much you had to fix.

The ZIP contains only text and Markdown files. The businesses in the examples are fictional. Running prompts uses your Claude plan or API credit; the playbook itself is free.

Download the playbook · Read the companion article · Watch the video

More free resources