On 24 July, the same day Claude Opus 5 launched, Anthropic announced that it had removed more than 80% of Claude Code’s system prompt for models like Opus 5 and Fable 5.
Three days later, Boris Cherny, the creator of Claude Code and now Head of Product at Anthropic, gave founders similar advice during Y Combinator’s Startup School:
Every six months, delete your CLAUDE.md, your skills and your hooks. Then see what the model can do without them.
Both examples came from Claude Code, but Anthropic’s warning isn’t limited to it. The same issue can appear whenever a Claude 5 generation model loads an older prompt or skill, whether you’re working in a regular Claude chat or Cowork.
A couple of weeks ago, in my guide to prompting Fable 5 and Sonnet 5, I shared free Amplifiers that rewrite your prompts for these models.
Now we need to apply the same thinking to the reusable instructions you rely on every day: your skills.
And I know how annoying that is after spending the past year carefully building skills and adding another instruction every time Claude ignored one. But if you’ve already noticed your skills getting worse with the newer models, you need to do this.
So today, we’re fixing your skills library so it works better with Sonnet 5, Opus 5 and Fable 5.
And because none of us wants to spend days rebuilding skills that used to work perfectly, I built everything you need to get the whole job done fast.
By the end of this article, you’ll leave with:
A skill that audits and rewrites any skill. It shows you what to improve and creates a separate version for the Claude 5 models that you can review, modify and test against the original before replacing anything.
A workflow that handles your entire skills library. Run it in Cowork or Claude Code and Claude will audit all your skills at once, then give you rewrites to review, modify or approve, saving you days of work.
P.S. Some of you have written to me that the newer Claude models no longer follow the Voice DNA skill you built, even though it worked like a charm before. It’s the same for me, so a dedicated update to the writing skill I shared previously is coming next week.
Here’s what we’ll cover today:
The new dos and don’ts for Claude skills on Claude 5
How to audit and rewrite any Claude skill for Claude 5 models
Step 1: Download the skill
Step 2: Upload it to Claude
Step 3: Audit and test your most important skill
How to fix your entire Claude skills library in bulk
Step 1: Paste the bulk workflow
Step 2: Work through the results with Claude
Step 3: Upload and test the new skills
Which Claude skills to update first
The new dos and don’ts for Claude skills
Here’s what you should know before opening a single skill: Claude still needs clear instructions. What changed is how much scaffolding it needs around them. The Series 5 models follow instructions more literally, verify more on their own and handle longer tasks with less help.
So we’re keeping the instructions that define good work and reviewing the ones we added to compensate for older models.
Dos
Keep the job and boundaries clear. Say what the skill does, when Claude should use it and what result you expect. The description controls whether the skill loads, while Sonnet 5 needs the scope of each instruction stated explicitly.
Control the length directly. Opus 5 tends to write longer responses and documents. A short instruction such as “keep the answer brief” still works, while lowering the effort level does not reliably shorten the final output.
Give Claude something concrete to follow. For writing, use full samples of your voice. For other skills, this might be a template, script, worked analysis or reference data. These often carry more useful information than another page of descriptive rules.
Split long skills into smaller files. Anthropic recommends keeping the body of
SKILL.mdunder 500 lines. Link reference files directly fromSKILL.md, and add a table of contents when one exceeds 100 lines.
Don’ts
Don’t ask for maximum thoroughness by default. Instructions such as “be exhaustive” can make Claude do more than the task needs. Broad filters such as “only report major issues” or “be conservative” can also suppress useful findings. Tell Claude how deep to go and where to stop.
Don’t add generic verification loops. Opus 5 already checks and corrects its work, so “double-check every answer” may only add time and tokens. Keep specific, evidence-based verification for critical operations and long-running Fable 5 workflows.
Don’t ask Claude to show its reasoning. On Fable 5, instructions to reveal, transcribe or explain its internal thought process can trigger a
reasoning_extractionrefusal and cause the task to fall back to Opus 4.8. Ask for the conclusion and supporting evidence instead.Don’t micromanage progress updates. Sonnet 5 already gives regular updates during longer tasks, while Opus 5 can narrate too much. Describe the timing and style you want only when Claude’s default does not suit the workflow.
And don’t replace the old version immediately. Test both skills on a few tasks you already know well, using fresh conversations, and keep the rewrite only when it performs better consistently.
Let’s get your skills ready for the Claude 5 generation models
Now that you know what to look for, let’s put it to work.
First, I’ll give you the skill I built to audit and rewrite your existing skills for the Claude 5 generation models.
Then I’ll give you the bulk workflow for Cowork or Claude Code, so you can repeat the process across your entire skills library without spending days doing it one by one.
How to audit and rewrite any Claude skill for Claude 5 models
Let’s start with one skill you know well, so you can see how the process works before applying it to your whole library. You’ll install the audit skill, use it to review and rewrite one of your existing skills, then test the new version alongside the original.
Step 1: Download the skill
Open the Google Drive link below and download the series-5-skill-audit skill. Keep the file zipped because that’s the format Claude needs.