Summary of DIGITAL STORM Weekly #173 – “Your Best Skill Never Fires” by Joerg Storm

This edition explores one of the biggest recent developments in AI automation: Anthropic’s new “Record a Skill” feature in Claude Cowork, which lets users create reusable AI workflows simply by recording themselves performing a task. However, Joerg Storm argues that creating AI skills is no longer the difficult part—making them trigger at the right time is.

Executive Summary

The newsletter’s central message is:

The success of an AI workflow depends far more on how it is described than on how well it is written.

Most users spend time perfecting their AI instructions, but Claude rarely sees those instructions unless the skill is selected first. Instead, Claude initially looks only at a skill’s:

• Name

• Description

These two fields determine whether the complete workflow is ever loaded.

Key Takeaways1. Claude Can Learn Workflows by Watching You

Anthropic introduced Record a Skill, allowing users to:

• Record themselves performing a task

• Let Claude convert the recording into a reusable workflow

• Run it automatically later

Examples include:

• Expense reports

• Email responses

• Monthly reporting

• LinkedIn posting

• Trading alerts

This dramatically lowers the barrier to creating automations.

2. Creating Skills Isn’t the Problem Anymore

Storm argues that automation creation has become almost effortless.

The real challenge is now:

Can Claude recognize when to use your saved skill?

Many workflows remain unused because Claude never identifies them as relevant.

3. The Description Is More Important Than the Instructions

Claude loads skills progressively.

Stage 1 (Always Loaded)

• Skill name

• Description

Stage 2 (Loaded Only if Matched)

• Full instructions

• Workflow logic

Stage 3

• Attached resources/files

This means:

The description functions like a search index.

A vague description means the skill may never activate, regardless of how good the underlying workflow is.

4. Think of Skills Like Books in a Library

Storm uses a memorable analogy.

Imagine a librarian who:

• can only read book spines

• cannot open any books

If your book’s spine says:

“Guide v3”

Nobody knows what’s inside.

If instead it says:

“Monthly Revenue Pack Builder”

the librarian immediately knows when to retrieve it.

The lesson:

Your description is the book spine—not documentation.

5. Write for Retrieval, Not Readability

Descriptions shouldn’t be elegant.

Instead, they should include the exact words users naturally type.

Good trigger phrases include:

• monthly report

• close the numbers

• revenue pack

• board deck

• MRR

• churn

The more closely your description mirrors real user language, the more reliably the skill activates.

6. One Skill = One Job

Many people create one large “super skill.”

Storm explains why this is ineffective.

Research cited in the newsletter shows:

• Two or three focused skills outperform one broad skill.

• Broad skills trigger too often.

• False activations waste context and reduce AI performance.

His recommendation:

Keep every skill focused on one specific task.

7. Recording Introduces New Risks

Recorded skills are inferred by AI rather than manually written.

Potential issues include:

• incorrect assumptions

• hidden logic

• security vulnerabilities

• unintended actions

Storm recommends reviewing every recorded skill before using it in production.

8. Security Shouldn’t Be Ignored

The newsletter highlights research showing many publicly shared AI skills contain security issues, including malicious payloads.

The advice:

• Never blindly install downloaded skills.

• Review them as carefully as you would third-party code.

Best Practices for Building AI Skills

Storm recommends:

• Use one skill for one job.

• Write descriptions using real user vocabulary.

• Include phrases people actually type.

• Clearly specify when not to use the skill.

• Test with positive and negative examples.

• Avoid catch-all workflows.

AI News Highlights

The newsletter also covers several recent AI developments:

• Claude Cowork now supports workflow recording through Record a Skill.

• Anthropic introduced per-skill usage and cost analytics, helping organizations identify which automations are actually used.

• A multimodal update showed that image cropping significantly improves chart-reading accuracy.

• Anthropic launched Claude for Teachers, offering premium access and education-focused resources.

• Enterprise security scanning features entered public beta.

• The Agent Skills specification has become an open standard supported by multiple vendors, improving portability across AI ecosystems.

AI Term of the Week

Progressive Disclosure

Instead of loading every instruction at startup:

1. Metadata loads first.

2. Instructions load only after a match.

3. Resources load only when required.

This approach reduces token usage while keeping AI responsive.

AI Tutorial

The newsletter walks readers through creating their first reliable AI skill in about 30 minutes:

1. Choose a repetitive weekly task.

2. Record or write the workflow.

3. Collect five real phrases users would use to request the task.

4. Collect three similar phrases that should not trigger it.

5. Rewrite the description using those trigger phrases.

6. Test the skill in fresh conversations.

7. Refine based on results.

Main Lesson

The biggest insight from this edition is:

The future bottleneck of AI automation isn’t building workflows—it’s making sure the right workflow activates at the right time.

As AI tools make workflow creation nearly effortless, organizations will accumulate large libraries of skills. Success will depend less on the quality of the instructions and more on how well each skill is indexed, described, governed, and tested.

Final Takeaway

This newsletter shifts the conversation from prompt engineering to skill engineering. Joerg Storm argues that the next competitive advantage in AI won’t come from writing better prompts—it will come from designing AI skills that are discoverable, narrowly focused, securely managed, and reliably triggered. For teams building internal AI automations, metadata, governance, and testing become just as important as the workflows themselves.

https://drstorm.substack.com/p/your-best-skill-never-fires-173?utm_campaign=email-half-post&r=5xcpdm&utm_source=substack&utm_medium=email