Summary: Stop Feeding Claude. Start Directing It (#172) by Joerg Storm

Joerg Storm’s latest newsletter shifts the conversation from using AI models to designing efficient AI workflows. The central message is simple: more context doesn’t make Claude smarter—it makes it slower, more expensive, and often less effective. Instead of stuffing Claude with large instruction files, users should organize their work using Skills and Projects.

Key Takeaways

1. More Context Is Not Better Context

Many users create massive folders containing personal profiles, company information, writing styles, and previous work, assuming Claude will produce better results.

Instead, Claude Cowork:

• Re-reads this information during every reasoning loop.

• Spawns multiple Claude sessions that repeatedly load the same files.

• Consumes significantly more tokens.

• Produces increasingly generic outputs.

The takeaway: Every unnecessary document becomes a recurring cost—not just financially, but also in performance and creativity.

2. Claude Cowork Works Like an AI Team

Claude Cowork isn’t a single chatbot.

It:

• Plans tasks

• Launches multiple AI agents in parallel

• Searches the web

• Reads files

• Runs code

• Revises outputs repeatedly until the task is complete

Since every agent repeatedly accesses your provided context, excessive information becomes a multiplier for both latency and cost.

3. Replace Huge Folders with Two Building Blocks

Storm recommends replacing complex folder structures with just two concepts:

Skills

A Skill contains how a task should be performed.

Examples:

• Writing proposals

• Creating weekly reports

• Formatting presentations

• Drafting blog posts

Skills:

• Load only when needed

• Are reusable across projects

• Can be shared with teammates

• Follow Anthropic’s open Skill standard

Think of a Skill as:

“Teach Claude the method.”

Projects

A Project stores what Claude needs to know for one specific context.

Examples:

• Client A

• Marketing Campaign

• Product Launch

• Research Project

Projects contain:

• Relevant documents

• Client files

• Notes

• Instructions

Nothing else.

Think of a Project as:

“Give Claude only today’s facts.”

4. The Biggest Hidden AI Cost Is Conversation Length

Many users assume expensive models drive AI costs.

Storm argues that the real cost often comes from:

• Long conversations

• Endless corrections

• Constant back-and-forth

Since Claude rereads the entire conversation every turn:

Long chats =

• More tokens

• Slower responses

• Higher costs

Better habits include:

• Restart conversations instead of endlessly editing

• Summarize long threads before continuing

• Batch related requests together

• Use premium models only for complex work

5. The “Skill vs Project” Rule

A simple decision framework:

Examples:

Skill

• Weekly board report

• Proposal template

• Blog writing workflow

Project

• Client contracts

• Product launch documents

• Campaign assets

This separation keeps Claude focused and minimizes unnecessary context.

6. Common Mistakes

Avoid:

• Giant “About Me” documents

• One massive folder for every client

• Long global prompts

• Using Skills for client-specific data

• Using Projects for reusable methods

Instead:

• Keep Skills generic and reusable.

• Keep Projects narrowly scoped and context-specific.

7. Build Your First Skill

Storm recommends creating Skills for tasks you repeat frequently.

Examples include:

• Weekly status reports

• Proposal writing

• Meeting summaries

• Social media posts

• Email drafting

His “Skill Architect” prompt interviews you about your process before generating a reusable Skill, helping capture tacit knowledge that teams can reuse consistently.

Other AI Updates

The newsletter also highlights several noteworthy developments:

• Cartesia topped both speech-to-text and text-to-speech leaderboards with its Sonic-3.5 and Ink-2 models, powered by State Space Models (SSMs), enabling faster, multilingual voice AI.

• Claude Cowork now supports background sessions across desktop, web, and mobile, allowing long-running tasks to continue seamlessly.

• Anthropic’s Skills are based on an open standard, making them portable across compatible AI agents instead of locking users into one platform.

Final Thoughts

This edition argues that the future of AI productivity depends less on choosing the most powerful model and more on how you structure work.

The key lessons are:

• Use less but more relevant context.

• Separate reusable methods (Skills) from task-specific information (Projects).

• Keep conversations short and focused to reduce costs.

• Think of Claude as a team of AI workers that needs clear organization rather than a chatbot that benefits from endless instructions.

The overall message is that effective AI systems are built through thoughtful workflow design, not by overwhelming models with information.

https://drstorm.substack.com/p/stop-feeding-claude-start-directing?utm_campaign=email-half-post&r=5xcpdm&utm_source=substack&utm_medium=email