A Skill Is A Folder, Not A Prompt: What Anthropic Learned Running Hundreds Of Them

📊 Full opportunity report: A Skill Is A Folder, Not A Prompt: What Anthropic Learned Running Hundreds Of Them on ThorstenMeyerAI.com — validation score, market gap, and execution plan.

TL;DR

Anthropic has demonstrated that modeling AI capabilities as ‘Skills’—folders containing instructions, scripts, and data—significantly enhances agent reliability and organizational knowledge. This approach shifts from prompts to structured assets, with broad implications for AI deployment.

Anthropic has introduced a new approach to building AI agent capabilities, emphasizing that a Skill is a folder containing instructions, scripts, and reference materials, not just a prompt. This redefinition aims to improve consistency, onboarding, and knowledge retention within organizations deploying AI agents, marking a significant shift in how AI workflows are structured and maintained.

According to a detailed write-up from Anthropic’s Claude Code engineer, a Skill is best understood as a container—like a folder—that holds instructions, reference documents, runnable scripts, templates, data, configuration, and hooks. This contrasts sharply with the common misconception that Skills are merely saved prompts or markdown notes. The folder-based model allows AI agents to discover, read, and execute the contents dynamically, creating a durable, reusable asset for organizational processes.

Anthropic’s internal experiments with hundreds of Skills across its engineering organization have shown that this structure leads to more consistent output regardless of who runs the agent, simplifies onboarding by codifying tribal knowledge, and enables compound improvement over time as Skills are refined with each edge case. The company emphasizes that developing a high-quality Skill justifies dedicating an engineer-week, as it becomes an asset that appreciates in value.

Anthropic identified nine core categories of Skills, ranging from library and API references to infrastructure operations, with verification and testing Skills considered the most valuable for ensuring output quality. The approach also involves careful design of the description and scripts within each Skill to trigger appropriate actions, avoiding redundancy and ensuring relevance.

At a glance
reportWhen: published recently, ongoing implementat…
The developmentAnthropic published insights from running hundreds of Skills across its engineering team, revealing a new framework for organizing AI agent capabilities as folders rather than prompts.
A Skill Is a Folder, Not a Prompt — Insights
AI Dispatch · Insights · 1 July 2026

A Skill is a folder, not a prompt

Anthropic published what it learned running hundreds of Skills across its own engineering org. Read as a business memo, the point is bigger than a coding trick: this is how ad-hoc prompting becomes durable institutional capability — the SOPs your agents actually follow, versioned and shared.

✕ The misconception

“A Skill is just a clever markdown prompt you save in a file.”

✓ What it actually is

A folder the agent can discover, read & run — instructions, scripts, references, templates, config & on-demand hooks.

Anatomy of a Skill — the file system is context engineering
my-skill/the unit you share & version
├─ SKILL.mdroot instructions + a description written for the model (its trigger)
├─ references/deep detail pulled in only when needed — progressive disclosure
├─ scripts/real code, so the agent composes instead of rebuilding boilerplate
├─ assets/templates & files to copy into the output
├─ config.jsonsetup the agent asks for if it’s missing (e.g. which Slack channel)
└─ hooks + memoryon-demand guardrails + an append-only log so it remembers
Why it matters: the folder itself is the knowledge base. The agent reads the root, then reaches deeper only when the task demands it — the same way you’d hand a new hire a one-pager that points to the detailed docs.
The nine types — a gap-analysis map for your own library
1Library / API reference
2Product verification ★ top impact
3Data fetching & analysis
4Business-process automation
5Code scaffolding & templates
6Code quality & review
7CI/CD & deployment
8Runbooks
9Infrastructure operations
By Anthropic’s own measurement, verification Skills — the ones that check the work — moved output quality the most. If you build one category well, build that one.
The craft — what separates a good Skill from a useless one
Gotchas = highest-signal section Describe for the model, not humans (it’s the trigger) Don’t state the obvious Ship scripts, not just prose On-demand guardrail hooks (/careful, /freeze) Let it remember (log / SQLite) Don’t railroad — leave room to adapt
The take

The knowledge of how your organization actually operates can be captured, versioned, shared & executed — and the thing capturing it is a humble folder with a script and a gotchas list inside. For the builder, that’s context engineering with real tools attached. For whoever owns the budget, it’s the difference between AI that starts from zero every morning and an asset that compounds. Caveats: best practices are still evolving, checked-in Skills cost context, and curation beats accumulation. Start with one Skill, one gotcha, and the category that catches your mistakes.

Source: “Lessons from building Claude Code: How we use skills,” Thariq Shihipar (Anthropic), Claude blog, 3 June 2026. Categories, examples & measured claims are Anthropic’s; framing is the author’s. Docs: code.claude.com/docs/en/skills.
thorstenmeyerai.com

Transforming AI Capabilities into Reusable Organizational Assets

This development signals a shift from ad-hoc prompting to structured, maintainable AI workflows that embed tribal knowledge directly into the agent’s capabilities. For organizations, this means more reliable, scalable, and explainable AI deployment, reducing the reliance on repeated manual instructions and fostering continuous improvement. The approach could set a new standard for operationalizing AI in enterprise settings, making AI agents more like integrated tools than ephemeral prompts.

Building AI-Powered Products: The Essential Guide to AI and GenAI Product Management

Building AI-Powered Products: The Essential Guide to AI and GenAI Product Management

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

From Prompt Engineering to Asset Building in AI Deployment

Until now, most teams using AI coding agents relied on manually re-entering instructions daily, leading to inconsistency and inefficiency. Anthropic’s recent publication builds on broader industry efforts to formalize AI capabilities, but its key innovation is framing these as persistent, versioned assets—Skills—rather than transient prompts. This approach aligns with ongoing trends toward operational AI, emphasizing reliability, scalability, and institutional memory.

Anthropic’s internal experiments with hundreds of Skills across nine categories reveal that refining and cataloging these assets can substantially improve output quality and reduce onboarding time. The company advocates for investing engineer time into developing high-quality Skills as a strategic asset, not a cost.

“A Skill is a container for how your organization actually does a thing—tribal knowledge, guardrails, tools—not just a prompt.”

— Thorsten Meyer, AI engineer at Anthropic

Programming Game AI by Example: .

Programming Game AI by Example: .

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Unclear Aspects of Skill Implementation and Adoption

While Anthropic’s internal results are promising, it remains unclear how broadly this approach has been adopted outside its organization or how easily other teams can implement similar systems. Details about the specific technical requirements, integration challenges, and scalability across different organizational contexts are still emerging. Additionally, the long-term impact on AI reliability and maintenance costs has yet to be fully evaluated.

Knowledge Management: Systems and Processes in the AI Era

Knowledge Management: Systems and Processes in the AI Era

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Next Steps for Broader Adoption and Validation

Anthropic is expected to publish more detailed case studies and technical documentation to facilitate adoption by other organizations. Further research and experimentation will likely assess how Skills integration impacts large-scale AI deployment, maintenance, and continuous improvement. Industry watchers anticipate that other AI developers may adopt similar frameworks if proven effective, potentially transforming enterprise AI workflows.

Agentic AI Playbook: A Practical Guide to Design, Build and Scale AI Agents

Agentic AI Playbook: A Practical Guide to Design, Build and Scale AI Agents

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Key Questions

How does a Skill differ from a prompt?

A Skill is a folder containing instructions, scripts, and reference materials, enabling dynamic discovery and execution, whereas a prompt is a static instruction or question sent to an AI model.

Why is organizing Skills as folders more effective?

Folders allow for reusable, versioned assets that include not just prompts but also tools, data, and guardrails, leading to more consistent and maintainable AI behaviors.

Can this approach improve AI reliability in enterprise settings?

Yes, by embedding tribal knowledge and operational procedures directly into Skills, organizations can reduce errors, improve consistency, and streamline onboarding.

What are the main categories of Skills identified by Anthropic?

They include library references, product verification, data analysis, business automation, code scaffolding, quality review, deployment, runbooks, and infrastructure operations.

Source: ThorstenMeyerAI.com

You May Also Like

ULA launches final Atlas 5 rocket supporting Amazon Leo’s broadband internet satellite constellation

United Launch Alliance successfully launched its last Atlas 5 rocket, supporting Amazon Leo’s broadband satellite constellation. The launch marks the end of an era for the rocket family.

Mobilised, Not Spent: What’s Left of Europe’s €200 Billion AI Offensive

Europe aims to mobilize €200 billion for AI, but only a fraction is committed or available now, raising questions about its actual impact and timeline.

PostgreSQL And The OOM Killer: Why We Use Strict Memory Overcommit

PostgreSQL adopts strict memory overcommit policies to avoid triggering the Linux OOM killer, balancing performance and stability in high-demand environments.

Micro-agency Proposal Scope Checker

A new AI tool for small web agencies to review and flag scope risks in fixed-scope proposals is entering initial testing, aiming to improve margins and clarity.