Anthropic Is Donating Agent Skills. Your Business Logic Should Travel Too.

Anthropic announced on October 22 that it is donating Agent Skills to the Agentic AI Foundation. For a business using AI, the interesting question is what can leave with you when you change tools. A reusable procedure can travel. Your authenticated connections, application behavior, and operating results still need separate work.

Disclosure: This article was generated by an autonomous AI system using Keith Azodeh’s documented work and the linked sources. The pricing-skill design below is a proposed example, not a claim that TMC has adopted Agent Skills or that its workflow has been tested across agent products.

The announcement concerns stewardship of an existing open standard. It is not the first release of Skills. Anthropic says the donation includes the name, logo, GitHub organization, and domain, and that using or contributing to the standard does not require joining the foundation. The practical opportunity is to make the procedure an asset you can inspect and maintain outside a single conversation.

The business rule is the part worth keeping

My TMC Price Intelligence work is a useful starting point. During the delivered 2023–2025 engagement, supplier emails and attachments became structured pricing records, connected to comparison, reporting, quotes, and invoice workflows. Product, size, price, region, and supplier had to become usable information. Interpretation was one stage of a larger process.

That distinction matters when someone proposes “moving the AI workflow” to another tool. Moving the instruction text would preserve only part of the system. The destination still needs a way to receive the files, represent the output, and connect it to the business application.

Consider an illustrative offer: tomatoes at $18 per case. A second supplier lists tomatoes at $0.90 per pound. The model may read both numbers perfectly. Without a known case weight and an agreed comparison rule, ranking them by the smaller number would be nonsense. This example uses invented prices; it does not describe a TMC transaction.

The knowledge worth retaining is the decision procedure: which units are comparable, what evidence supports a conversion, and what to do when information is missing. Rewriting that procedure in every new chat creates more opportunities for it to drift.

Package a narrow procedure

The Agent Skills overview describes a folder with a required SKILL.md file and optional scripts, references, and other resources. Agents discover the name and description, then load the fuller instructions when needed. That gives the procedure a more durable home than an expanding chat history.

For the proposed pricing example, I would define one task: turn a supplier offer into a candidate record for review. The instructions would identify the required fields and explain how to preserve the supplier’s original unit. A reference file would hold approved product and pack-size conventions. Synthetic examples would show an ordinary offer, an ambiguous unit, and a missing region.

Keep the task’s output narrow. Producing a candidate record should not quietly include permission to email a customer, change a live quote, or issue an invoice. Those actions belong to the surrounding application and its controls.

This separation also leaves room for experimentation. Try a different extraction approach against the same examples. If it invents a pack weight, retain that failure as a new example with an explicit expected result: unresolved conversion, requiring review. The procedure improves through specific corrections instead of another paragraph demanding perfection.

Port the procedure, then adapt the plumbing

The format specification includes an optional compatibility field for environment requirements. Bundled scripts depend on what the agent implementation can run. A familiar folder structure therefore does not establish that two products have the same runtime or tool access.

I would separate the reusable procedure from a small adapter for each destination. The procedure describes the candidate record and the comparison rules. The adapter handles receiving a document and passing the validated result to the application. Credentials remain in the destination’s approved credential system.

Before using a second agent product, run the same synthetic offers through both integrations. Compare whether the records preserve the source unit, whether missing information remains missing, and whether the output can be consumed by the application. A result that looks correct in chat may still use a field name the importer does not recognize.

If a rule can be checked deterministically, put that check in code. Instructions can tell an agent to provide a currency, but the receiving application can reject a record that has none. Portability becomes easier to assess when success has a concrete shape.

An open format changes the purchasing question

For a small team choosing an AI product, the useful question becomes: which parts of our procedure can we take elsewhere, and which parts depend on this integration? That makes a tool evaluation more specific than comparing the quality of two demonstrations.

There is still maintenance work. Someone must update the approved conventions, review changes to the procedure, and rerun examples when an adapter changes. A shared format makes that work easier to locate; it does not perform it automatically.

Nor does packaging instructions establish a token-cost saving. Loading relevant material on demand may reduce unnecessary context, but retries, tool output, and model behavior still affect usage. Measure cost per accepted candidate record, including correction work, before claiming an improvement.

Anthropic’s donation is a reason to inspect where your operating knowledge lives. Start with one recurring procedure, preserve the examples that taught you its edge cases, and keep the destination-specific connection small. Then changing an AI tool becomes a bounded integration decision rather than an attempt to reconstruct the business from old chats.