Self-improving skills

The assistant can preserve a useful procedure after it has actually carried it out, then reuse that procedure when a similar task comes up. The result is a managed skill that stays visible and under your control.

This is not retraining the model. Facts, preferences, and decisions stay memories. Repeatable procedures can become skills. See Memory and context and Skills and tools.

What are skills?

Skills are reusable sets of instructions, tools, and supporting files that teach the assistant how to handle a kind of task. Cadence ships skills for common work, and you can add more. When a request matches, the assistant loads those instructions instead of starting from scratch.

What self-improving means

After eligible work is complete, memory can preserve the procedure the assistant actually used. It does not turn an untested suggestion, or a procedure it only described, into a skill.

How a skill is learned

Learning starts with completed work: the sequence followed, the tools used, the outcome reached. When that work contains a reusable procedure, Cadence can write it as a managed skill with instructions and supporting files. Future conversations can load it when the request calls for the same kind of work.

What qualifies

A procedure qualifies when it is useful, repeatable, actually carried out, and not already covered by an existing skill. A one-time fact stays memory. Example: “this project uses a particular host” is a fact. The concrete deploy steps the assistant ran can be a skill.

Create vs refine

Before creating anything, Cadence checks skills already available. It creates a new skill only when the procedure is distinct. If an assistant-authored skill already covers the work, it may refine that skill with what the new conversation verified.

It does not overwrite skills you authored or skills you installed.

When you see it

A learned skill is created after the work is complete, then announced for review. You do not need to approve it before it is created. A new skill shows an “I just learned how to do...” card. Opening the card takes you to the skill so you can inspect what was preserved. Routine memory updates that do not create a distinct skill do not produce this card.

Review, edit, or remove

Learned skills are managed skills, not hidden weight changes. From Skills you can view files, edit the text, inspect the source conversation when lineage was recorded, and delete the skill. Removing it stops the procedure from loading as that skill. The original conversation stays.

Boundaries

A learned skill is not active in every conversation. It loads only when the current request is relevant. It does not gain broader access than the tools it uses. Those tools keep their normal permissions.

Authoring walkthrough: Your first skill. Catalog: Skills reference.