Introduce shared-patterns/ with voice-and-tone

[MLatzke]

This PR creates a new top-level shared-patterns directory that houses patterns that are shared across skills.

Currently, it holds:

• voice-and-tone.md — Asta's shared communication conventions: neutral/professional register, no emotional rapport, no cheekiness or judgmental tone, plain language with technical-term definitions, intellectual humility with categorical (not numeric) confidence labels, process rules. Adapted from @psesito https://docs.google.com/document/d/1OzhY7gVmtTwlnuR63vQF7fPurCWI7PCZ1IGM6gD2Pfc/edit?tab=t.0
• SKILL-BOILERPLATE.md — starting point for new Asta skills with frontmatter scaffolding, voice-and-tone reference, and optional sections. Developer-only; excluded from built plugins.

It also updates the build script to copy shared-patterns/ into each plugin, so the relative path resolves consistently across the source and the built artifacts. Each plugin remains self-contained.

First consumer: skills/generate-theories/SKILL.md references voice-and-tone at the top and in the References section.

[jbragg]

I think we need a mechanism outside the skills to load an agent with identity tone etc.
Since skills can be plugged into any agent, they don't have the right to override the user's agent identity etc.
cc: @rodneykinney

[rodneykinney]

Agree. These instructions are more appropriate at the CLAUDE.md level. Agent skills are typically invoked with a clean context. In that case, the top-level agent is responsible for reporting the result of an invoked skill back to the user, so adding tone to skills is not enough.

[MLatzke]

Hmm, I think we're already affecting the voice and tone, though. We're just not doing it intentionally. Multiple agents have the ability to write things. @charliemcgrady showed me an example of a theory output, and the agent sounded sarcastic and patronizing.

We need to figure out a way to guide that output. If the user's agent decides to rewrite something, that's fine, just as long as we're not sending biased text in the first place.

[jbragg]

@MLatzke it sounds like you might be describing a problem with the theorizer agent that should be solved there?

[charliemcgrady]

I agree that something like personality is best modeled via Claude.md, et. al. instead of as a skill in the core plugins. Asta Bot should behave differently in the Slack vs. Web surface.

My recommendation would be to publish this in https://github.com/allenai/design.md. Then our handler + users can optionally adopt this specific personality.