docs(release): generate value-prop-first release notes from the gaia-release skill

[kovtcharov-amd]

Why this matters

GAIA's release notes read dry and engineering-first — they describe what changed, not why a user should care or want to try it. We ship genuinely useful local agents and then sell them poorly: Agent UI launched with no demo, and per-release the agents that are the actual product get buried under SDK/infra detail.

This upgrades the gaia-release Claude Code skill so it generates value-prop-first notes the first time, instead of validating tone after the fact. No new validator, no LLM judge, no extra tests — the fix is better generation guidance where the notes are written.

What the skill now enforces when drafting:

• Value-prop first — lead with what the user can now do and why, not the implementation.
• Local agents are the headline — agents that solve real problems come first; SDK/infra is supporting detail.
• One agent/command per highlight — gaia browse and gaia analyze each stand alone with their own utility line.
• Plain language, engaging but factual, no emoji/fluff — banned-phrase list promoted from prose to an explicit rule.
• A bad/good worked example (email triage) gives the model a concrete target.
• The Phase 6 Discord announcement reuses the same parameters, so it stops reading dry too.

Test plan

• Read .claude/skills/gaia-release/SKILL.md Phase 1 "Generation parameters" block and the bad/good example for accuracy and house-style fit.
• On the next release, run the skill and confirm the drafted ## What's New is per-agent and value-prop-first.

[github-actions]

Clean, targeted fix for a real problem — the generation guidance is now actionable enough that a model can apply it without drift. Approve.

Summary

The gaia-release skill was telling the model not to do bad things (banned phrases) but not what good looks like, which meant every release required after-the-fact tone correction. This PR promotes the banned-phrase list to an explicit ruleset, adds a concrete bad/good worked example, and cross-references the same parameters for the Phase 6 Discord draft — closing the loop so both artifacts get the same voice.

No code changes, no tests required (the test plan is correct: guidance is validated by reading the result, not by running pytest).

Issues Found

🟢 Historical framing in the "Generation parameters" intro (.claude/skills/gaia-release/SKILL.md)

"GAIA's notes have historically read dry and engineering-first" reads as past-tense commentary rather than instructional guidance. A future reader skimming for rules could miss the transition into the parameter list. Tiny rewording:

   **Generation parameters (apply to every entry — this is the point of the skill).**
   Without these, notes default to engineering-first — they say *what changed* but
   not *why a user should care or want to try it*. Generate against these every time:

Strengths

• Concrete worked example is the right call. A bad/good pair with real prose is more useful than five more abstract bullets — the model has something to pattern-match against, not just a list of negatives.
• Cross-referencing Phase 6 (Discord) ensures both the release notes and the announcement announcement use the same voice. Previous version only constrained the notes; the announcement was left to drift.
• Scope-clean. Exactly one file, no drive-by edits, no half-finished additions. The key-rules bullet at the top points to the new section rather than duplicating it inline — good information hygiene.

Verdict

Approve. One optional nit on the intro framing; nothing blocking.