stack-contextlisted
Install: claude install-skill SilverAssist/agents-toolkit
<!--
Note on the `model: haiku` pin above: this skill's own work — reading package.json,
running `npm view`, diffing semver, skimming a release-notes page for bullet points — is
mechanical. It runs at the *start* of nearly every dev-flow prompt (see "When to use"),
so keeping it cheap matters more than for a once-per-PR pass like `core-review`. Escalate
to the invoking prompt's own model only for judging *relevance* of a new API to the task
at hand — that reasoning happens in the caller, not in this skill.
Same portability caveats as `core-review` apply: Copilot ignores unknown skill
attributes (cosmetic linter warning only), skills have no independent model on Copilot
(they inherit the invoking prompt's), and Codex has no per-skill `model:` at all.
-->
# Silver Assist — Stack Context
Reads what is **actually installed** in this project — exact `next`/`react` versions, not
"Next.js" as a generic label — before any suggestion, plan, or fix is proposed. Every site in
this org runs Next.js 15/16 and React 19, but a suggestion drafted from general training data
routinely regresses to older patterns (Pages Router idioms, pre-App-Router data fetching,
`getServerSideProps`) or misses a capability that shipped two minor versions ago because
nothing in the conversation ever looked at `package.json`.
This skill closes both directions of that gap:
1. **Don't suggest what's stale.** Ground every recommendation in the exact versions installed.
2. **Don't miss what's