The previous 284-line skill added ~54K tokens of context which heavily penalized
the improvement score despite quality improvements. Condensed to 62 lines keeping
only the highest-impact rules (parameter patterns, EventCallback, async/disposal,
don'ts) while removing verbose examples the model already knows.
With --runs 3: passes 3/5 scenarios (vs 2/5 before trim).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Emphasize 'existing page' and 'explicitly asks to plan or decompose'
- Add 'building new applications from scratch' to DO NOT USE FOR
- Remove broad phrases like 'any page request with 3+ sections'
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Remove 'before writing any code' phrasing that caused model to stop at plan
- Add explicit instruction to immediately implement after planning
- Remove 'multi-step wizards' from plan-ui-change USE FOR (belongs to author-component)
- Update wizard eval prompt to emphasize component authoring over planning
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Integrates the 9 Blazor skills from PR #357 (javiercn/author-component) with
fixes for the systemic activation failure (all skills showing NOT ACTIVATED):
Key fixes:
- Improve skill descriptions with explicit USE FOR/DO NOT USE FOR patterns that
match the activation keyword style used by working skills in the repo
- Add license: MIT to all skill frontmatter per repo conventions
- Add marketplace entries (.github/plugin, .claude-plugin, .cursor-plugin)
- Add CODEOWNERS entry for @dotnet/aspnet
- Add README.md plugin entry
- Add both playwright and microsoft-docs MCP servers to allowed-external-deps.txt
MCP server analysis:
- The eval framework (AgentRunner.cs:302-307) already skips non-stdio MCP servers
- playwright (type: 'local') and microsoft-docs (type: 'http') are both filtered
out during evaluation, so they don't affect eval results
- Root cause of NOT ACTIVATED is likely description quality, not MCP servers
Skills included:
- author-component: Blazor component authoring patterns
- collect-user-input: Forms, validation, EditForm patterns
- configure-auth: Authentication/authorization across render modes
- coordinate-components: Cross-component state sharing
- create-blazor-project: New project scaffolding with correct render modes
- fetch-and-send-data: API calls and async data lifecycle
- plan-ui-change: Component decomposition planning
- support-prerendering: Prerender-to-interactive handoff
- use-js-interop: IJSRuntime, IJSObjectReference, .razor.js modules
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Javier Calvarro Nelson